Global Divergence in Nuclear Power Plant Construction: The Role of Political Decentralization

Abstract

Lead time—the duration of construction and commissioning—is an important determinant of the capital cost of nuclear power plants (NPPs). For an industry dominated by a handful of multinational firms, the degree of cross-national variation is surprising. NPP lead times have historically trended upward over time in Western nations, and yet they are comparatively quick and stable in East Asia. I theorize that the institutional capacity and autonomy of subnational governments can partially explain these patterns in the data. Having assembled a novel data set on the design specifications of the global population of NPPs, I empirically document a positive association between political decentralization and NPP lead time that is not explained by observed cross-country differences in NPP design. The results are suggestive of the hypothesis that political decentralization creates conditions that slow NPP construction for nontechnical reasons. However, the findings are not robust to certain robustness checks and fail to rule out the possibility that unobserved differences in design explain this association.

Keywords:

• Economics
• construction
• lead time
• politics
• decentralization

I. INTRODUCTION

It is a stylized fact that the capital costs of nuclear power plants (NPPs) have historically trended upward in Western developed nations. Some scholars have characterized this as “negative learning-by-doing.”^1,2 This trend is often contrasted sharply with the steady downward trajectory of the cost of other electric generation technologies (“positive” learning-by-doing), particularly photovoltaic (PV) solar panels, wind turbines, and gas combustion turbines.³ Budget overruns and schedule slippage in the construction of the AP1000 in the United States and the EPR in Europe indicate that the nuclear industry’s economic woes have yet to be properly addressed. The problematic economics of NPP construction are representative of “megaproject syndrome,”⁴ a theory that applies to massive infrastructure projects broadly, such as airports, urban public transit, high-speed rail, hydroelectric dams, and sports venues.

Academics and industry observers have offered numerous explanations for the root causes of the cost problem for the nuclear industry: construction project mismanagement,⁵ evolution in the political environment and regulatory regime,⁶ lack of standardization in design,⁷ reliance on immature or incomplete designs before beginning construction,⁸ diseconomies of scale,⁹ and added complexities in design arising from innovation in nuclear safety.¹⁰ However, outside the West, historic trajectories and recent results in NPP construction suggest that an upward cost trend is not inevitable and lower costs are possible,¹¹ although this interpretation and the credibility of the underlying data are disputed.^12,13 The present work wades into this fierce debate with two primary contributions: (1) novel, rich data on the design specifications of NPPs (see Appendix A), and (2) a quantitative analysis that connects the study of the nuclear industry to the literature of institutional political economy.

Previous studies of this industry have been haunted by the specter of omitted variable bias: simple cross-country and time-trend analyses of NPP construction outcomes are not necessarily valid for causal inference given that the technical characteristics of NPPs vary across countries and over time.¹⁴ The present work is the first of its kind (to the author’s knowledge) to incorporate detailed data that “look inside” a nuclear reactor. These include such variables as the operating temperature and pressure of the primary coolant, the number of primary coolant loops, the size of the reactor pressure vessel, the choice of cooling technology, and the design of the containment structure. Previous work has been largely limited to power output in megawatts and categorical classifications of the make and model of the reactor. Unfortunately, due to the terms under which I accessed these data from the International Atomic Energy Agency (IAEA), much of the underlying data cannot be publicly made available for replication. Nevertheless, all of the analyses I present herein—except the estimation of Eq. (3(3) $ln(\mathrm{L}{\mathrm{T}}_i)=\sum _{s\in S}{\theta }_{s}Spe{c}_{s,i}+{\delta }_t+\gamma M_i+{\mu }_c+{\nu }_y+{\epsilon }_i.$(3) )—can be replicated with the data I have provided in the online data appendix (https://github.com/a-g-benson/Global-NPP-Database).

In seeking to explain the high degree of cross-national variation, I observe the long and storied history of local opposition as a factor in the siting, regulation, construction, and cancellation of NPPs. I argue that the political economy of nuclear power is characterized by locally concentrated risks and diffuse national (and global) benefits, in an inversion of the classic problem formalized by Ref. 15. Hence, NPPs are expected to face greater regulatory hurdles and political constraints in countries whose subnational governments have greater autonomy and institutional capacity. This generates a suite of hypotheses regarding how the degree of federalism or regional autonomy (“decentralization,” for brevity) influences the design characteristics of nuclear reactors, the speed of their construction, and the industry’s ability to improve upon past performance through learning.

To perform the quantitative analysis, I combine the technical data on reactors with economic and political data regarding the nation in which the NPP was constructed, including democracy, regime change, decentralization, national level of economic development, and utility ownership (public or private). In the present work, I take lead time (LT) as the sole outcome of interest, due to data availability and quality issues associated with overnight capital cost (OCC). The headline findings of the analysis are as follows:

I find no significant association between a nation’s political conditions and the expected LT of its NPPs given their observed design specifications. In other words, highly decentralized countries do not systematically choose design characteristics of NPPs that would tend toward longer LTs.

Instead, I find a statistically significant and economically substantive association between decentralization and actual LT, when holding design characteristics constant. The estimated effects imply that a one standard deviation increase in a nation’s political decentralization is associated with approximately a 9.5% increase in LT, which amounts to 8 months of additional LT for the typical 1-GW reactor.

However, this second finding is not robust to the strictest possible test, whereby I only compare reactors of identical models through the inclusion of fixed effects. This analysis fails to reject the null hypothesis of no association between LT and decentralization. One possibility is that unobserved technical differences in design may explain the empirically observed raw correlation between LT and decentralization. Alternatively, because many models of reactors are observed in only one country or a few similar countries, high reliance on fixed effects may sap the data of statistical power. Richer data on technical specifications—particularly those related to safety systems—are needed to resolve this question.

My final finding relates to the theory of “megaproject syndrome.”¹⁶ As previous research has shown, larger NPPs take longer to build, and I replicate that result here. I extend this finding in two ways: I generate a more comprehensive measure of scale and project complexity, namely, an NPP’s expected LT conditional on its size and design specifications, and validate it as having predictive power in explaining observed LT. Then I show that expected LT does not correlate with observed LTs in a one-to-one relationship in all countries. In particular, I find that East Asian nations—Japan, South Korea, mainland China, and Taiwan—have historically completed construction of their NPPs much faster than would otherwise have been expected on account of the “megaproject-iness” of their NPPs. However, I find no evidence for the hypothesis that decentralization mediates the relationship between NPP scale and LT.

The outline of this paper is as follows. Section II reviews the literature and elaborates on the theory that motivates the empirical analysis. Section III summarizes the data set assembled for this paper. Section IV formally lays out the econometric specifications. Section V presents the results. Section VI discusses the results and proposes directions for future research. The data sources, cleaning and coding procedures, and restrictions on the availability of the data are detailed in Appendix A. Appendix B addresses several methodological issues and assumptions.

II. BACKGROUND, THEORY, AND PRIOR WORK

II.A. Measurement of Capital Cost in the Electricity Sector

The two most widely studied outcomes in the literature on NPP construction are OCC and LT.

Overnight capital cost consists of all outlays on materials, manufactured components, construction equipment, construction labor, engineering services, land, and permitting fees. These are what economists call accounting costs. The designation “overnight” refers to the hypothetical case of a power plant constructed from start to finish over the course of a single night. Effectively, no interest would accumulate during construction. While not a complete measure of capital cost, OCC enables comparisons of the capital costs of different NPPs independently of financing parameters, which can vary due to macroeconomic conditions, government policies to subsidize the cost of capital, and other factors outside the control of the firm building the plant. Figure 1 presents the OCC of NPPs in the United States.

Fig. 1. OCC of NPPs in the United States.²²

In this work, I use LT to denote the length of time between the initiation of major construction activities and the start of commercial operation. By convention in the nuclear industry, the initiation of major construction activities is considered to begin with the pouring of concrete for the foundation of the plant.¹⁷ The start of commercial operation is usually “declared” after several weeks to months of test operations have been completed and the plant begins operating full time. Because NPPs require a considerably longer amount of time to construct than competing technologies in the electricity sector, financing costs account for a comparatively greater proportion of capital costs for NPPs, around 17% under ideal conditions^18,19 and even higher when delays stretch out construction schedules. The opportunity cost of capital during the construction period is commonly called “allowance for funds used during construction” in the electric utility sector. Figure 2 presents the LT of NPPs in the United States.

Fig. 2. LT of NPPs in the United States (source: IAEA).

II.B. Prior Quantitative Studies of OCC and LT

In this section, I primarily review studies that estimate the effect of underlying casual determinants of OCC and LT for NPPs. But first I will briefly mention the prior works that collected and presented the necessary data on which subsequent analyses rely. These works have successively expanded data availability from the United States,^6,20 to France,^1,21 to several other Organisation for Economic Co-operation and Development (OECD) nations,¹¹ and finally 82%^a of the global population of reactors.²² However, most of the foregoing works (with the exception of Ref. 21) do not analyze the underlying causal determinants of LT or OCC in a quantitative or systematic way.

Reference 10 estimates a system of equations for the OCC and LT in the United States and France. They conclude that the French policy of standardization helped reduce cost escalation and schedule slippage relative to the U.S. experience. Their estimated learning effects are conditional on experience from previous NPP construction being accumulated by the same architect-engineer (AE) firm with the same reactor model. Notably, the U.S. market for nuclear reactor design was contested by four major suppliers of nuclear reactors whose designs were routinely customized by approximately 20 different AE firms to meet the requirements of different utilities. In contrast, the French market was monopolized by Framatome as reactor supplier and monopsonized by the state-owned national utility, EDF, which performed in-house architect-engineering for its plants.

In addition, Berthélemy and Escobar Rangel estimate a model of LT alone on a larger sample, adding observations from Canada, the United Kingdom, Japan, and South Korea. This analysis lends further support for the hypothesis that standardization of reactor design helps to reduce LT.

The LT of the global population of NPPs was investigated by Csereklyei et al.⁷ using duration analysis.^b The authors find several economic conditions influence NPP construction: higher levels of gross domestic product (GDP) per capita, higher expectations of future economic growth, and higher oil prices are associated with shorter LTs. Furthermore, they find partial evidence for the benefits of standardization. They show some—but not all—reactors of certain standardized designs tended to be built faster compared to those of nonstandardized design.

Regarding political factors, Csereklyei et al. find both autocracy and democracy are associated with faster construction, where anocracy (Polity IV^c score between −5 and +5) is the reference category. But the standard errors on the effects are very large; they find a statistically significant effect of democracy in only one econometric specification. They find no statistically significant effect of the accidents at Three Mile Island (TMI) or Chernobyl on LT, which contrasts sharply with the conventional wisdom among industry observers, prior academic findings,¹⁰ and the results I find in Table VIII in Sec. V.B.

TABLE I Count of Completed Reactors by Country as April 6, 2021

Western Bloc		Eastern Bloc		East Asia		Other
United States	133	USSR	69	Japan	59	India	22
France	70	Czechoslovakia	13	China	49	Pakistan	5
United Kingdom	45	Russia (post 1991)	8	South Korea	26	Argentina	3
West Germany	30	East Germany	6	Taiwan	6	Mexico	2
Canada	25	Bulgaria	6			Brazil	2
Sweden	13	Hungary	4			South Africa	2
Spain	10	Romania	2			Yugoslavia	1
Belgium	8					Iran	1
Switzerland	6					UAE	1
Italy	4
Finland	4
Netherlands	2
Subtotal	350	Subtotal	108	Subtotal	140	Subtotal	39

TABLE II Descriptive Statistics of NPP LT by Region*

Region	Mean	Standard Deviation	N
Western Bloc	7.6	3.4	350
Eastern Bloc	7.7	2.9	106
East Asia	5.5	1.4	139
Global South	10.6	5.0	38
World	7.4	3.3	633

*In years.

TABLE III Lead Time by Type of Reactor*

Acronym	Reactor Type	Mean	Standard Deviation	N
PWR	Pressurized water reactor	7.5	3.2	359
BWR	Boiling water reactor	6.5	3.0	116
PHWR	Pressurized heavy water reactor	8.3	3.4	57
GCR	Gas-cooled reactor	7.6	4.9	52
LWGR	Light water graphite reactor	6.5	1.5	30
Other	Miscellaneous reactor types	8.3	3.4	23
Total		7.4	3.3	637

*In years.

TABLE IV Descriptive Statistics of Decentralization by Global Region

	Has Federal Constitution		RAI Self-Rule Index
Region	mean	n	mean	n
East Asia	0.00	140	11.3	140
Western Bloc	0.58	350	17.0	350
Eastern Bloc	0.83	108	11.6	8
Global South	0.92	39	17.7	36
World	0.51	637	15.5	534

TABLE V Marginal Effects of Design Characteristics on NPP LT*

	Marginal Effect	(Standard Error)
100 MW(electric) increase in power output	5.7%	(0.8%)
1°C increase in reactor outlet temperature	0.2%	(0.04%)
1 MPa increase in operating pressure of the primary coolant	−0.6%	(1.9%)
One additional primary coolant loop	3.1%	(1.3%)
One additional pump/circulator in the reactor coolant system	−0.2%	(1.1%)
OTC to discharge waste heat	−13.2%	(3.4%)
Standardized reactor design	−10.3%	(3.9%)
Measurement error related to multi-unit construction	10.0%	(2.2%)
Reactor Type Fixed Effects
PWR: pressurized water reactor (reference category)	—	—
BWR: boiling water reactor	−27.4%	(12.5%)
PHWR: pressurized heavy water reactor	−13.3%	(13.9%)
GCR: gas-cooled reactor	−28.4%	(21.3%)
LWGR: light water graphite reactor	−22.8%	(17.9%)
Other: miscellaneous reactor designs	−22.6%	(23.3%)
Containment Design Fixed Effects
Large dry (reference category)	—	—
Subatmospheric	−6.2%	(12.8%)
Vacuum building	2.0%	(14.4%)
Ice condenser	16.5%	(11.2%)
Mark I	41.6%	(21.4%)
Mark II	45.3%	(23.0%)
Mark III	42.2%	(23.1%)
Other BWR containment design	36.4%	(22.7%)
No containment	35.0%	(29.7%)
Other containment/missing data	53.6%	(20.4%)
Country fixed effects	✓
Year fixed effects	✓
Observations	582

*Dependent variable: $ln(L{T}_i)$.

TABLE VI Predictors of Reactor Design Standardization*

	(1)	(2)	(3)
GDP per capita	1.93	0.93	5.15
One standard deviation increase in $ln(GDPp{c}_{c,y})$	(1.30)	(−0.18)	(1.64)
Democracy	0.75	3.43	2.57
One standard deviation increase in $Dem$	(−0.97)	(3.23)	(2.21)
Decentralization	0.94
1 if country has a federal constitution	(−0.10)
One standard deviation increase in $De{c}_{c,y}$		0.67	1.31
		(−1.21)	(0.46)
Population			2.76
$ln(Po{p}_{c,y})$			(2.21)
Electricity sector fragmentation			0.26
$ln(NumUti{l}_c)$			(−2.02)
Investor-owned utility	0.19	0.31	1.04
$IO{U}_i$	(−2.36)	(−1.77)	(0.05)
Time trend	1.12	1.17	1.09
Year construction began	(3.59)	(5.27)	(2.33)
Observations	774	626	626

*Dependent variable: 1 if the reactor is standardized, 0 otherwise; odds ratio shown with t-statistics in parentheses.

TABLE VII Estimation Results of Eq. (4(4) $ln(\widehat{\mathrm{L}{\mathrm{T}}_i})={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}+{\epsilon }_i.$(4) )*

	(1)	(2)	(3)
GDP per capita	12.2%	13.1%	−2.15%
One standard deviation increase in $ln(GDPp{c}_{c,y})$	(1.79)	(1.51)	(−0.41)
Democracy	1.11%	−0.35%	4.57%
One standard deviation increase in $De{m}_{c,y}$	(0.40)	(−0.09)	(1.10)
Decentralization	−0.31%
1 if country has a federal constitution	(−0.05)
One standard deviation increase in $De{c}_{c,y}$		0.12%	−1.07%
		(0.03)	(−0.23)
Capacity			3.94%
100 MW(electric) increase in power output			(3.73)
Observations	612	513	513
Adjusted $R^2$	0.074	0.063	0.269

*Dependent variable: $ln(\widehat{L{T}_i})$; transformed marginal effects on $\widehat{L{T}_i}$ in in percent and t-statistics in parentheses.

TABLE VIII Estimation Results of Eq. (5(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) )*

	(1)	(2)	(3)	(4)
TMI accident	17.7%	19.4%	20.1%	23.6%
Under construction on March 28, 1979	(3.20)	(2.31)	(4.13)	(3.10)
TMI accident $\times$ USA	37.0%	40.2%	31.4%	35.1%
Under construction on March 28, 1979 in the USA	(3.98)	(3.47)	(5.07)	(4.17)
Chernobyl disaster	2.8%	−2.5%	13.2%	6.0%
Under construction on April 26, 1986	(0.29)	(−0.22)	(1.92)	(0.85)
Chernobyl disaster $\times$ USSR	−31.6%	Not	−31.7%	Not
Under construction on April 26, 1986 in the USSR	(−2.92)	estimable	(−3.63)	estimable
Fukushima Daiichi disaster	−11.9%	−18.1%	9.9%	0.8%
Under construction on March 11, 2011	(−0.73)	(−0.86)	(1.49)	(0.10)
Regime change	31.8%	19.1%	50.3%	48.1%
Under construction during regime change	(2.39)	(1.04)	(3.13)	(2.10)
GDP per capita	−0.9%	4.5%	−9.2%	−11.4%
One standard deviation increase in $ln(GDPp{c}_{c,y})$	(−0.20)	(1.34)	(−2.38)	(−2.34)
Democracy	−4.1%	−6.9%	1.0%	0.4%
One standard deviation increase in $De{m}_{c,y}$	(−1.50)	(−1.64)	(0.41)	(0.12)
Decentralization	7.1%		22.6%
1 if country has a federal constitution	(1.08)		(5.00)
One standard deviation increase in $De{c}_{c,y}$		1.4%		9.8%
		(0.40)		(3.17)
Investor-owned utility	−10.4%	−10.5%	−16.0%	−14.9%
$IO{U}_i$	(−2.44)	(−2.44)	(−6.05)	(−3.41)
Expected LT, conditional on design	0.85%	0.86%	0.68%	0.71%
1% increase in $\widehat{L{T}_i}$	(5.32)	(4.94)	(8.17)	(7.30)
Reactor model fixed effects	✓	✓
Observations	556	476	590	500
Adjusted $R^2$	0.720	0.740	0.663	0.660

*Dependent variable: $ln(L{T}_i)$; transformed marginal effects on $L{T}_i$ in percent and t-statistics in parentheses.

In a series of three closely related papers,^2,23,24 Sovacool et al. analyze a sample of 401 projects in the electricity sector, consisting of several different types of power plants (fossil, nuclear, solar PV, solar thermal, wind, biomass) and high-voltage transmission lines. They present data on budget overruns and schedule slippage (i.e., increases in the OCC and LT relative to original estimates). Comparing all the types of projects studied, they find that (1) NPPs most frequently exhibit budget overruns and (2) NPP budget overruns are, on average, the largest as a percentage of initial budget relative to all other technologies considered. Another noteworthy finding is that budget overrun and schedule slippage are positively correlated with each other for NPPs.^d This is consistent with the findings of Portugal-Pereira et al.,²² who report a correlation of $r=0.48$ between OCC and LT. Reference 25. (an American AE firm involved in several NPP projects) attribute the relationship between time and cost to the effect of delays on labor productivity. For example, failed inspections and design changes are said to have a “triple penalty”: the cost of the initial work, the cost of removing the initial work, and the cost of performing the work again. Such work also comes at the cost of a longer LT.

For the present work, I have selected LT as the sole outcome of interest for several reasons. First, the data are available for the global population, which bolsters statistical power. Second, LT is a more transparent and consistently recorded metric, whereas OCC data are subject to disputes regarding accounting practices, inflation adjustment, currency conversion, and trustworthiness of data sources. Third, LT is an economically important outcome per se, as it plays an essential role in the accumulation of financing costs during construction and schedule slippage tends to correlate with budget overruns. Last, modeling the endogenous interactions between the OCC and LT is beyond the scope of the present work. Future research could extend the present work by modeling the simultaneous determination of the OCC and LT as in Ref. 10 while using the OCC data compiled by Portugal-Pereira et al.²²

II.C. Learning-By-Doing

Learning-by-doing is a theory of endogenous technological change that ascribes cost reductions and quality improvements to the accumulation of practical experience with a production process.²⁶ The conventional model of learning-by-doing posits the following relationship between some outcome $Y_t$ (typically, cost per unit) and cumulative experience, $Ex{p}_t$, based on the work of Wright²⁷:

(1) $ln(Y_t)=\alpha +\beta ln(Ex{p}_t)+{\epsilon }_t.$(1)

Assuming lower values of the outcome are more desirable, the production process is said to exhibit learning-by-doing when $\beta <0$. In practice, as a technology matures, the level of the outcome over time ceases to be characterized by Eq. (1)(1) $ln(Y_t)=\alpha +\beta ln(Ex{p}_t)+{\epsilon }_t.$(1) and reaches some relatively stable level. This level would be determined exogenously by physical limits to the production process and the price of inputs.

A common method for contextualizing the magnitude of $\beta$ is the progress ratio (PR) or learning rate (LR):

(2) $1-2^{\beta }=1-\mathrm{P}\mathrm{R}=\mathrm{L}\mathrm{R},$(2)

where PR is interpreted as the relative level of the cost (or other outcome) after a doubling of cumulative production as compared with the prior level and LR is the percentage reduction in cost (or other outcome) arising from a doubling of cumulative production. For example, $\beta =-.32$ generates PR = 80% (a cost equal to 80% of the prior level) and LR = 20% (a 20% reduction in cost).

Several improvements to the operating performance of NPPs have been documented, such as increased reliability,^28,29 increased power output,^29,30 reduced occupational exposures to radiation,³¹ and reduced rates of initiating events (precursors of more serious safety problems).³² However, the empirical evidence regarding learning in NPP construction paints a more dismal picture. Rubin et al.³ survey the literature on learning-by-doing in the capital costs of energy technologies, reporting mean one-factor^e learning rates of 15% for natural gas combustion turbines, 12% for wind turbines, 23% for solar PV, and 11% for biomass generation, inter alia. Their review of learning rates for nuclear power captures only four studies, which report values ranging between −38% (Ref. 1) and 5.8% (Ref. 34). Subsequent to the public release of more authoritative data on the costs of France’s nuclear reactor fleet, Rangel and Lévêque²¹ argued that the cost estimates underlying the calculations of Grubler¹ were too high for later reactors. The findings of Berthélemy and E. Rangel¹⁰ correspond to a learning rate of 10%,^f conditional on the same design of plant being built by the same AE.

One hypothesis for the poor rate of learning in NPP construction is the high degree of onsite construction work as a share of the total cost. Estimates from United Engineers and Constructors²⁵ suggest that equipment manufactured off-site accounts for approximately 21% of the base cost^g of a typical American pressurized water reactor (PWR) built in the 1980s. Factory fabrication is theorized to better facilitate learning-by-doing,³⁵ for reasons such as assembly line production methods, a stable workforce, and consistent and well-controlled workplace conditions. Lessons learned at one construction site may not disseminate as readily to the next site, such as when different workers are employed at the two sites.

A strong contrast can be drawn between nuclear fission and solar PV in this respect. The price of PV modules constituted 74% of the total cost of rooftop solar panel installations in Germany in 2007; following dramatic declines in global module prices, that share fell to 39% as of 2019 (Ref. 36). This decline is consistent with evidence for faster learning in PV module manufacturing than in PV module installation. Elshurafa et al.³⁷ estimate a learning rate of 11% for balance-of-system costs of solar PV installations, whereas the median learning rate for PV modules among the studies included in Rubin et al.³ is 20%.^h Furthermore, the high initial share of cost associated with the module provided a greater scope for manufacturing-based learning effects to reduce the overall capital cost of solar PV.

Many commentators emphasize the role of standardization in fostering beneficial learning effects in the nuclear industry.^10,11,21,38 However, technologies such as solar panels, wind turbines, and gas combustion turbines appear to have achieved considerable learning despite a much larger number of firms engaged in each industry, with each firm offering competing designs, relying on proprietary innovations, and regularly introducing new product lines. Why is cumulative industry experience a meaningful predictor of cost reductions for these technologies but not for nuclear power?

To illustrate this, consider the example of General Electric (GE), a large multinational firm engaged in a variety of industries, including several different energy technologies. General Electric currently advertises 21 different models of gas combustion turbine on its website,³⁹ many of which come in two different versions depending on whether the customer’s grid runs at 50 or 60 Hz. This high diversity of product offerings—and the development costs that each product entails—is sustained by a large volume of orders. General Electric boasts that over 1100 of its F-class turbines have been installed at power plants to date,³⁹ the first of which entered commercial operation in 1990 (Ref. 40). General Electric claims sales of over 3000 units of its smaller B and E class turbines. Such a high volume of sales can sustain the serial manufacture of several different, standardized models.

Now consider GE’s involvement in the nuclear industry. General Electric was the first commercialize boiling water reactor (BWR) technology, beginning with Dresden Unit 1 in 1960. To date, a mere 99 commercial-scale BWRs have been built by GE and firms to which it licensed its technology. These 99 reactors consist of several different product lines, and most of these exhibit a staggering degree of internal diversity.⁴¹ BWR-1 is a designation retroactively applied to a hodgepodge of early designs, which is perhaps to be expected in the early stages of technological development. The BWR-2 was obsolete before the first one had entered commercial operation,ⁱ as GE quickly returned to the drawing board and the first BWR-3 began construction several years earlier.^j BWR-4s and BWR-5s have been mixed and matched with the Mark I and Mark II containment designs.^k The BWR-6 started to exhibit more standardization; it was exclusively paired with Mark III containment, and GE applied to the U.S. Nuclear Regulatory Commission (NRC) for approval of a “Standard Safety Analysis Report.” Yet the BWR-6 was offered in three different sizes of reactor pressure vessel, each requiring its own safety analysis. The first truly standardized BWR was the ABWR, of which four have been completed to date.^l While the scale of the GE BWR installed base is impressive in terms of megawatts (roughly 82.5 GW), the scope for learning through repetition of a standardized design has been historically quite narrow.

Of course, learning-by-doing is not limited to improvements in the ability of workers and firms to perform a production process more efficiently. It also encompasses improvements in the design of the product. For example, a reduction in the number of external recirculation loops from five to two was a major breakthrough in the design of the BWR-3 and a reason for the quick discontinuation of the BWR-2. David and Rothwell⁴² consider the question of how firms balance between the competing considerations of standardization and experimentation through diversity. On one extreme, consider repeated construction of identical plants, which permits learning only to occur in the efficiency of the manufacturing and construction process. On the other extreme, imagine iterated construction of one-of-a-kind plants. Such diversity provides fertile ground for experimentation and allows for the possibility of improvements to the design of future plants. However, it comes at the cost of workers and managers constantly readjusting to a new production process, as well as fixed development costs for each new design. Of course, between these two extremes exists a continuum of possibilities. The appropriate balance between experimentation and standardization is a problem of dynamic optimization under considerable uncertainty.

II.D. Megaproject Syndrome

An alternative hypothesis regarding learning in NPP design and construction is the view learning did indeed occur, but the cost-reducing and time-saving effects of learning were swamped by countervailing factors. Prime suspects for countervailing factors include the upward ratcheting of safety requirements,⁴³ regulatory delays in the granting of operating licenses,¹⁹ and diseconomies of scale.^9,38 One theoretical explanation for diseconomies of scale concerns the dispersal of decay heat after a reactor is shut down:

…[C]ore power (and decay power) is proportional to the volume of the core, which varies as the cube of the effective core radius. On the other hand, heat removal from the vessel is proportional to the vessel surface area, which varies roughly as the square of the core radius.³⁸

Thus, as reactors grew in size, ever more powerful and elaborate systems were needed to ensure the control of decay heat under emergency conditions.

However, if diseconomies of scale are present in nuclear reactors beyond a certain size, then it is puzzling why some firms in the industry continue to pursue even larger designs, such as the EPR (1650 MW) and the APR-1400 (1340 MW). Surely identifying optimal scale is part of the learning process. The promotion of small modular reactors (SMRs) and the proliferation of venture capital–backed firms pursuing SMR development implies a lack of consensus within the industry regarding what lessons should be learned from the scale-up of NPPs in the twentieth century.

A large academic literature on so-called “megaproject syndrome” theorizes that persistent economic problems in the construction of large-scale infrastructure is not merely a failure of technical optimization.^4,44–46 Nuclear power plants are but one category of megaprojects; examples of others include dams, airports, bridges, tunnels, harbors, public transit, and high-speed rail. The uniting characteristics of megaprojects include a budget above $1 billion (although some authors argue for lower thresholds in certain sectors or in the context of less developed nations); customization as necessitated by unique geographic conditions or customer requirements; extensive involvement of the public sector in matters such as planning, permitting, and financing; and complex management challenges arising from a large number of subcontractors.

Several theories have been considered in the literature regarding the high propensity of megaprojects to run over budget, fall behind schedule, be abandoned prior to completion, and fail to deliver the level of benefits promised once in operation. The classical view is that the incentive structure faced by politicians and project managers produce optimistically biased and/or strategically underestimated estimates of cost and schedule.¹⁶ Alternative views emphasize, inter alia, scope change,⁴⁷ corruption,⁴⁸ cross purposes and infighting among project partners,⁴⁹ and relations with external stakeholders (i.e., parties other than the project owner and the firms delivering the project).⁵⁰ I take the view that all of these theories are in no way mutually exclusive; in some cases, they could be mutually reinforcing. However, in this paper I focus on the role of external stakeholders—the local community, civil society organizations dedicated to the environment or advocacy for utility ratepayers, and enterprising politicians—in contributing to megaproject syndrome. I theorize that a higher degree of political decentralization enables external stakeholders to more substantively impact the design, permitting, and construction of megaprojects such as NPPs.

II.E. Decentralization

Decentralization has been in vogue as a development strategy promoted by major international institutions (e.g., the World Bank and International Monetary Fund) since the closing decades of the twentieth century; the recommendation has been increasingly accepted by a variety of countries.^51–53 The advice is motivated by a large and well-established literature that spans political economy, economic history, and development. Purported benefits of decentralization include greater public sector efficiency,⁵⁴ greater accountability,⁵⁵ lower corruption,⁵⁶ opportunities for yardstick competition,⁵⁷ and self-enforcing government commitment to markets.⁵⁸

At first glance, there may be limited applicability of the lessons from this literature to the case of nuclear power. Historically, national governments have assumed sole authority for the regulation of safety at NPPs, with the notable exception of West Germany (and reunited Germany post 1990), where authority is shared between the länder and the federal government. National control of nuclear safety regulation limits the scope of subnational regulation of the industry to policy areas such as land use, environmental permitting, and rate setting for regulated electric utilities. These are important aspects of the regulatory environment faced by firms in the nuclear industry, and they have a long history as the setting for political conflict over nuclear power,⁵⁹ as will be discussed further in Sec. II.F.

The consequences of what might be considered “inefficient regulation”—such as delaying or canceling the construction of NPPs and discouraging investment in the nuclear supply chain—are often intentional. The literature on decentralization primarily studies outcomes that are valence issues for voters—that is, issues on which all voters agree on the desired outcome, even if they may disagree on the optimal policy to achieve that outcome. Examples of valence issues include economic growth (faster is better), crime rates (lower is better), and corruption (lower is better). How does decentralization operate when the issue in question is a controversial technology over which opinions differ?

II.F. Local and Regional Opposition to NPP Siting

The politics of nuclear power has historically featured opposition by citizens, civil society, and politicians who are geographically near the site of proposed and existing NPPs. This has been documented in the United States,^59–61 France,⁶² West Germany,⁶³ the United Kingdom,⁶⁴ several separatist regions in Western Europe,⁶⁵ Japan,⁶² and even the Soviet Union in its final years.⁶⁶ Such opposition is often characterized by the acronym NIMBY (not in my backyard).^62,64 Some scholars view the term as inherently pejorative, conveying a normative disapproval of opponents’ position and motivations.⁶⁷ To avoid the appearance of passing an unnecessary normative judgment within the context of a positive analysis, hereafter I characterize the phenomenon as local and regional opposition to NPP siting, or “local opposition” for brevity.

The success of local opposition to NPPs has varied widely across nations, regions, and communities. A first-order explanation is to attribute siting outcomes to the magnitude and persistence of mobilization campaigns. In Site Fights: Divisive Facilities and Civil Society in Japan and the West, Aldrich⁶² provides a comparative history of local opposition to NPP siting in Japan and France. Meaningful contestation of pronuclear policy in the national halls of power was almost entirely absent in both countries in the late twentieth century. Furthermore, both Japan and France are unitary nations, meaning all sovereignty is vested in the national government. Thus, the ability of local and regional governments to conduct policy at cross purposes with the central government is necessarily circumscribed.

However, France and Japan contrast sharply with respect to actions taken by their central governments to ameliorate or overcome local opposition. Initiatives by the Japanese central government tended toward “soft social control”: propaganda, public meetings, offering tours of other NPPs, and most especially, generous transfer payments à la Coase to municipalities, fishermen, and farmers. France, by contrast, engaged in the methods of “hard social control,” such as police presence (and police violence), expropriation of land, surveillance, secrecy, restrictions on public participation, and simply ignoring local opinion. Aldrich argues that the difference in approaches resulted from the persistence of opposition in Japan and the withering away of opposition in France. In the face of persistent opposition, the state is obliged to “win hearts and minds.” Conversely, if opposition demobilizes after a proverbial “whiff of grapeshot,” the state sees no need to take another approach. Comparing the results of the French and Japanese nuclear programs, Aldrich writes,

Analysts point out that without only a few exceptions, “the government [of France] implemented its initial plans” for siting reactors (Rucht 1994, 153), an accomplishment far surpassing Japan’s record, where close to half the sitings failed.

While Japanese utilities regularly withdrew proposals in response to local opposition, it seems likely that they benefited considerably from only moving forward with construction in communities that had agreed to host NPPs. Once regulatory approval is granted and construction begins, the LT for constructing and commissioning an NPP in Japan has historically been extraordinarily fast and stable, averaging 4.7 years^m and showing modest declines from the 1970s to the 1990s. By comparison, the global average LT is 7.4 years. Construction in France was once faster than the global average as well, averaging 6.2 years for plants starting construction prior to 1980. That figure has trended upward, averaging 9.0 years for the plants beginning construction in 1980 or later, and it is certain to rise further with the eventual completion of Flamanville 3.

Circumstances in Japan and France contrast sharply with those in United States, where local opposition has historically been neither placated nor denied political and legal avenues by which to obstruct NPP construction. Cohen et al.⁶⁸ argue that a multiplicity of veto points in the constitutional design of the United States laid the groundwork for vigorous contestation of nuclear policy, including at the state and local levels. Emphasizing the federal nature of the United States, Joppke⁵⁹ points to three specific issues for which local opposition played an important role in delaying and canceling NPP construction:

The three predominant issues of the U.S. nuclear power controversy in the 1980s—emergency planning, utility rate regulation, and waste disposal—are all similar in this regard. In each case, local citizen groups formed effective alliances with local and state authorities in opposition to particular nuclear facilities or federal regulatory agencies.

Critical Masses: Opposition to Nuclear Power in California, 1958–1978, by Wellock,⁶⁰ is instructive of the causal mechanisms by which political decentralization would tend toward lengthening NPP LTs globally. For example, the Diablo Canyon Power Plant in California was the target of public protests throughout its construction period, drawing record-breaking crowds, celebrities, and Governor Jerry Brown. Seismic safety was among the activists’ leading concerns about the plant. State bureaucracies, such as the Natural Resources Agency, the State Lands Commission, and the Public Utilities Commission, offered ample opportunities for local opposition groups to intervene in the process, demand transparency from the utility, and force it to adjust its behavior. While construction of Diablo Canyon had begun in 1968 and was effectively completed in 1973, it was not permitted to enter commercial operation until 1985 after major seismic retrofits. While formally licensing decisions were in the hands of the federal bureaucracy, Wellock presents a strong case for the role of state government and local activists in pushing for stricter regulatory scrutiny.

A principal theme of Critical Masses is the emergence of a post-materialist environmentalist ethos. This ethos places little weight on economic concerns, distrusts technocrats and technocratic institutions, and emphasizes values such as local control, preserving the esthetic character of natural vistas, and opposition to war. Berndt and Aldrich⁶¹ report empirical evidence from the United States that proposed and under-construction NPPs were more likely to be abandoned in counties with higher incomes, which they consider to be a proxy measure of post-materialist values. On the other hand, Berndt and Aldrich find no relationship between local political affiliation and siting outcomes. They conjecture that ideological stances on environmental issues had not yet been mapped onto polarized partisan identities as they are in the present day.

Several authors have commented on the importance of a coherent, stable, long-term policy commitment to the nuclear industry in enabling its success. Delmas and Heiman⁶⁹ argue that fragmentation of power prevented the United States from making such a commitment. In case studies of China, India, South Korea, and Japan, Sovacool and Valentine^70,71 conclude that “centralization of national energy policymaking and planning” is one of six key factors for successful NPP deployments. They note, for example, that “in South Korea, the Office of Atomic Energy was placed directly under the President and the nuclear program was structured as a monopoly under the Korea Electric Power Corporation.” However, even South Korea—arguably the world leader in centralization, standardization, and successful learning-by-doing in the nuclear industry¹¹—offers a lesson in how decentralization can impede timely NPP construction:

Yonggwangⁿ was one of the first of the state-owned utility (Korea Electric Power Co –- KEPCO) projects to attract serious local opposition. Political reform in South Korea has devolved some power from the centre. Local politicians in Yonggwang used their new strength to slow down construction.

Hanjung (Korea Heavy Industries and Construction) was due to begin construction in December 1995, but a delay was brought on by the cancellation of construction permits for the site by Yonggwang County, South Cholla Province.⁷²

II.G. The Logic of Local Democratic Control

In this section, I draw on the framework of Mançur Olson’s seminal work, The Logic of Collective Action,¹⁵ to argue that the spatial distribution of costs and benefits from NPPs tends to generate a pattern of support by national governments and opposition by local and regional governments. The reasoning herein follows along the same lines as those in the introductory chapter of Site Fights.⁶²

The standard problem considered by Olson posits some policy provides concentrated benefits to a small group and diffuses costs to the rest of society. Lobbying the government to advocate for or against the policy requires overcoming a collective action problem, as no one individual can meaningfully influence the outcome. Olson argues that this situation inherently favors small groups for two reasons. First, the costs of overcoming collective action problems (such as building sufficient solidarity to overcome free-riding incentives and coordinating on a common strategy) are increasing in group size. Second, the benefits of a policy change can be quite large on a per-person basis for the sorts of small groups and policies typically considered.

To analyze the political economy of NPP construction, I modify Olson’s problem in three ways. First, I give a spatial dimension to group identity and interest: proximity to a proposed NPP. Those who live within the range of a hypothetical evacuation or exclusion zone in the event of a catastrophic nuclear accident are the small group; those who live farther away and yet would still benefit from the plant in some way are the rest of society.

Next, I invert the distribution of costs and benefits. The small group faces a geographically concentrated risk while the rest of society stands to gain geographically diffuse benefits. Of course, there is also a geographically concentrated benefit in the form of increased local economic activity. However, it is not unheard of for residents to regard this benefit as a cost. Local opponents of a proposed NPP near Bodega Bay, California, argued that a large industrial facility would ruin the rustic charm of their small fishing community by attracting further development (pp. 25–28 of Ref. 60).

The primary diffuse benefit is the electricity produced by the plant, which can be transmitted by the electricity grid to households and firms hundreds of miles away. The electricity may not be particularly valuable if substitute sources of electricity can be had at little, zero, or negative additional cost. However, other diffuse benefits include clean air and water,^o lessening of national dependence on expensive energy imports,^p complementarities with national nuclear weapons development,^q and interregional technological spillovers arising from learning-by-doing.

In a final modification of Olson’s original framework, I observe that democratic subnational government is a ready-made solution to the collective action problem faced by local residents who oppose a nearby NPP. Elected politicians are strongly incentivized to care about the interests of constituents in their jurisdiction and may take on the cause of opposing NPP construction as an electoral strategy. Even when the issue does not immediately arouse the attention of subnational politicians or those politicians favor the plant, the subnational government offers a more convenient forum with lower transaction costs in which local opponents of a nearby NPP can mobilize and seek to effectuate policy. A subnational government with sufficient autonomy and institutional capacity can directly intervene to regulate NPP construction on issues such as land use, environmental protection, or economic regulation of utilities without ever needing to lobby or influence the national government.

Of course, the reasoning here can be applied to a variety of political economy problems of a spatial nature, such as residential zoning, routing of high-speed rail lines, or the provision of services to the mentally ill and homeless. In the case of nuclear power, I propose it may explain partially patterns we see in the data on NPP LTs.

III. DATA

I assembled a database of all commercial nuclear power reactors that have ever initiated construction as of April 6, 2021. The observations are identified by the PRIS of the IAEA. While certain basic information about each NPP is available on the IAEA’s public website and through their various publications, I was granted temporary access to a private version of the PRIS database restricted to authorized users. The data set I have assembled offers considerably more detail and comprehensiveness than any other prior work on this topic, to my knowledge. Past studies are typically limited to variables such as size of the reactor in megawatts, general type of reactor (PWR, BWR, etc.), the identity of the firm responsible for the design of the nuclear steam supply system (NSSS), and a coarse coding of reactor models (e.g., 7). The most fine-grained coding of reactor models can be found in the data appendix to Portugal-Pereira et al.²² However, it suffers from the same inconsistencies present in the raw IAEA PRIS data. For example, American BWRs are coded as a concatenation of the design of the NSSS (BWR-1, BWR-2, etc.) and the design of the containment structure (Mark I, Mark II, Mark III), whereas BWRs in other countries are purely coded by their design of the NSSS. Similar inconsistencies abound for other types of reactors.

Unfortunately, the terms and conditions of my access to the PRIS prohibit me from sharing any of its data that are not otherwise publicly available. This primarily limits the sharing of data on the technical specifications of reactors.^r In the following, I will briefly describe the key variables central to the analysis and present certain summary statistics. A more detailed description of the data sources, cleaning, and variables can be found in Appendix A.

Table I tabulates the observations by country and geopolitical region. Countries were assigned to regions based on a constellation of factors, primarily their alliances, form of government, and economic system during the Cold War. A detailed discussion of the coding scheme is reserved to Sec. A.XII in Appendix A.

III.A. Lead Time

Lead time is computed as the time between the date on which construction began—the first day on which concrete for the foundation was poured—and the date of commercial operation, less any time during which construction was totally suspended.

The mean LT in the data set is 7.4 years, with a standard deviation of 3.3 years. There are clear geographic patterns to the data, as summarized in Table II and plotted in Fig. 3. Notably, East Asian nations construct NPPs significantly more quickly and consistently, with a mean of 5.5 years and a standard deviation of 1.4 years. The mean LT in Western nations does not differ substantially from the global mean, which is perhaps unsurprising given that NPPs in Western nations account for 55% of the sample.

Fig. 3. Divergent regional trends in NPP LT.

III.B. Reactor Typology

The PRIS uses the term “type” to encapsulate broad similarities in the principles of a reactor’s design. The most common types are PWRs, BWRs, pressurized heavy water reactors (PHWRs), gas-cooled reactors (GCRs), and light water graphite reactors (LWGRs). I aggregated all other types into a single category called “other” due to a sparsity of observations.

Summary statistics by type of reactor are presented in Table III. Light water graphite reactors, which were exclusively built in the Soviet Union, exhibit the quickest average LT, as well as the lowest standard deviation. In a close second place are BWRs, which are largely found in the Western Bloc and Japan. Pressurized water reactors are exactly at the global average, which is unsurprising given that they account for 56.3% of the global population. Pressurized heavy water reactors perform relatively poorly, although this average is heavily influenced by three countries: Argentina and Romania (which suspended construction on theirs for many years due to economic and political conditions) and India (whose nuclear power program developed with little international support as a consequence of the international response to India’s acquisition of nuclear weapons). Excluding these three countries, which account for roughly 40% of PHWR observations, the LT of the remaining PHWRs is 6.7 years.

I use the term “family” to classify reactors that share an evolutionary heritage. This classification is narrower than reactor type in that a family encompasses only reactors by a single firm or a small set of firms that have a history of licensing intellectual property and collaborating with one another. The classification scheme is detailed in Sec. A.X of Appendix A.

The most granular typology of reactor is the model. Where applicable, I use the model names assigned by the manufacturer, such as AP-1000, CP1, P4, OPR-1000, CNP-300, VVER-213, and ABWR. For standardized reactor designs, this identification comes as close as realistically possible to identifying “identical” reactors. However, for nonstandardized designs, the PRIS provides an abbreviated, generalized description of the reactor’s design in place of a model name. For example, “WH 4LP (DRYAMB)” indicates that the reactor is a Westinghouse design with four primary coolant loops and the containment structure operates at ambient atmospheric pressure. Information about the containment design was divorced from the name of the reactor model and used to populate a separate categorical variable relating to containment.

III.C. National Political and Economic Characteristics

Decentralization is my independent variable of interest. As my primary measure, I adopt the “self-rule” subindex from the Regional Authority Index (RAI) by Hooghe et al.⁷⁴ They evaluate the constitutions and political histories of individual countries, and they systematically score them on matters such as the role of subnational governments in approving constitutional change, whether the central government holds a veto over subnational decisions, and the autonomy of subnational jurisdictions in setting their tax base and rates. For robustness, I also test my hypotheses against a binary indicator of whether a country has a federalist or unitary constitution. Table IV reports the descriptive statistics for decentralization for both measures.

A t-test of the difference in mean LT between federalist and unitary countries rejects the null hypothesis of no difference, finding federalist nations take 18 months longer on average ($t=-2.12$).^s Lead time correlates with the continuous measure of decentralization at $r=0.167$.

Because decentralization may correlate with other important country characteristics, I also include measures of GDP per capita, democracy, and regime change. I rely on the historical estimates of GDP per capita from the Maddison Project.⁷⁵ For democracy, I use the “Polyarchy” index of electoral democracy generated from the Varieties of Democracy (V-Dem) Project.⁷⁶ To identify the dates and magnitudes of changes of a country’s constitutional structure or regime type, I rely on data from Polity IV (Ref. 77). I assign a value of 1 to a reactor if it was under construction (or in a period of temporarily suspended construction) during an episode of major regime change, and zero otherwise.

IV. ECONOMETRIC SPECIFICATIONS

Appendix B discusses an assortment of econometric issues that are common to many or all of the specifications that follow. Here, I summarize its conclusions briefly. In Sec. B.I, I argue that political institutions (democracy, decentralization, and regime change) are exogenous to NPP construction. In Sec. B.II, I account for a special type of measurement error that arises from serial construction. In Sec. B.III, I investigate possible selection bias arising from abandoned construction and conclude that it is negligible. In Sec. B.IV, I explain how I control for the effect of major nuclear accidents and political events on LT using an instrumental variables strategy. In Sec. B.V, I define cumulative experience as the count of reactors of the same family as reactor $i$ that began construction prior to reactor $i$. Table B.IV lists all symbols used in the equations for this section.

For lack of quantitative measures of cross-nationally comparable, site-specific local opposition, the hypotheses tested in this paper assume the presence of local opposition. Given the literature I reviewed in Sec. II.F documenting the presence of local opposition in both unitary and federalist nations, I argue that this is a reasonable, albeit imperfect, assumption. My analysis focuses on identifying the channels through which political decentralization operates. While the credibility of the analysis qua causal inference is limited, the results can help guide future research by narrowing the range of likely explanations for the raw correlation between decentralization and NPP LT.

IV.A. Mechanism 1: Politically Constrained Design

While summary statistics show that NPP LTs tend to be longer in federalist nations than in unitary nations, we must ask whether they build comparable NPPs. Federalist and unitary nations may systematically choose different designs of reactors that have differing technical, safety, and economic characteristics. Do LTs differ because of these differences in design, or is it because of factors beyond the design of the plant? To test this hypothesis, I conduct the analysis in two steps.

First, I investigate which design characteristics have meaningful impacts on LT in a regression with country fixed effects. The country fixed effects are intended to generate credible estimates of the average treatment effects of design characteristics on LT by leveraging only within-country variation in design characteristics. The econometric specification is as follows:

(3) $ln(\mathrm{L}{\mathrm{T}}_i)=\sum _{s\in S}{\theta }_{s}Spe{c}_{s,i}+{\delta }_t+\gamma M_i+{\mu }_c+{\nu }_y+{\epsilon }_i.$(3)

The selection of $Spe{c}_{s,i}$ variables was guided by five-fold cross validation.^t An additional consideration was sample size; including design characteristics for which too many observations have missing values would limit the sample size of the subsequent analyses. Refer to Table V for the list of variables ultimately included. ${\delta }_t$ represents fixed effects by type of reactor (e.g., BWR, PWR).

The year fixed effects ${\nu }_y$ control for any number of time-related variables that might otherwise be spuriously correlated with the regressors. For example, gas-cooled, graphite-moderated reactors have fallen out of favor in the two countries that have historically built them in meaningful numbers: the United Kingdom and France. Without year fixed effects, the estimated coefficient for this type of reactor—which both nations eventually judged to be technically and economically inferior to PWRs—could be biased downward due to most of these reactors having been built prior to the emergence of mass movements against and stricter regulation of nuclear power. Sources of longitudinal variation are not explicitly modeled because (1) they do not relate to the hypothesis being tested and (2) there is sufficient within-year dispersion in design characteristics to generate well-powered estimates.

In the second step of the analysis, I generate the predicted values of a reactor’s LT conditional on its design characteristics, type of reactor, and $M_i$ while omitting the country and year fixed effects. This represents a measure of a reactor’s expected LT in a hypothetical “average country” and “average year” conditional on its design characteristics.

I regress these expected values of LT on country-level characteristics. Past research has found that nations with higher GDP per capita tend to complete their NPPs faster ceteris paribus.⁷ In light of the strong correlation of GDP per capita with political institutions,⁷⁹ I control for the natural log of GDP per capita in order to avoid any possible spurious correlation between level of economic development and form of government. I estimate the following equation by ordinary least squares:

(4) $ln(\widehat{\mathrm{L}{\mathrm{T}}_i})={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}+{\epsilon }_i.$(4)

This regression tests whether economic development and political institutions are associated with choices in the design of NPPs that entail longer or shorter LTs.

IV.B. Mechanism 2: Regulatory Delays

I hypothesize that political decentralization generates conditions that cause construction to be temporarily halted or to proceed more slowly than would otherwise occur. This hypothesis proposes that, on average, otherwise identical reactors built in politically decentralized nations will tend to take longer to build than those in politically centralized nations, holding all else constant. The difficulty is in credibly identifying “otherwise identical reactors.”

To begin, I include fixed effects for the model of reactor. I argue that this is a sufficient control for reactors that are of a standardized design ($n=311$), which share a common designation supplied by the lead designer of the NSSS. Eight pairs of reactors built as twins at the same site are classified with a unique model name, although they are not classified as standardized because they were never replicated elsewhere. As a general rule, twin reactors at the same site are identical. This group presents no econometric concern but offers no cross-country variation to exploit. Reactors of models that were only built once ($n=50$) are automatically dropped by the estimation procedure due to misleading causal inference that arises from singleton fixed effects.⁸⁰

The more challenging case is that of nonstandardized “models” of reactors that have been built in more than one country ($n=212$). To account for the technically differing features of nonstandardized models that may cause them to have shorter or longer LTs, I control for the predicted LT (conditional on design characteristics) that was generated in step 1 of the procedure outlined in Sec. IV.A. This approach maintains the parsimony of the econometric specification, as contrasted with controlling for several dozen design characteristics. Furthermore, while the design characteristics data cannot be publicly released due to IAEA data sharing restrictions, no such restriction applies to the predicted values of LT I generated from them by Eq. (3)(3) $ln(\mathrm{L}{\mathrm{T}}_i)=\sum _{s\in S}{\theta }_{s}Spe{c}_{s,i}+{\delta }_t+\gamma M_i+{\mu }_c+{\nu }_y+{\epsilon }_i.$(3) . Thus, the data necessary to replicate this analysis have been made available.

I omit country fixed effects for two reasons. First, within-country, over-time variation in decentralization is exceedingly limited when considering how few countries have built NPPs entirely before and entirely after major changes in their political institutions. Second, the cross-national variation in decentralization is of greater interest, as cross-national differences in NPP LT is the primary puzzle. Since the treatment of interest—political decentralization—is more or less assigned by country rather than by reactor, the standard errors are clustered by country.

I do not include year fixed effects. Instead, I explicitly model the major events that are widely believed to have caused lengthy delays, per the instrumental variables methodology described in Sec. B.IV of Appendix B. The controls for these events take the form of binary indicator variables that indicate whether a reactor was under construction during a given event. I further allow a separate coefficient for the nations in which the accident occurred, namely, the United States in the case of TMI and the Soviet Union in the case of Chernobyl.^u

Last, I control for whether the reactor was built for an investor-owned utility (IOU) or a publicly owned utility.^v Several possible hypotheses may point toward one form of ownership structure favoring faster or slower construction given the differing economic incentives, regulatory treatment, and cost of capital associated with each business model. My preferred hypothesis is that, given the higher cost of capital for IOUs, I expect that IOUs generally complete construction faster.

The econometric model is given by

(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5)

IV.C. Mechanism 3: Megaproject Syndrome

Megaprojects such as NPPs have a natural tendency toward schedule slippage. I theorize that political decentralization exacerbates megaproject syndrome by initiating more instances of scope change mid construction and increasing the number of external stakeholders who may intervene in the project.

To quantitatively measure such an effect, I take as a measure of complexity and scale the variable $ln(\widehat{L{T}_i})$ generated from step 1 of the analysis in Sec. IV.A. This variable primarily reflects the size of the reactor in megawatts, but it also incorporates several other specifications and design choices that are associated with longer or shorter LTs, such as whether the reactor is of a standardized design. I hypothesize that, if decentralization exacerbates megaproject syndrome, then the penalty to LT arising from a higher degree of “megaproject-iness” should be stronger in decentralized nations. I model this with an interaction between $ln(\widehat{L{T}_i})$ and decentralization.

I build the econometric specification as follows. I include country fixed effects, as there is sufficient within-country dispersion in $ln(\widehat{L{T}_i})$ to generate well-powered estimates. These fixed effects control for differing national characteristics; cross-national differences in the level of LT are not of interest for this hypothesis. Next, I include year fixed effects as there is sufficient dispersion within years to generate well-powered estimates. This removes any global time trends in LT.

However, two-way fixed effects cannot account for the possibility that time trends differ by country for reasons unrelated to the interaction of $ln(\widehat{L{T}_i})$ and decentralization. While fixed effects by country-year would be ideal, the number of degrees of freedom would greatly diminish with the introduction of so many fixed effects. Furthermore, in 155 cases, there were no other reactors that began construction in the same country in the same year, so there is no dispersion in size within those country-year pairs. As a next-best control for the possibility of differential trends by country, I include instrumented indicator variables for events that likely had a disproportionate effect on a particular country (TMI in the United States, Chernobyl in the Soviet Union)^w or which occurred in different countries at different points in time (regime change). I also control for GDP per capita, which exhibits differing trends over time across countries.

I do not control for any design characteristics or measurement error $M_i$, as these variables are embedded in the value of $ln(\widehat{L{T}_i})$. I do control for whether an IOU or publicly owned utility is building the reactor, for the same reasons as in Sec. IV.B. I test several specifications, so the equation that follows is of a generalized nature, allowing for several specifications of $\beta$:

(6) $\begin{array}{rl}& ln(L{T}_i)=\beta ln(\widehat{L{T}_i})+{\gamma }_{1}ln(GDPp{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\delta }_f+{\mu }_c\\ & +{\nu }_y+{\epsilon }_i.& (6)\end{array}$(6)

In the first specification, $\beta$ is simply a constant that estimates the global average relationship between “megaproject-iness” and LT. In the next specification, I allow $\beta$ to vary as a linear combination of a nation’s democracy and decentralization. While my hypothesis concerns decentralization, the intensity of megaproject syndrome could just as well vary with the level of democracy as with decentralization. Therefore, I include both variables in estimating $\beta$. In the final specification, I estimate separate values of $\beta$ by geopolitical region, as defined in Table I.

IV.D. Modeling Mechanism 4: Resetting the Learning Curve

I theorize that political decentralization inhibits learning-by-doing through regulatory instability, jurisdictional diversity, and electricity market fragmentation. These factors oblige firms to abandon gains from proceeding down an established learning curve and begin exploring the learning curve of a more novel design. To estimate this effect empirically, I propose an econometric specification that allows the learning rate to vary according to the degree of decentralization of a country’s political institutions.

I operationalize cumulative experience as the inverse hyperbolic sine transformation of the count of reactors of the same family as reactor $i$ that began construction prior to reactor $i$. Further details regarding the measurement of cumulative experience are available in Sec. B.V of Appendix B.

I distinguish between two possible dimensions along which experience may matter. The first is the “within” dimension: the effect of cumulative experience on LT that results from continuing to build more reactors within the same family. The second is the “between” dimension: the effect of cumulative experience on LT that results when choosing between families of reactors with differing levels of cumulative experience.

I argue that the between dimension contains information regarding “learning-by-searching,”⁸¹ as opposed to learning-by-doing. When utilities are deciding between different designs of NPPs to build, they face choices ranging from experimental reactor designs of uncertain future potential to reactors from families with an established track record and large experience base to draw from. The more established design should, in expectation, present fewer challenges in the construction process, even if the less-experienced design has a greater, long-term techno-economic potential.⁸² In settings with weak, inefficient, or impeded learning-by-searching, the benefits to adopting a more established design should be less evident.

In both cases, the methods herein do not generate strong causal inference. They should be understood as descriptive partial associations between cumulative experience and LT, holding constant several other factors that might otherwise explain the correlation between experience and LT. In particular, because cumulative experience is endogenous—families that are inherently better for techno-economic reasons are liable to gain more experience—estimation along the between dimension is especially suspect. Improving causal inference is an opportunity for future research.

For the econometric specification to capture “within family” learning, I naturally include fixed effects by reactor family. This means the econometric model assumes there are constant, unexplained differences between the level of LT across different reactor families. Next, I include country fixed effects.^x Political factors may cause differences in the average level of LT across countries; these differences are investigated with the methods of Sec. IV.B but are not of interest here.

These sets of fixed effects combine to form an econometric specification in which the only remaining variation to be explained is changes in LT over time, within families of reactors, controlling for cross-national average differences in the level of LT. Fixed effects by year of construction start would sap the model of nearly all remaining variation. Instead, I control for the major events affecting the nuclear industry per the instrumental variables strategy laid out in Sec. B.IV in Appendix B.

In general, I do not control for design specifications in these regressions because an important component of learning-by-doing is using the information gained to redesign the product better next time. Holding a design constant would limit the estimated learning effects to only learning arising from repetition of identical or nearly similar designs. That said, I make three exceptions in controlling for the following design specifications:

First, I control for rated power output in megawatts. The trend toward increasingly large reactors over time is unambiguous; furthermore, size tends to be correlated with a reactor family’s cumulative experience. In a regression of size in megawatts on cumulative experience with year fixed effects (i.e., removing any time trends and only looking at cross-sectional variation), I find that a doubling of cumulative experience is associated with a 74-MW increase in the size of a reactor ($t=10.1$). In other words, new concepts for reactors are implemented at a small scale first and then gradually scaled up as experience accumulates.

Given the economic costs of long LTs, I must conclude that NPP designers are not deliberately choosing larger capacities for the sake of longer LTs. Instead, they are reportedly choosing larger capacities in order to reduce the OCC. Berthélemy and Escobar Rangel¹⁰ find a strong negative association between size and OCC when controlling for LT; in unreported regressions, I replicate that finding with the larger sample provided by Portugal-Pereira et al.²² However, given the likely causal effect of LT on the OCC and the certain effect of LT on financing costs, this strategy of ever-increasing scale may not be wise.

Two additional design characteristics I control for are whether the reactor was used for the coproduction of electricity and plutonium and the cooling technology. I argue that reactors which coproduced electricity and plutonium exhibit exceptionally fast LTs because they were built in haste for military purposes during the Cold War. Regarding cooling technology, the use of once-through cooling (OTC) or some other method to discharge waste heat is determined by environmental conditions and environmental regulations. Cooling towers are not unique to the nuclear industry.

As in Secs. IV.B and IV.C, I control for whether the reactor is being built for an IOU or publicly owned utility for the same reasons described there. Given the substantial within-country, over-time variation in GDP per capita, I control for it. Conversely, there is very little within-country, over-time variation in the level of democracy and decentralization in my sample, so I do not control for those. To summarize, the econometric specification for “within family” learning-by-doing is given by

(7) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_{1}M{W}_i+{\theta }_{2}1\{OT{C}_i\}\\ & +{\theta }_{3}1\{\mathrm{P}{\mathrm{u}}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+{\gamma }_3{M}_i+{\delta }_f+{\mu }_c+{\epsilon }_i.& (7)\end{array}$(7)

Similar to the approach in Sec. IV.C, I allow $\beta$ to vary across countries according to a linear combination of its political characteristics. I standardize $ln(GDPp{c}_{c,y})$, $De{m}_{c,y}$, and $De{m}_{c,y}$ such that they are centered on their global average values and scaled by their global standard deviations.^y

To estimate the effects of cumulative experience when comparing between different families of reactors, the first step is to omit the family fixed effects. I retain the country fixed effects from before, as there is plenty of within-country dispersion in cumulative experience to work with. This time, I add year fixed effects, so that the comparison is between different families of reactors with differing levels of experience at the same point in time. Year fixed effects render unnecessary most of the controls for major events affecting the nuclear industry, except those representing effects specific to one country, for the same reasons as in Sec. IV.C.

As with the “within family” estimation, I control for capacity in megawatts, plutonium coproduction, investor ownership, and GDP per capita. The econometric specification for “between family” learning-by-searching is given by

(8) $\begin{array}{rl}& ln(L{T}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_1\mathrm{M}{\mathrm{W}}_i+{\theta }_{2}1\{\mathrm{O}\mathrm{T}{\mathrm{C}}_i\}\\ & +{\theta }_{3}1\{P{u}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_3{M}_i+{\mu }_c+{\nu }_y+{\epsilon }_i.& (8)\end{array}$(8)

V. RESULTS

V.A. Results for Mechanism 1: Politically Constrained Design

The estimation results of Eq. (3(3) $ln(\mathrm{L}{\mathrm{T}}_i)=\sum _{s\in S}{\theta }_{s}Spe{c}_{s,i}+{\delta }_t+\gamma M_i+{\mu }_c+{\nu }_y+{\epsilon }_i.$(3) )—the regression of LT on design specifications—is presented in Table V. I will define the variables and interpret the coefficients for the benefit of readers not familiar with the technical terms related to nuclear reactor design.

I will contextualize and interpret some of the results in Table V for the benefit of readers not familiar with the technical terms related to nuclear reactor design.

The relationship between power output and LT was ascertained to be best approximated as log-linear through rigorous testing of alternative specifications. The estimated effect size implies that a typical 1-GW reactor (which is the approximate order of magnitude of nearly all reactors being built today) would take about 69% longer to build than a hypothetical 50-MW SMR. For reference, the global average LT for gigawatt-scale reactors is around 86 months, so on the basis of the scaling factor alone, the estimated LT of the SMR would be 51 months. Further applying the bonus to standardized designs brings the estimate to around 45.6 months.

Reactor outlet temperature refers to the temperature of the primary coolant upon exit from the reactor.^z The median reactor outlet temperature is 328°C. The minimum observed value is 220°C; the maximum of 950^°C. Hotter outlet temperatures enable greater thermal efficiency in the conversion of steam to electricity, but they also present greater safety challenges.

The number of primary coolant loops ^aa does not vary quite so widely, being typically between two and four, with a maximum observed value of eight. The results suggest that more loops represent more complexity to deal with in construction, but the effect sizes are not statistically significant. That said, in unreported regressions I find there is a clear trend toward a reduction in the number of these loops, which is consistent with the idea of the industry trying to streamline design.

Once-through cooling refers to the practice of discharging waste heat directly into a nearby body of water. This obviates the need for cooling towers or other structures designed to dissipate waste heat into the atmosphere; it also enhances the efficiency of the conversion of heat to electricity. There is a sizable reduction in LT associated with OTC, estimated to be 13.2%. In unreported regressions, no statistical difference was found when comparing natural draft to forced draft cooling towers; both add nearly the same amount to the construction schedule relative to OTC.

I find that reactors of standardized designs finish construction and commissioning about 10% faster than their custom-built peers, on average. This suggests that custom ordering an NPP is generally a mistake, except perhaps for experimental purposes. To explore this further, I ran a logistic regression to understand the determinants of reactor standardization, the results of which are presented in Table VI. At first, it appears that decentralized nations and IOUs are less likely to adopt standardized designs (as indicated by odds ratios less than one). However, column (3) reveals that this finding is probably an artifact of the fragmentation of the electricity sector in such countries, and not necessarily related to political conditions. I theorize that countries with more utilities fail to coordinate on a standardized design.

The predicted values of LT generated after estimating Eq. (3)(3) $ln(\mathrm{L}{\mathrm{T}}_i)=\sum _{s\in S}{\theta }_{s}Spe{c}_{s,i}+{\delta }_t+\gamma M_i+{\mu }_c+{\nu }_y+{\epsilon }_i.$(3) are displayed in Fig. 4, where they are graphed against the observed values of LT. The estimated slope coefficient is 0.836 (standard error = 0.072, N = 590). The $R^2$ of this bivariate regression—30.2%—implies that observed design characteristics only account for a modest fraction of the global dispersion in LT. Furthermore, I find predicted LT does not meaningfully trend upward over time.^ab This is suggestive evidence for the view that the escalation over time in LT cannot be solely attributed to changes in design, whether arising from regulation or industry mismanagement. However, further research should revisit this question when more data concerning safety features can be collected.

Fig. 4. Goodness of fit of Eq. (3).

Table VII displays the results of the second stage of the analysis, wherein the fitted values of LT (as predicted solely by design characteristics) are regressed on national characteristics. The only national characteristic meaningfully correlated with design-related LT is GDP per capita, and this correlation becomes statistically and substantively insignificant when controlling for the tendency of richer countries to build larger reactors. This is strong evidence against the hypothesis of “politically constrained design.” That is to say, there is no evidence that democratic or decentralized nations exhibit longer LT in NPP construction on account of differences in the design of the plants as compared to those in undemocratic or centralized nations.

V.B. Results for Mechanism 2: Regulatory Delays

In Table VIII, I report the results of Eq. (5)(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) . Columns (1) and (2) present estimation results as originally specified in Eq. (5)(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) . The coefficients on the two measures of decentralization are statistically insignificant; that is, they fail to reject the null hypothesis of no relationship between decentralization and LT, conditional on reactor model and other controls.

The null results arising from the estimation of both Eqs. (4)(4) $ln(\widehat{\mathrm{L}{\mathrm{T}}_i})={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}+{\epsilon }_i.$(4) and (5(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) ), at first glance, may appear to rule out two mutually exclusive and exhaustive channels by which political decentralization might correlate with NPP LT. Either decentralization correlates with NPP LT on account of differences in design, or decentralization correlates with NPP LT holding design constant, or there is no correlation. However, as discussed in Sec. III.C, there is, in fact, a meaningfully large and statistically significant bivariate correlation.

Hence, the question arises of “where did the correlation go?” In columns (3) and (4) of Table VIII, I exclude the fixed effects by reactor model originally specified for Eq. (5)(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) . The coefficients on decentralization become empirically large and highly statistically significant. From this, I conclude that the correlation was absorbed by the reactor model fixed effects of columns (1) and (2). There are two possible interpretations of these findings.

One possibility is that reactor model fixed effects behave approximately like country fixed effects, on account of the fact that certain models historically were only built in one country or a handful of geopolitically allied nations. This would tend to sap the model of statistical power by limiting the cross-national comparison to relatively few cases where identical models were built in countries with large differences in their level of political decentralization.

The second possibility is that there are technical specifications that are unobserved in my data set yet are relevant to the hypothesis of politically constrained design. If this possibility is true, then it would mean that the null results in Table VII are false negatives. Reactor model fixed effects are “black boxes” that collect and hide the effect of any omitted technical specifications which have a bearing on NPP LT.

In short, with the present data it is impossible to distinguish between these two hypotheses—politically constrained design and regulatory delays—in ascertaining the channel by which decentralization correlates with NPP LT. That said, we can reject the observed design characteristics listed in Table B.III as explaining the correlation. Furthermore, the results in columns (3) and (4) present some reassurance that the raw correlation is not patently spurious, given the controls for nuclear accidents, regime change, GDP per capita, democracy, investor ownership, and the expected LT conditional on observed design characteristics.

The interpretation of the effect sizes in columns (3) and (4) are as follows. A federalist constitutional design is associated with 22.6% longer LT, relative to unitary constitutions. The range of globally observed values in the continuous measure of decentralization (RAI) spans approximately three standard deviations, so the comparable effect size from column (4) would be on the order of a 30% increase. A three-standard-deviation increase in decentralization is equivalent to the difference between federalism in the United States and the centralism in Finland present in the 1970s.^ac

Table VIII also present results of the estimated effect on LT arising from various events that impacted the politics and regulation of the nuclear industry. As these are not central to the present work, I will not discuss them at length. However, I will note a few issues that likely undermine the accuracy of the estimates. First, the finding that the Chernobyl disaster supposedly accelerated NPP construction in the USSR is almost surely spurious. Being under construction during the Chernobyl disaster in the USSR is correlated with being under construction during regime change, as the USSR dissolved a few years later. Per the results in column (3), the combined effect of both events is roughly a 32%^ad increase in LT. Soviet NPPs that were under construction on April 26, 1986 but still finished construction prior to 1991 represent those which began their construction relatively earlier [recall that Eq. (5)(5) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)={\beta }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})+{\beta }_{2}De{m}_{c,y}+{\beta }_{3}De{c}_{c,y}\\ & +\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_{1}ln(\widehat{\mathrm{L}{\mathrm{T}}_i})\\ & +{\delta }_m+{\epsilon }_i.& (5)\end{array}$(5) includes no year fixed effects] and therefore were closer to completing construction sooner, thereby avoiding the upheaval of the 1990s.

Regarding the Fukushima Daiichi disaster, only reactors that have commenced commercial operation as of April 6, 2021 are included in the sample, so the coefficient is necessarily biased downward by the exclusion of as-of-yet incomplete reactors.

V.C. Results for Mechanism 3: Megaproject Syndrome

Table IX displays the results of regressions which test the hypothesis that the impact of scale and complexity of design on NPP LT is mediated by political decentralization. Column (1) reports the finding that, globally on average, the correspondence between a reactor’s predicted LT and its actual LT is fairly close to a one-to-one relationship.^ae In columns (2), (3), and (4), I allow the parameter governing this relationship to vary according to national political characteristics. Note that the $De{m}_{c,y}$ and $De{c}_{c,y}$ variables are standardized according to their z-values, so they are centered on the global averages. Thus, the uninteracted coefficient on $\widehat{L{T}_i}$ can be interpreted as the marginal effect in a country with average values of both variables.^af

TABLE IX Estimation Results of Eq. (6(6) $\begin{array}{rl}& ln(L{T}_i)=\beta ln(\widehat{L{T}_i})+{\gamma }_{1}ln(GDPp{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\delta }_f+{\mu }_c\\ & +{\nu }_y+{\epsilon }_i.& (6)\end{array}$(6) )*

	(1)	(2)	(3)	(4)
Expected LT, conditional on design	0.91%	0.88%	0.85%
1% increase in $\widehat{L{T}_i}$	(13.04)	(9.60)	(11.03)
Expected LT $\times$ Democracy		−0.01	−0.00
… $\times$ one standard deviation increase in $De{m}_{c,y}$		(1.00)	(−0.06)
Expected LT $\times$ Decentralization		0.03%
… $\times$ 1 if country has a federal constitution		(0.26)
… $\times$ one standard deviation increase in $De{c}_{c,y}$			0.01%
			(10.1)
Expected LT $\times$ East Asia				0.49%
1% increase in $\widehat{L{T}_i}$ in East Asia				(4.15)
Expected LT $\times$ Western Bloc				0.97%
1% increase in $\widehat{L{T}_i}$ in the Western Bloc				(11.13)
Expected LT $\times$ Eastern Bloc				1.06%
1% increase in $\widehat{L{T}_i}$ in the Eastern Bloc				(4.03)
Expected LT $\times$ Global South				1.70%
1% increase in $\widehat{L{T}_i}$ in the Global South				(3.90)
GDP per capita	−15.7%	−19.7%	−15.1%	−10.3%
One standard deviation increase in $ln(GDPp{c}_{c,y})$	(−0.95)	(−1.14)	(−0.92)	(−0.63)
Investor-owned utility	−6.5%	−6.5%	−4.3%	−6.3%
$IO{U}_i$	(−1.14)	(−1.15)	(−0.71)	(−1.13)
Reactor family fixed effects	✓	✓	✓	✓
Country fixed effects	✓	✓	✓	✓
Year fixed effects	✓	✓	✓	✓
Nuclear accidents and regime change	✓	✓	✓	✓
Observations	571	571	483	571
Adjusted $R^2$, within effects	0.344	0.348	0.311	0.362

*Dependent variable: $ln(L{T}_i)$; transformed marginal effects on $L{T}_i$ in percent and t-statistics in parentheses.

Columns (2) and (3) concur in rejecting the hypothesis that national political characteristics contribute to megaproject syndrome by sharpening the penalty to LT that results from NPPs of larger and more complex designs. The results are both statistically insignificant and not of an economically substantive magnitude.

Column (4) allows for the relationship between $\widehat{L{T}_i}$ and $L{T}_i$ to vary according to geopolitical regions, which are defined and justified in Sec. A.XII in Appendix A. Some interesting patterns emerge, although I will note that the only coefficient that statistically differs from one^ag is the coefficient for East Asia ($t=4.29$). It appears that East Asian countries—mainland China, Taiwan, South Korea, and Japan—are exceptionally competent at managing large and complex NPP construction projects, much more so than the rest of the world. This may have something to do with their highly unitary political regimes, although surely that is not the only factor at play. It should not go without mention that all three of the Western-aligned East Asian nations began their NPP programs under eras of weak or absent democratic institutions involving rule by military dictatorships (e.g., the regime of Park Chung-hee in South Korea) or a single political party (KMT in Taiwan, LDP in Japan). The increasing political contestation of nuclear power policy in these now firmly democratic nations might dismantle the conditions that made their earlier NPP deployments so successful. However, if democracy matters, the present methods are not sufficient to identify any such effect.

V.D. Results for Mechanism 4: Resetting the Learning Curve

Table X displays the results of regressions estimating the effect of cumulative experience on LT. Columns (1) and (2) report the raw parameters that form a linear combination in estimating $\beta$ in Eq. (7)(7) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_{1}M{W}_i+{\theta }_{2}1\{OT{C}_i\}\\ & +{\theta }_{3}1\{\mathrm{P}{\mathrm{u}}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+{\gamma }_3{M}_i+{\delta }_f+{\mu }_c+{\epsilon }_i.& (7)\end{array}$(7) . Recall that Eq. (7) is designed to capture learning-by-doing within reactor families by examining trends in LT, holding constant any level effect of the reactor family. Similarly, columns (3) and (4) report the raw parameters that form a linear combination in estimating $\beta$ in Eq. (8)(8) $\begin{array}{rl}& ln(L{T}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_1\mathrm{M}{\mathrm{W}}_i+{\theta }_{2}1\{\mathrm{O}\mathrm{T}{\mathrm{C}}_i\}\\ & +{\theta }_{3}1\{P{u}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_3{M}_i+{\mu }_c+{\nu }_y+{\epsilon }_i.& (8)\end{array}$(8) . Recall that Eq. (8) is designed to capture learning-by-searching across reactor families by comparing how reactor families with more experience fare in terms of LT relative to those with less experience, compared at the same years in history.

TABLE X Learning Parameters Estimated per Eqs. (7)(7) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_{1}M{W}_i+{\theta }_{2}1\{OT{C}_i\}\\ & +{\theta }_{3}1\{\mathrm{P}{\mathrm{u}}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+{\gamma }_3{M}_i+{\delta }_f+{\mu }_c+{\epsilon }_i.& (7)\end{array}$(7) and (8(8) $\begin{array}{rl}& ln(L{T}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_1\mathrm{M}{\mathrm{W}}_i+{\theta }_{2}1\{\mathrm{O}\mathrm{T}{\mathrm{C}}_i\}\\ & +{\theta }_{3}1\{P{u}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_3{M}_i+{\mu }_c+{\nu }_y+{\epsilon }_i.& (8)\end{array}$(8) )*

	(1)	(2)	(3)	(4)
Cumulative experience	0.006	−0.018	−0.073	−0.110
$sin{h}^{-1}(Ex{p}_{i,f})$	(0.19)	(−0.58)	(−3.37)	(−5.80)
Cumulative Experience $\times$ Democracy	0.015	−0.000	0.029	0.009
… $\times$ one standard deviation increase in $De{m}_{c,y}$	(1.77)	(−0.05)	(2.98)	(0.76)
Cumulative Experience $\times$ Decentralization	−0.003		−0.048
… $\times$ 1 if country has a federal constitution	(−0.08)		(−1.75)
… $\times$ one standard deviation increase in $De{c}_{c,y}$		0.014		0.019
		(1.38)		(1.74)
Reactor family fixed effects	✓	✓
Year fixed effects			✓	✓
Country fixed effects	✓	✓	✓	✓
Additional controls per Eqs. (7)(7) $\begin{array}{rl}& ln(\mathrm{L}{\mathrm{T}}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_{1}M{W}_i+{\theta }_{2}1\{OT{C}_i\}\\ & +{\theta }_{3}1\{\mathrm{P}{\mathrm{u}}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{IO{U}_i\}+{\gamma }_3{M}_i+{\delta }_f+{\mu }_c+{\epsilon }_i.& (7)\end{array}$(7) and (8(8) $\begin{array}{rl}& ln(L{T}_i)=\beta sin{h}^{-1}(Ex{p}_{i,f})+{\theta }_1\mathrm{M}{\mathrm{W}}_i+{\theta }_{2}1\{\mathrm{O}\mathrm{T}{\mathrm{C}}_i\}\\ & +{\theta }_{3}1\{P{u}_i\}+\sum _{x\in X}{\xi }_{x,i}+{\gamma }_{1}ln(\mathrm{G}\mathrm{D}\mathrm{P}p{c}_{c,y})\\ & +{\gamma }_{2}1\{\mathrm{I}\mathrm{O}{\mathrm{U}}_i\}+{\gamma }_3{M}_i+{\mu }_c+{\nu }_y+{\epsilon }_i.& (8)\end{array}$(8) )	✓	✓	✓	✓
Observations	595	503	601	509

*Dependent variable: $ln(L{T}_i)$; untransformed regression coefficients shown with t-statistics in parentheses.

Because variables $De{m}_{c,y}$ and $De{c}_{c,y}$ are standardized according to their z-values, the uninteracted coefficient on cumulative experience can be interpreted as the marginal effect of experience within a country with average values of both variables.^ah The estimated learning-by-doing rate is not statistically different from zero in the average nation, whereas the benefit of learning-by-searching appears to be considerable. Averaging the two results in columns (3) and (4) and transforming them according to Eq. (1)(1) $ln(Y_t)=\alpha +\beta ln(Ex{p}_t)+{\epsilon }_t.$(1) implies a learning-by-searching rate of 6.2%. The interpretation is as follows. Consider two reactors from two different reactor families that are built in otherwise identical national conditions at the same point in time. Reactor A’s family has a cumulative experience double that of reactor B’s family. Holding all else equal, we expect reactor A to finish construction 6.2% faster than reactor B. This may be because the family of reactor A is technically superior (hence why it has accumulated more experience) or it may be because of a greater experience base, which facilitates more timely construction. Future research could try to establish causal identification along this “between” dimension.

Next, we examine the interaction terms and consider how these estimated learning rates vary according to political conditions. Columns (1) and (2) provide no support for the hypothesis political decentralization mediates learning-by-doing. Columns (3) and (4) return coefficients with the opposite sign; neither are statistically significant at conventional levels. I conclude that there is no significant mediating role of political decentralization in learning-by-searching.

The result in column (4) contrasts sharply with my finding in an earlier draft of this work, which reported a statistically significant ($t=2.84$) parameter of 0.029 for the experience-decentralization interaction. To contextualize the magnitude of such a parameter, I calculate that a nation with decentralization two standard deviations above the global average would experience a learning rate of −5.5% (i.e., LT would increase by 5.5% for every doubling of cumulative experience). As compared to those of the earlier draft, the finding published in Table X reflects the latest available data from the RAI (Ref. 74), which notably expanded data coverage to India and (mainland) China, inter alia, in early 2021. In an unreported regression, I replicate the earlier finding by excluding India and China from the sample.

This is not to argue that China and India should be disregarded. Instead, I believe their influence on the result is itself a remarkable finding. Unlike most countries that have built NPPs, China and India exhibit considerable within-country, over-time variation in decentralization according to the self-rule index constructed by Hooghe et al.⁷⁴ In India, while the average is 2 standard deviations above the global mean, it ranges from a minimum of 0.5 to 3.3 standard deviations above the global mean. In China, observed decentralization ranges from 1.0 to 1.6, although the preponderance of observations is in the vicinity of 1.

China’s degree of decentralization, while modest relative to federal countries such as the United States (2.1 standard deviations above the global mean), Canada (2.1), West Germany (2.5), and Switzerland (1.5), make it an outlier compared to its neighbors Taiwan (0.0), South Korea (−0.2), and Japan (0.6).

A final observation is the clear learning-by-doing trend in India’s indigenous PHWR family, which I plot in Fig. 5.^ai India is formally a federal nation; its federalism has deep historic roots. The data from the RAI are consistent with this. Thus, India’s substantial reductions in LT are stark evidence against the supposition that federalism is incompatible with progress in the construction economics of NPPs. Recognition of India’s achievements is far from a novel contribution,^11,83 but I highlight them here to acknowledge the challenge that India presents to my theory.

Fig. 5. Lead time for PHWRs in India.

VI. CONCLUSION

VI.A. Discussion

This paper hypothesized and investigated several mechanisms by which decentralization influences the LTs of NPPs. The findings are as follows.

The design specifications of NPPs, in so far as they relate to LT, do not appear to be correlated with political factors. Richer countries have a tendency to build larger reactors. Decentralized countries have a higher propensity to build reactors of a nonstandardized design, but this association is explained by their higher degree of electricity market fragmentation. Without a single national electric utility, coordination on a standardized design tends not to happen, except by explicit national policy, as in Japan.

Conditional on observed design specifications, I find that NPPs tend to take longer to build in politically decentralized nations. Further research is needed to improve the data coverage of safety-related technical characteristics of NPPs to ensure that the comparison is truly between “otherwise identical” NPPs. When the comparison is restricted to reactors of the same model built in different countries, the apparent partial association between decentralization and LT disappears. One possible explanation for this result is that most reactor models appear in only one or a few similar countries, so the statistical power may be too weak to detect an effect.

I find that East Asian countries are unique in their capacity to manage construction of NPPs with comparatively little penalty arising from scale and more complexity. However, the evidence does not support the hypothesis that decentralization mediates the relationship between scale/complexity and NPP LTs.

The difference in the average learning-by-doing rate (effectively zero) and the average learning-by-searching rate (modest, but statistically significant and indicative of beneficial learning) merits some comment. The cross-sectional dimension and the time dimension of the cumulative experience, as I have defined it, may reflect two different underlying data-generating processes. Over time, as more is learned about the technology of a reactor family, additional time-intensive measures become necessary to implement in reactor design in order to satisfy new safety requirements or perhaps to improve the operational reliability of the plant. This may be for reasons that simply trend upward over time that are unrelated to learning about a specific technology, or perhaps it is a byproduct of learning.

Advocates of SMRs will find much to cheer in my work, as there are clear benefits to small size and standardized design with respect to LT. Applying both the estimated scaling factor and the bonus from design standardization based on the results of Table V, it can be conjectured that a 50-MW reactor of standardized design could achieve LTs on the order of 46 months. Such a LT would go a long way toward improving the economics of NPP construction.

VI.B. Directions for Future Research

I see several opportunities for extending and improving this area of research. In particular, more data on safety-related systems, structures, and components should be collected. The IAEA PRIS and my supplementary data collection provided insufficient data coverage to incorporate such technical specifications into the analysis. Such data would generate greater confidence in the cross-national comparison of technically identical reactors without reliance on fixed effects. This would resolve the ambiguity regarding whether the correlation between decentralization and LT is driven by differing design characteristics or other factors (such as regulation or construction sector productivity), irrespective of design.

A more clean-cut test of the “logic of local democratic control” would require direct measurements of the intensity of local opposition and regulatory burden on individual NPP construction projects. In chapter 3 of my dissertation,¹¹⁰ I conduct such an analysis for the United States. However, comparable measures of regulatory burden appropriate to the regulatory context of other nations would be desirable to improve the generalizability of findings.

The present work has largely taken the reactor family and the reactor model—as defined by the NSSS—as the primary unit of categorizing similar designs. However, NPPs consist of several other important features, such as the containment, the turbo-generator, and the balance of plant. Study of learning by firms involved in other aspects of plant design and construction, such as the AE, the turbo-generator supplier, and the constructor could be another fruitful area for investigation. Berthélemy and Escobar Rangel¹⁰ show a clear role for the AE in learning, but the sample is limited to a handful of countries. More data collection on the identity of these firms could expand the analysis to the global population.

Another avenue for improvement would be to extend the present work by modeling the simultaneous determination of the OCC and LT as in Berthélemy and E. Rangel¹⁰ while using the large sample of OCC data compiled by Portugal-Pereira et al.²² This would be important for determining the extent to which decentralization drives up the OCC as a consequence of longer LT, or if it impacts the OCC directly.

The results of Table V confirm past findings and conventional wisdom among industry observers that reactor standardization improves the economics of NPP construction. While I show that fragmentation of the electricity market is associated with nonstandardization, the issue would benefit from more formal modeling of the decision to standardize. Indeed, why do firms ever redesign NPPs given the heavy upfront development and licensing costs? The answer surely involves regulation and learning, but there are likely to be industrial organization explanations for why customization was so historically prevalent in the U.S. nuclear industry.

The finding that East Asian countries suffer the least from megaproject syndrome in NPP construction, while nations of the Global South suffer the most, could be considered further. If not political decentralization, what attributes of these regions explain the difference? How can East Asian success be exported globally?

Acknowledgments

This research was supported in part with funding from the University of California Irvine Department of Economics. Special thanks are owed to the IAEA for access to restricted data, and to Linda Cohen, Ami Glazer, Damon Clark, Stergios Skaperdas, Dan Bogart, Tony Smith, Aditi Verma, Richard DiSalvo, and three anonymous reviewers for detailed comments. I offer the standard disclaimer that any remaining errors are my own. My website is https://github.com/a-g-benson/.

Disclosure Statement

No potential conflict of interest was reported by the author(s).

Notes

^a While Portugal-Pereira et al.²² limit their analysis to light water reactors, their data appendix provides the OCC for 521 reactors.

^b Duration analysis is also known as “survival analysis,” so called because it is classically used to estimate patient survival after a medical treatment. However, the method extends naturally to modeling the length of time between any two events.

^c See Sec. A.XII in Appendix A for a discussion on the Polity IV democracy-autocracy index.

^d The authors report an $R^2$ of 0.316 in regression of schedule slippage on budget overrun, using a polynomial fit. The estimated fit is nearly linear, so the implied coefficient of correlation is approximately 0.56.

^e The foregoing discussion has been solely of one-factor (cumulative experience) learning. Two-factor learning encompasses cumulative experience and the stock of knowledge. See Wiesenthal et al.³³ for further discussion.

^f $1-2^{-.152}=10\mathrm{\%}$.

^g The author’s own calculations from Table 5-3 of Reference 25. “Base cost” includes all costs in overnight cost except for the contingency allowance (p. 78 of Ref. 19), the amount budgeted to cover unexpected expenses.

^h The author’s own calculations from Table A3 of Rubin et al.³

ⁱ Oyster Creek, December 1, 1969.

^j Dresden Unit 2, January 10, 1966.

^k Seven BWR-5s were built with Mark I containment in Japan by Toshiba and Hitachi, licensees of GE technology.

^l Two ABWRs began construction in Taiwan but were never permitted to operate due to political decisions.

^m The author’s own calculations from the IAEA Power Reactor Information System (PRIS). This average is for plants that have been completed as of the time of writing. Thus, two reactors that remain under construction are excluded.

ⁿ Yonggwang NPP was renamed Hanbit NPP in 2013.

^o Assuming the substitute sources of electricity are polluting. Historically, this has been the case.⁷³

^p Even for nations that depend on uranium imports, importing uranium is much cheaper per unit of final electricity generated than fossil fuels. Provided the nation is a signatory to the Nonproliferation Treaty, availability of supply is a nonissue.

^q Of course, nuclear weapons programs generate negative externalities globally but plutonium recovered from spent nuclear fuel is often considered a benefit by national policymakers who desire nuclear weapons.

^r The data available for public dissemination can be found at https://github.com/a-g-benson/Global-NPP-Database.

^s The bootstrap procedure was employed to calculate a standard error clustered by country.

^t See Arlot et al.⁷⁸ for an introduction to the method.

^u No such interaction term can be estimated for Japan, as the two reactors under construction in Japan on March 11, 2011 have not been completed.

^v In the case of fractional ownership among multiple utilities, I code the variable according to the ownership structure of the lead utility.

^w No reactors under construction as of March 11, 2011, have entered operation in Japan as of the time of writing, so the parameter cannot be estimated.

^x Where country is defined as the country in which construction began, e.g., the Soviet Union and Russia are two separate “countries” for this purpose. Reactors that began construction under the Soviet Union and finished after its collapse are coded as belonging to the Soviet Union.

^y Global averages and standard deviations are computed from the global population of countries, not just those which have built NPPs.

^z The primary coolant is the fluid that conveys heat away from the core, where the heat is generated, to the remainder of the plant where it is converted to steam.

^aa These are the independent piping systems through which primary coolant flows.

^ab A bivariate regression finds that predicted LT increased by about 0.1% per year (p = 0.21) over the sample period.

^ac Disregard the special autonomous status of the Å land Islands, where no NPPs have been built.

^ad $13.2\mathrm{\%}-31.7\mathrm{\%}+50.3\mathrm{\%}=31.8\mathrm{\%}$.

^ae The upper bound of the 95% confidence interval is 1.05.

^af There is one exception, namely, in column (2). Having a federal constitution is not normalized; it is a binary variable. The interpretation of the uninteracted coefficient is that of the marginal effect in nations with unitary constitutions.

^ag The t-statistics in Table IX refer to the coefficient’s statistical difference from zero.

^ah This is with the exception of columns (1) and (3), as having a federal constitution is not normalized; it is a binary variable. The interpretation of the uninteracted coefficient is that of the marginal effect in nations with unitary constitutions.

^ai For historical context, I also include Rajasthan 1 and 2, although I classify them as part of the CANDU family on account of the Canadian involvement in their design and construction. These were delayed due to the sudden termination of Canadian support in response to India’s first test of nuclear weapons. I code all subsequent PHWRs in India as belonging to a separate family.

^aj https://pris.iaea.org/PRIS/home.aspx.

^ak https://pris.iaea.org/signin.

^al https://pris.iaea.org/pris/.

^am Thanks, Dad.

^an https://www.iso.org/iso-3166-country-codes.html.

^ao https://de.nucleopedia.org/.

^ap All reactor models in this family begin with the letters VVER, a Russian acronym that basically translates to “light water reactor.”

^aq This Indian family inherits the cumulative experience of the CANDU family associated with the two reactors in India for which Canada initially provided support.

^ar A nuclear reactor’s capacity may change over time as a result of uprates and downrates—modifications to the original design and/or changes in regulatory permissions

^as Gross capacity is the amount of electrical power produced by the generator. Some of that power is used to operate the reactor and power other facilities at the plant. The amount of power exported to the grid is the net capacity.

^at Specifically, I exclude all events for which the absolute value of the Polity IV variable REGTRANS is less than or equal to 1. This retains “major democratic transitions,” “minor democratic transitions,” “adverse regime transitions,” and “state failures.”

^au With the exception of Romania, which imported a Canadian heavy water reactor design.

^av For example, Czechoslovakia, a federal nation, dissolved and become two unitary nations on the basis of ethnic differences.

^aw Gorbachev, Mikhail, “VIEW: Turning point at Chernobyl” (Apr. 17, 2006); https://www.gorby.ru/en/presscenter/publication/show_25057/.

^ax Suspension and completion are not mutually exclusive outcomes, as 14 reactors have been suspended but were later completed.

^ay Further, assume that these reactors are located at separate sites, and therefore are not being built according to a staggered schedule.

^az The reactors are Watts Bar 2, Bushehr 1, Atucha 2, and Kalinin 4.

^ba Where I define success as having the most completed reactors associated with a family.

APPENDIX A

THE BENSON DATABASE OF NUCLEAR POWER PLANTS

This appendix documents the compilation of the database upon which I rely for all analyses in this dissertation. The data, excepting any data that are subject to IAEA data sharing restrictions, have been made available at https://github.com/a-g-benson/Global-NPP-Database.

A.I. UNIT OF OBSERVATION

The unit of observation is the “nuclear generating unit,” which is a discrete collection of equipment and structures that are jointly necessary to safely and efficiently generate electricity from nuclear fission. A NPP may host multiple nuclear generating units, which may be constructed and operated independently.

For brevity and to avoid confusion with other several other uses of the word “unit,” I refer instead to “reactor” as a metonym for “nuclear generating unit.” The reactor is a central component of a nuclear generating unit, although it would not be able (or permitted) to function without the other components, such as the reactor coolant system, containment, steam turbine, structures for discharging waste heat, generator, and electrical switchyard.

A.II. TYPES OF DATA

My database consists of a diverse mix of cross-sectional, time series, and panel data. The primary structure of the database is cross sectional in nature, recording for each reactor pertinent data about its identity, location, ownership, technical specifications, important dates (of construction, operation, and retirement), and so on. In general, in cases where a variable may take on different values at different points in times (e.g., a reactor’s ownership may change hands), the reactor is assigned a value for that variable that corresponds to the date it began construction.

In certain cases, discussed in the body text or the following, time series or panel data may be transformed into a cross-sectional variable by computing the sum or average value of that variable within a range of specified dates appropriate for the reactor, such as the dates on which construction began and finished.

A.III. SAMPLE DEFINITION

The foundation of the database is provided by the PRIS of the IAEA. While certain basic information about each NPP is available through a public website^aj and through various IAEA publications,¹⁷ I was granted temporary access to a private version of the PRIS database restricted to authorized users.^ak

The IAEA PRIS database consisted of 1056 reactors as of the month of access (July 2018). The reactors in this sample consist primarily of the global population of all nuclear reactors that have ever entered operation for the purposes of commercial electricity generation. It furthermore includes all commercial reactors that have begun construction, including those which were never finished and those that are presently under construction. The PRIS classifies a reactor as having begun construction if “the first major placing of concrete, usually for the base mat of the reactor building, was carried out.”⁸⁴ Construction activities involving site preparation do not qualify.

In addition to all reactors which meet this definition of having begun construction, the PRIS includes a fairly large sample ($n=284$) of “planned” but never (or not yet) built reactors. The IAEA states that a reactor is categorized as “planned” during the period “when a construction licence application has been submitted to the relevant national regulatory authorities” but construction has not yet begun.¹⁷ However, inclusion into the PRIS is evidently inconsistent with this criterion. For example, the Ninh Thun NPP, which Vietnam had contemplated for nearly a decade until it was canceled by a vote of Vietnam’s National Assembly in 2016 (Ref. 85), is included in the PRIS yet no news reports indicate that a construction license application was ever submitted to the national regulatory authority.^86–88

Conversely, the PRIS omits 26 proposed reactors that Berndt and Aldrich⁶¹ include from the historical record in the United States, which I append to my database. Of these 26 reactors absent from the PRIS, I identify 12 with valid docket numbers from the U.S. Atomic Energy Commission, implying that an application for a construction permit was submitted. I conclude that the PRIS most likely does not contain the global population of proposed but never built commercial power reactors.

In the time subsequent to my temporary access to the restricted data of the PRIS, I have manually updated my database with data from the public-facing version of the PRIS^al and Reference Data Series No. 2 (Ref. 17). In most cases, this entailed updates to the dates related to plant construction and retirement. However, it was also necessary to append four observations from China, which were not present in the PRIS as of July 2018 (even as “planned” reactors). These are Zhangzhou 1 and 2 and Taipingling 1 and 2, which began construction in 2019 and 2020.

One final reactor which I have appended is an observation representing the original Bushehr 2 in Iran. Bushehr 1 and 2 were originally designed by Kraftwerk Union and began construction in 1975. Due to circumstances arising from the Islamic Revolution and the Iran-Iraq War, only Bushehr 1 was ultimately completed after considerable delay and Russian assistance. The PRIS contains an observation that refers to the Russian-designed reactor presently under construction which bears the official designation “Bushehr 2.” This is a brand new reactor, built from a clean slate, rather than an effort to complete the original Bushehr 2. Therefore, I treat these two instances of “Bushehr 2” as separate observations.

The PRIS does not clearly distinguish between “commercial” and “research” power reactors. For example, BOR-60—a sodium-cooled, fast breeder reactor (FBR) in Russia with a nameplate electric capacity of 12 MW(electric)—is included, while the Experimental Breeder Reactor II—a 20-MW(electric) American FBR—is excluded. Both were designed, built, and operated by state-owned scientific laboratories for noncommercial purposes. While both generated electricity that was exported to the grid, the electricity was fundamentally a byproduct of the research.

The boundary between “research” and “commercial” is somewhat blurry. For example, the Shippingport Atomic Power Station was not economically justified on its own commercial merits. Rather, its purpose was as a proof of concept for future commercial PWRs and to provide operating experience for the electric utility industry (p. 421 of Ref. 89). I argue that, while many aspects of a reactor like Shippingport are unrepresentative, its inclusion is necessary to view the full picture of the historical evolution of commercial NPPs.

Another category of reactor that blurs the lines of “commercial” are those which historically served dual purposes of electricity generation and plutonium production for nuclear weapons. The earliest GCRs built by Great Britain and France were designed explicitly for this purpose at some penalty to their economics; they nevertheless generated commercial quantities of electricity that were exported to the grid. State ownership of the electricity sector further blurs the line between “commercial” and “noncommercial” in these cases. Reactor designs in these families were later refined to improve their economics but there is not a clear “break” at which such reactors became purely peaceful endeavors in their fundamental design.

Lacking a clear definition by which to differentiate commercial power reactors from others, I elected to retain reactors present in the PRIS but to not append any which were absent except those identified previously, all of which were unambiguously commercial in purpose and scale. To account for unexplained properties of reactors of questionable “commercial” nature, I manually coded various binary indicator variables. My primary justification for retention of such reactors was to ensure accurate measurement of cumulative experience.

Therefore, the final data set consists of 1087 observations (reactors) at 404 sites (power plants) in 50 countries (as defined by their present-day boundaries). Table A.I tabulates the count of observations according to their status (planned, under construction, operational, etc.) as of April 6, 2021.

A.IV. DATA CLEANING PROCEDURES

Because the data in the PRIS were provided by the owner of the NPP in question or by a governmental representative of the country in which it is located, the PRIS suffers from internal inconsistencies in the coding of many of its variables. I employed my knowledge of the subject matter to clean up the data where inconsistencies were obvious. For example, Framatome changed its name to Areva during a restructuring in 2001, only to later change it back in 2018 after another restructuring. Reactors designed and manufactured by this company are not consistently labeled under a single name in the raw PRIS data set. Similarly, I treat Rosatom—Russia’s state-owned monopoly in NPP construction and operation—as the one-in-the-same firm as the Soviet Ministry of Medium Machine-Building, which was responsible for the Soviet nuclear power program. Rosatom came about through a series of restructurings after Chernobyl and the collapse of the Soviet Union. Rosatom retains the intellectual property and the Soviet manufacturing infrastructure related to nuclear power in Russian territory; it even occupies the same headquarters in Moscow as the old Soviet ministry.

Furthermore, missing data are a pervasive problem in the raw PRIS data set. Where possible, I filled in missing data by referring to publications in nuclear engineering journals and documents released by nuclear regulatory agencies. Major supplementary sources of data are noted in Sec. A.XI. In a handful of cases, missing data concerning particular reactors were supplied directly to me by personal contacts^am in the nuclear industry.

A.V. SITE DATA

Site data terms are defined as follows.

Name of the Site: As discussed in Sec. A.I, multiple reactors may be colocated at the same site. Generally, the string of text designating a shared site was taken directly from the PRIS without alteration. However, in a handful of cases I merged sites listed separately into a single site that better reflects the colocation of certain reactors. For example, the PRIS lists the site for the Shippingport Atomic Power Station as “Shippingport” and the site for Beaver Valley Units 1 and 2 as “Beaver Valley.” In light of the fact all three reactors were built immediately adjacent to each other, I edited the name of the Shippingport reactor’s site to “Beaver Valley” to unify all three reactors with a single coding. The purpose of this coding is to properly account for spatial autocorrelation in regressions that cluster reactors by site.

Total Reactors: For each reactor, I generate a count of the total number of reactors whose construction has been completed at that the same site.

N^th Reactor: For each reactor, I identify whether it is the first, second, etc., reactor to be built at that site.

Tuplet Group: Many reactors were built as twins, triplets, or (in rare cases) higher-order tuplets at the same site. I encode a variable that groups reactors together according to whether they are identical reactors built around the same time as a combined project.

Total Tuplets: For each reactor, I generate a variable that equals one if the reactor has no identical siblings, two if it has a twin, three if it belongs to a set of triplets, and so on.

N^th Tuplet: For each reactor, I identify whether it is the first, second, etc., reactor to be built within its tuplet group.

Shared Start Dates: 149 reactors are listed as having begun construction on the same day as one or more other reactors at the same site; 132 of these are twin reactor units, along with one set of triplets, two sets of quadruplets, and one set of sextuplets. When multiple reactors are reported to have begun construction in tandem at a site, it is atypical for those reactors to be completed on or around the same date. This reflects the fact that NPP construction management usually economizes on equipment and labor by not performing the same tasks for both reactors at the same time. Thus, the second reactor is liable to finish, approximately, 1 year after the first, the third 1 year after the second, and so on. This pattern can be almost perfectly predicted by the number assigned to each unit. For example, Calder Hall Units 1 and 2 are both listed as having begun construction on August 1, 1953, but Unit 1 became operational 4 months earlier than Unit 2.

To account for this, I generate a control variable, which I abbreviate $M_i$, which ranks reactors at the same site that share the same start date. The reactor with the smallest unit number (or alphabetically earliest letter) is assigned a value of one on $M_i$, the second smallest (or alphabetically earliest) is assigned a value of two, and so on. A reactor which (A) has no twin or higher-order tuplet or (B) whose twin is listed as having begun construction on a different day is also assigned a value of one on $M_i$. Therefore, the interpretation of any coefficient on $M_i$ refers to the marginal effect of increasing by one the number of reactors that began construction on the same date and the same site as reactor $i$ but were prioritized over reactor $i$ in the construction process.

A.VI. GEOGRAPHIC DATA

Geographic data terms are defined as follows.

Subnational Region: Many (but not all) sites are matched with their current International Organization for Standardization (ISO) 3166-2 code for principal subdivision (e.g., province or state). This work remains ongoing; special attention is needed regarding the matter of changes in subnational boundaries over time.

Country: Construction of the first observation in the data set commenced in 1951, and several major changes in international borders have occurred since that time. For the purposes of the analysis in this work, a reactor’s “country” is whichever national entity had territorial sovereignty over the site as of the year construction began. In particular, this means several reactors which began construction under the Soviet Union and Czechoslovakia but were finished after the dissolution of those countries are considered to be “in” their former countries. However, in post-Soviet countries, work on 14 new reactors has begun, 12 in Russia, and 2 in Belarus. For the purposes of country fixed effects and standard errors clustered by country, Soviet successor states are treated as distinct countries in these cases.

The identity of the country with territorial control over the site as of the years 1950 and 2020 are also coded for expositional purposes and the benefit of future users of the data. In all cases, counties are coded according to their ISO 3166 Alpha-3 abbreviations.^an

Latitude and Longitude: Latitude and longitude data were primarily drawn from the Global Power Plant Database by the World Resources Institute.⁹⁰ Missing data were supplemented from Nucleopedia, a German-language wiki on NPPs,^ao and a series of reports from Oak Ridge National Laboratory^91–95 on American NPPs.

Local Climate: Chen and Chen⁹⁶ provide a global map of Köppen climate classifications at a resolution of 0.5 × 0.5 deg. I match every NPP site with the nearest centroid of this grid, except where the centroid lies in the ocean or other large body of water. The data from Chen and Chen⁹⁶ do not classify the climate of such cells, and I instead match the site to the nearest centroid over land.

Cooling Water Salinity: The local source of cooling water is categorized, principally with respect to its salinity: ocean, brackish sea (principally the Caspian and Baltic Seas), estuary, freshwater, or municipal wastewater (a water source unique to the Palo Verde NPP).

A.VII. DATES OF SIGNIFICANCE

Dates of significance are defined as follows.

Construction Start: These dates come directly from the PRIS, which defines construction as having begun “when the first major placing of concrete, usually for the base mat of the reactor building, is carried out.”¹⁷ For brevity, this event is sometimes called “first concrete.” Site preparation proceeds first concrete, but data regarding the commencement of site preparation are not widely available, particularly because it can proceed regulatory approval.

Construction Suspension: 195 reactors in the sample have had their construction suspended (whether temporarily or permanently). The dates of such occurrences are partially available in the PRIS; however, the data availability is incomplete and it is provided in the same field as the date of retirement from commercial operation. I have created a separate variable for the date of construction suspension to rectify this coding issue. Missing dates have been supplemented manually through case-by-case historical research, principally through consultation of issues of Nucleonics Week. I also retain a variable indicating the level of precision by which this date is known (day, month, or year) from my research.

Construction Restart: 22 reactors in the sample have resumed construction following a suspension. No dates of construction restart were provided by the PRIS; all were gathered manually through case-by-case historical research. A variable indicating the level of precision by which these dates are known is also retained.

First Criticality: Criticality is “the state of a nuclear chain reacting medium when the chain reaction is just self-sustaining.”⁹⁷ For NPPs, the first instance of criticality occurs prior to commercial operation, as part of the commissioning procedures and tests. These dates are provided by the PRIS with no supplementation from other sources.

Grid Connection: The date of grid connection refers to “the date when the plant is first connected to the electrical grid for the supply of power.”¹⁷ However, typically some time elapses after this event before the plant is officially declared to be in commercial operation; trial operations and further tests are usually carried out. These dates are provided by the PRIS with no supplementation from other sources.

Commercial Operation: The date on which a reactor is considered to be in commercial operation is “the date when the plant is handed over by the contractors to the owner and declared officially in commercial operation.”¹⁷ These dates are provided by the PRIS with no supplementation from other sources.

Retirement: The retirement date is defined as “the date when the plant is officially declared to be shut down by the owner and taken out of operation permanently.”¹⁷ These dates are provided by the PRIS with no supplementation from other sources.

A.VIII. Construction Economics

Construction economics terms are defined as follows.

Gross Lead Time: I compute gross lead time as the difference in days between the construction start date and the date of commercial operation. For ease of exposition, I usually present this number in months ($\frac{days}{30.44}$), years ($\frac{days}{365.25}$), or the natural logarithm of months.

Net Lead Time: I subtract the number of days during which a reactor’s construction was suspended (if any) from the gross lead time to generate the net lead time.

To account for the problems arising from suspension of construction, I retain an indicator variable that takes on the value one for reactors which were suspended and a continuous measure of the number of months during which construction was suspended. In unreported regressions, I find that—after subtracting the months of suspension from the gross lead time—the length of the suspension period has no statistically significant marginal effect on net lead time over and above the predictive power of a binary indicator of whether construction was ever temporarily suspended for any length of time.

For all summary statistics, analyses, and graphs in this work, I use net lead time, although I refer to it as LT for brevity. I wish to emphasize that net lead time is not intended to represent a “complete” measure of LT for purposes such as estimating the levelized cost of energy or comparing the LTs of NPPs to the LTs of other technologies. Gross lead time, including periods of construction suspension, is the appropriate metric for those purposes. It may also be desirable to include planning and permitting phases, as in Aldrich,⁶² for certain purposes. My purpose in defining LT in this way is to generate an outcome metric that improves apples-to-apples comparisons of NPPs in order to understand why LT varies cross nationally and over time. It would be unfair to compare on the basis of gross lead time, for example, French and Soviet PWRs that began construction in the 1980s. Many Soviet NPP projects were put on hiatus for macroeconomic and political reasons. If policymakers and industry participants wish to improve the economics of NPP construction, one simple change they can make is to avoid suspending construction, insofar as they can help it.

Overnight Capital Cost: I append to my database the OCC data of Portugal-Pereira et al.,²² which is in purchasing power parity–adjusted U.S. dollars, inflation-adjusted to the year 2010.

A.IX. Firms Involved in Design, Construction, and Operation

Firms are defined as follows.

NSSS Designer: The PRIS provides the name of the designer(s) of the NSSS. Extensive editing was performed by hand to ensure a single, consistent name for each firm. In cases where more than one firm is listed for a single reactor, the firm with more experience or holding the intellectual property is identified as the “primary” designer.

Turbo-Generator Manufacturer: The PRIS provides the name of manufacturer(s) of the steam turbine/generator set (turbo-generator). Extensive editing was performed by hand to ensure a single, consistent name for each firm. The identity of the manufacturer of the turbo-generator was ultimately not used in any of the analyses described previously.

Architect-Engineer: The AE is the firm that was responsible for the design of the overall plant, unifying the NSSS with the steam turbines, generator, other major infrastructure, and auxiliary buildings. This information is not provided by the PRIS. Instead, I compiled the data provided by Berthélemy and E. Rangel¹⁰ and Gavrilas et al.,⁴¹ which provide coverage for light water reactors of Western design. Remaining gaps were filled in with data from the World Nuclear Industry Handbook.⁹⁸

Constructor: The constructor is the firm responsible for day-to-day management and supervision of construction at the site, including the hiring and managing of many subcontractors for specific tasks. In some cases, one firm serves as both the AE and constructor.

Lead Utility: The PRIS only identifies the current, primary owner of each reactor. Therefore, I consulted other sources to identify the original utility in the case of NPP divestments in jurisdictions that underwent liberalization of their electricity markets. From this information, I generated a variable indicating whether the lead utility—typically, the single largest owner—was investor owned (1) or state owned (0) as of the date construction began. For utilities of mixed ownership, this variable takes on the value 0.5.

A.X. Reactor Typology

Reactors are classified according to the following typologies.

Reactor Type: I use the term “type” to encapsulate broad similarities in the principles of a reactor’s design. The most common types are PWR, BWR, PHWR, GCR, and LWGR. All other types were aggregated into a category called “other” due to a sparsity of observations.

Reactor Family: I use the term “family” to classify reactors that have a shared evolutionary heritage. For example, all PWRs of Soviet or Russian origin are grouped into the VVER family.^ap The largest family is the Westinghouse family, which includes not only PWRs designed by Westinghouse, but those designed by firms which licensed Westinghouse’s intellectual property, notably Framatome, Siemens, and Mitsubishi. The identification of families was based explicitly on the “family trees” provided in Gavrilas et al.⁴¹ for Western light water reactors and Sidorenko⁹⁹ for the Soviet VVER and RBMK families. The CANDU family is identified in Garland¹⁰⁰; I treat India as having branched off and established a separate family of heavy water reactors after Canada (the originator of the CANDU design) canceled its cooperation on nuclear power in response to India’s first nuclear weapons test in 1974.^aq Future research could improve upon this classification scheme by properly accounting for cross fertilization in reactor design that has occurred in recent decades.

Reactors of unconventional and experimental designs that were never iterated upon are coded as belonging to a family equal to their reactor model.

Sister Group: I draw on the “sister unit group” classifications of the Information System on Occupational Exposure, a project of the OCED Nuclear Energy Agency.^101,102 Where possible, I extend these classifications to reactors that are absent from the aforementioned sources on account of retirement or abandoned construction. These classifications occupy a middle ground of granularity between family and model. They are more specific than family in that sister groups are based on the firm that designed the NSSS, whereas family is based on the firm that originated the intellectual property for the NSSS. In addition, sister groups also specify the vintage of the plant (e.g., BWR-1, BWR-2, BWR-3, and so on); for PWRs, they further specify the number of primary coolant loops.

Reactor Model: I use the term for the name of the model assigned by the manufacturer, where applicable. Examples of model names assigned by the manufacturer include AP-1000, CP1, P4, OPR-1000, CNP-300, VVER-213, and ABWR. For standardized reactor designs, this classification comes as close as realistically possible to identifying “identical” reactors. However, for nonstandardized designs, the PRIS provides an abbreviated, generalized description of the reactor’s design in place of a model name. For example, “WH 4LP (DRYAMB)” indicates that the reactor is a Westinghouse design with four primary coolant loops and the containment structure operates at ambient atmospheric pressure. Information about the containment design is inconsistently included in the IAEA coding of models, so I remove it and place it in a separate variable.

A.XI. Technical Specifications

Technical specifications are defined as follows.

Capacity in Megawatts: The PRIS offers four measures of the rated capacity: the rated net electric capacity as originally designed, the current rating of the net electric capacity,^ar the current rating of gross electric capacity,^as and the current rating of the thermal capacity of the reactor core. My ideal specification would select the rated thermal capacity as originally designed. Thermal capacity is a more precise indicator of the inherent safety challenges of a larger reactor, whereas electrical capacity—while primarily a function of size—also reflects the thermodynamic efficiency of the plant. However, original thermal capacity is not available from the PRIS. As second best, I use the original net electricity capacity because it is a measure of the “original” size of the plant (prior to uprates) and because it lends itself to a more intuitive interpretation of the results. In any case, robustness checks revealed that none of the results presented herein are sensitive to the specification of this variable.

Design Characteristics: The PRIS includes over 150 variables that quantify or characterize technical details of a reactor’s design. Notable variables include cooling method (e.g., cooling towers versus OTC), height and diameter of the reactor pressure vessel, average density of power per unit volume of the core, reactor outlet and inlet temperature, average core power density, number of steam generators, and number of steam turbines per reactor. A handful of variables are not particularly informative, as they are necessarily implied by a reactor’s type, such as choice of moderator and coolant. Unfortunately, many other variables were left blank for a large number of the observations. Most notably, safety-relevant design characteristics are sparsely provided and inconsistently coded.

To supplement the PRIS, I draw from a variety of sources. The most comprehensive of these is the World Nuclear Industry Handbook,⁹⁸ which has global (albeit imperfect) coverage. For the United States, richer detail is available from a series of reports by Oak Ridge National Laboratory,^91–95 as well as a report prepared for the NRC (Ref. 103). Beyond these sources, gaps in the data were generally filled in with sources particular to the reactor or site, such as regulatory records or passing references found in industry periodicals. Where possible, multiple sources were consulted to identify and reconcile discrepancies between sources.

Standardization: I code every reactor as either standardized (1) or nonstandardized (0). A reactor was determined to be standardized if the preponderance of the literature characterized it (or all reactors of its model) as standardized. Sources consulted include Gavrilas et al.,⁴¹ Goldberg and Rosner,¹⁰⁴ Lovering et al.,¹¹ Csereklyei et al.,⁷ and back issues of Nucleonics Week. This dichotomous coding of standardization is not ideal, as standardization is arguably better characterized by a continuum of similarity or dissimilarity between two reactors. I generated such a continuous measure, drawing from within-model variation in design characteristics. However, in robustness checks, continuous measures of standardization were not found to contribute any meaningful explanatory power above and beyond that provided by a dichotomous indicator of standardization. Therefore, I adopt the dichotomous coding as my preferred measure of standardization.

Containment Design: I classify containment as falling into one of ten categories. These are listed in Table V. Data coverage here is imperfect, as 108 reactors are classified as having an “unknown or other” design of containment. These are primarily early and experimental reactors, but it also includes 12 commercial-scale BWRs that cannot be classified as either Mark I, Mark II, or Mark III. Further research is needed to close these gaps in the data.

A.XII. Country-Level Data

Country-level data are defined as follows.

GDP per Capita: I draw from the Maddison Project Database⁷⁵ for its historical estimates of GDP per capita. While the Maddison Project reports data for the former USSR and Yugoslavia, it does not desegregate East and West Germany. For these countries, I rely on data from Broadberry and Klein.¹⁰⁵

Democracy: The “Polyarchy” index of electoral democracy provided by the V-Dem Project⁷⁶ is my preferred measure of democracy. The V-Dem project is an ongoing collaboration of “six Principal Investigators (PIs), seventeen Project Managers (PMs) with special responsibility for issue areas, more than thirty Regional Managers (RMs), 170 Country Coordinators (CCs), Research Assistants, and 3,000 Country Experts (CEs)” who generate quantitative measures of the characteristics of government. It is currently headquartered at the University of Gothenburg.

In unreported robustness checks, I also use the Polity score of democracy/autocracy from Polity IV, a project of the Center for Systemic Peace.⁷⁷

Decentralization: I test three measures of decentralization. The first is a binary indicator of whether the country has a federal or unitary constitution as of the year in which construction begins. This is a fairly coarse measure, failing to capture more complex cases like Spain. Spain formally declares itself a unitary nation, but in practice has operated with a high degree of regional autonomy ever since the end of the Franco regime and the restoration of the monarchy. Conversely, the USSR considered itself a federation of several constituent republics, but—as a totalitarian regime—operated in a highly centralized manner in practice, up until its final years, over the course of which it ultimately dissolved.

A more fine-grained metric is the “division of power index” from V-Dem. This index measures whether local and regional governments exist, whether they have elected offices, and the extent to which elected local and regional governments can “operate without interference from unelected bodies at the local [and regional] level[s].” The V-Dem codebook is careful to stress that this variable does not measure the power of local and regional governments relative to the national government. It is better conceptualized as the degree of democratic control at the local and regional levels of government. However, the primary benefit of using this measure of decentralization is that it provides complete data coverage; no observations are dropped from the analysis on account of missing data from V-Dem. A severe downside is that it is highly collinear with Polyarchy (r = .91).

The richest measure of subnational political autonomy is from the RAI by Hooghe et al.⁷⁴ They evaluated the constitutions and political histories of individual countries and they systematically scored them on matters such as the role of subnational governments in approving constitutional change, whether the central government holds a veto over subnational decisions, and the autonomy of subnational jurisdictions in setting their tax base and rates. These scores are summed to generate indices along two dimensions of decentralization: self-rule (“the authority exercised by a regional government over those who live in the region”) and shared rule (“the authority exercised by a regional government or its representatives in the country as a whole”). These two indices are then summed to generate a single, generalized measure of decentralization, which they call the RAI. However, I only use the self-rule index, as it more closely pertains to the theory I elaborate in Sec. II.G.

To increase coverage of the RAI data, I rely on the coding of self-rule from Sorens¹⁰⁶ for South Africa. My final data set matches an RAI self-rule score to 534 completed reactors, out of 637 total.

Regime Change: I rely on data from Polity IV to identify the dates and magnitudes of regime changes. I assign a value of 1 to a reactor if it was under construction (or suspended) during an episode of major regime change, and zero otherwise. I exclude relatively minor^at “regime transition events,” such as the resignation of U.S. President Richard Nixon, which corresponds to a small increase in the Polity score for the United States of America. The resulting binary indicator largely reflects the fall of communism in Eastern Europe. However, it also captures the Iranian Revolution and the beginning and/or ending of military dictatorships in Spain, Latin America, and Asia.

Geopolitical Region: Table I disaggregates summary statistics by four geopolitical regions. In assigning countries to these regions, I applied the following judgments in ambiguous cases.

Certain capitalist countries in Europe are not members of the North Atlantic Treaty Organization (NATO) (Switzerland, Sweden, and Finland) or were not members of NATO as of the year construction began (Spain prior to 1982). These nations are nonetheless classified as part of the Western Bloc due to broad similarities to NATO nations in their political, economic, and cultural characteristics, as well as their choice of Western nuclear technology.

Slovenia, while under communist rule as part of Yugoslavia during the period when the Krško NPP was built, was not classified as part of the Eastern Bloc. As a result of Tito’s diplomatic “split” with Stalin and his role in the foundation of the Nonaligned Movement, Yugoslavia imported a Westinghouse design for its reactor rather than a Soviet one. Therefore, Yugoslavia/Slovenia was assigned to the reference region.

Twenty-three reactors in Eastern Bloc countries have entered commercial operation after the collapse of communist regimes (including 15 that were under construction during episodes of regime change). Although some of these countries subsequently joined NATO, observations in such countries are still classified as Eastern Bloc because Soviet technology was employed^au and/or construction began prior to the collapse of communism.

East Asian countries were grouped separately from South Asian countries due to the relatively high LTs of NPPs built in India and Pakistan and relatively low LTs in East Asian nations, as compared to the global average. Because this categorization was explicitly motivated by patterns in the outcome variable, it is more of a descriptive than explanatory variable. However, it should be noted that the cultural, historical, and economic differences between East Asia and South Asia are tremendous, beginning with their independent development as “cradles of civilization,” separated by the largest mountain range on Earth.

Any country not assigned to the Western Bloc, the Eastern Bloc, or East Asia was assigned to the reference category, which may be conceptualized as the Global South or the Nonaligned Movement. Note that Argentina, Brazil, and Mexico are observers but not members of the Nonaligned Movement.

APPENDIX B

METHODOLOGICAL APPENDIX

B.I. Exogeneity of Political Institutions

In all of the specifications described in Sec. IV, I take democracy and decentralization to be exogenous. Political institutions are almost surely exogenous to NPP design and construction activity. For most nations in the sample, the constitutional design was chosen long before the discovery of nuclear fission in 1938 and it has continued with only modest changes up to the present day. In the rare cases where it changed during the sample period, the LT in constructing NPPs was almost certainly unrelated to the change.^av One may argue that the dissolution of the USSR was meaningfully hastened by the Chernobyl disaster—a theory which has been endorsed by ex-President Mikhail Gorbachev.^aw However, modeling this historical trajectory is beyond the scope of the present work. All regime changes are assumed to be exogenous for my purposes.

In theory, countries which undergo regime change or constitutional reform should offer fertile ground for causal inference. However, too few of the observations lie on both sides of major regime changes or constitutional reforms within a single country, limiting the statistical power of a hypothetical event study. Furthermore, for NPPs that began construction under one regime and finished under another (e.g., the Soviet Union and Soviet successor states), it is hard to disentangle the effect of economic upheavals commonly associated with regime change from the effects of the new regime per se.

B.II. Serial Construction

A total of 149 reactors are listed as having begun construction on the same day as one or more other reactors at the same site; 132 of these are twin reactor units, along with one set of triplets, two sets of quadruplets, and one set of sextuplets. When multiple reactors are reported to have begun construction in tandem at a site, it is atypical for those reactors to be completed by the same date. This reflects the fact that NPP construction management usually economizes on equipment and labor by not performing the same tasks for both reactors at the same time. Thus, the second reactor is liable to finish, approximately, 1 year after the first, the third 1 year after the second, and so on. This pattern can be almost perfectly predicted by the number assigned to each unit. For example, Calder Hall Units 1 and 2 are both listed as having begun construction on August 1, 1953, but Unit 1 became operational 4 months earlier than Unit 2.

To account for this, I generate a control variable, $M_i$, which ranks reactors at the same site which share the same start date. The reactor with the smallest unit number (or alphabetically earliest unit letter) is assigned a value of one on $M_i$, the second smallest (or earliest) is assigned a value of two, and so on. A reactor which (A) has no twin or higher-order tuplet or (B) whose twin is listed as having begun construction on a different day is also assigned a value of one on $M_i$. Therefore, the interpretation of any coefficient on $M_i$ refers to the marginal effect of increasing by one the number of reactors that began construction on the same date and the same site as reactor $i$ but were prioritized over reactor $i$ in the construction process.

B.III. Abandoned Construction and Possible Selection Bias

The PRIS lists 95 reactors where construction has begun but has never been completed, as of April 6, 2021, due to suspensions or cancellations. This suggests the possibility of selection bias, as reactors which are taking longer to build for reasons related to decentralization (or any other explanatory variable of interest) are more liable to have their construction abandoned due to poor economics. Table B.I summarizes these observations by country and lists known or likely explanations for the abandonment of construction. Abandoned construction can be broadly grouped into three typologies: conditions in federalist democracies (43 observations), the fall of communism and its geopolitical fallout (35 observations), and regulatory/political decisions at the national level in democracies (11 observations).

Nations transitioning out of communist regimes tended to suspend or abandon construction on their reactors for the same set of reasons: shortfalls in financing, a collapse in electricity demand, and the fresh memory of Chernobyl in the minds of voting publics. In such cases, I argue that the noncompletion is attributable to regime change. In former East Germany, the newly reunited German government shut down the operating reactors and canceled those under construction on the grounds that Soviet-designed reactors did not meet West German safety standards. The abandoned reactor in North Korea was being supplied by the United States as a condition of a 1994 agreement to convince North Korea to remain a party to the Nonproliferation Treaty. Construction began in 2002 and ended a year later when the agreement broke down.

However, the slew of cancellations by utilities in the United States, primarily in the 1970s and 1980s, do present a serious selection concern. The proximate motive for these voluntary cancellations, by and large, were economic factors: budget overruns, schedule slippage, and downward revisions in electricity demand forecasts. However, the effect of the political and regulatory environment on schedule slippage is a precisely the causal mechanism under study.

The abandoned reactor in West Germany presents a similar selection concern. The SNR-300, a fast breeder reactor, began construction in 1973 near Kalkar, North Rhine-Westphalia. While its cancellation can be formally attributed to the decision in 1990 of the state government to deny permission to operate, substantial delays had already occurred due to local public protest and regulatory intervention by the state government. Had it instead been permitted to operate, it would register in the data as another observation with long LT in a nation with high decentralization.

The expected selection bias due to the United States and West Germany is negative. In general, utilities are more likely to abandon construction on reactors that are behind schedule than those for which construction is proceeding smoothly. To the extent that the treatment (decentralization) has a causal effect on the outcome (LT), it is expected that higher levels of the treatment cause higher rates of attrition from the study (failure to complete construction). Reactors that finish construction are in this sense a selected sample of “survivors.”

In Table B.II, I report the results of two probit models, the first taking suspension of construction as the outcome of interest and the second evaluating completion.^ax I find that neither democracy nor decentralization are statistically meaningful predictors of either outcome, although GDP per capita is meaningfully associated with the probability that a reactor is suspended. Moreover, suspension and completion are much more strongly predicted by momentous events, namely, nuclear power accidents and regime change. I take this as evidence that selection bias—insofar as it might bias downward the coefficients on democracy and decentralization—is of minimal concern. Selection bias is almost certainly present in the coefficients on regime change and nuclear power accidents, the correction for which I discuss in Sec. B.IV. The regressions presented in Sec. V were also estimated with the Heckman correction,¹⁰⁷ but these are not reported here because the differences in the results are quantitatively negligible.

B.IV. Modeling the Effect of Major Events

The three largest nuclear accidents—namely, those at TMI, Chernobyl, and Fukushima Daiichi—are widely recognized among industry observers as producing episodes of regulatory instability and political difficulty for NPPs under construction. Additionally, I consider the effect of regime change, which is a leading cause of construction suspension and cancellation, as noted previously.

It would be desirable to control for these events, even if they are uncorrelated with the variables of interest, for the sake of improving the precision of the model. However, there is a problem of endogenous selection into treatment (i.e., being under construction during an event). Consider two reactors that are identical on all observable characteristics and began construction on the same date.^ay If one reactor finished construction prior to the TMI accident while the other finished after, there necessarily must exist some unobserved characteristic of the second reactor that caused it to take longer and therefore be exposed to the political/regulatory aftermath of the accident. For this reason, the estimated effect on LT is necessarily biased upward.

To resolve this endogeneity issue, I instrument for selection into treatment with a nonlinear function of the date on which construction began. To construct this instrument, I first set aside reactors that began construction after a given event. With the remaining reactors, I estimate a binary probit model that regresses selection into treatment on the date construction began. I then generate the predicted probabilities of having been still under construction as of the date of the event. For the reactors that began construction after the event, I assign a predicted probability of zero. For such reactors, the event is not an unanticipated shock.

This procedure generates the instrumental variables for selection into treatment by major events. The F statistics for the first stage of the regression reported in column (2) of Table VIII are reported on Table B.III. Nearly all of them are extremely large (greater than 50), with the exception of the instrument for being under construction during the Fukushima Daiichi disaster. The weak relevance of the instrument may be an artifact of four long-delayed reactors that began their construction prior to the year 2000, had their construction suspended for a decade or longer, and only resumed construction much later.^az Thus, the probit model estimates that reactors that began construction decades before 2011 have a nontrivial probability of exposure to the Fukushima Daiichi disaster.

I argue that the exclusion criterion is satisfied for two reasons. First, the events in question are unanticipated, so they cannot have a casual relationship that flows backward in time to influence the start date. Second, the instruments have an unusual nonlinear and stepwise relationship with time; they are unlikely to correlate with other possible unobserved variables that may trend over time.

B.V. Measuring Cumulative Experience

The decision of how to quantify cumulative experience for the purpose of estimating learning-by-doing raises numerous issues. By convention in the literature on electricity generation technologies, the unit of measure of cumulative experience is the megawatt.^3,10 For example, utility-scale solar and wind farms consist of so many wind turbines and solar panels that it is not particularly important to count the discrete number of panels and turbines. However, I argue that the megawatt is a less theoretically applicable unit of measure for NPP construction. Nuclear reactors are quite lumpy in nature due to their (traditionally) massive size. In my view, a firm that has built ten 200-MW reactors has had five times as many opportunities for learning as a competing firm that has built a pair of 1000-MW reactors. By contrast, whether 2000 MW of solar panels are divided up into two or ten solar farms does not matter at all to the factory that produced the panels; the only difference is that there may be some modest economies of scale in the installation process for larger solar farms.

I draw on the work of Gavrilas et al.⁴¹ and Sidorenko⁹⁹ to conclude that a credible measure of cumulative experience should (1) be global in scope, (2) recognize technological spillovers between associated firms, and (3) account for the common evolutionary heritage of related reactor models. I argue that reactor family, as I define it in Sec. A.X of Appendix A, best fits these criteria.

A global, rather than national, measure of cumulative experience is appropriate because most firms involved in nuclear reactor design and component supply are multinational corporations. Eight of the top ten most successful^ba families have “offspring” in more than one country; these eight families account for 85% of the observations. Experience gained by a firm in one country should, for the most part, be transferable by that firm to the business it does in another country. Furthermore, knowledge disseminates globally through organizations such as the IAEA and the OECD Nuclear Energy Agency.

Firms in the nuclear industry frequently license intellectual property to one another and even collaborate in reactor design, but they tend to do so within small networks that are, for the most part, stable. Reactor type is too broad of a criterion, as it would imply technological spillovers between an American firm like Westinghouse and the Soviet Ministry of Medium Machine Building. Both built PWRs, but due to geopolitics, each firm developed its own PWR design independently.

Reactor model would be too narrow a criterion, because that would imply cumulative experience is entirely forfeited when a firm develops a new model. While economies in serial production of identical models almost surely enhance productivity, I am primarily interested in the learning that has occurred (if any) over the seven decades during which nuclear fission has been deployed for commercial electricity generation. Reactor models are continuously revised and replaced on comparatively shorter timescales. For the 91 models that were built more than once, the average gap between the date on which construction on the first reactor of that model began and the date on which the last reactor of that model began construction was 4.5 years. By comparison, the average reactor takes longer than that to build, at a global average of 7.4 years. This implies learning-by-doing has a very short time window within which to be relevant to other reactors of the same model. Instead, I argue that the benefits of learning-by-doing (if they exist) have the greatest impact on newer models within the same family.

In unreported regressions, I tested whether the effect of cumulative experience is sensitive to defining cumulative experience with a delay period between when construction begins on a reactor $j$ and when “knowledge” is gained for the purposes of reactor $i$. The results were found to be robust to several possible delay periods, but the best model fit was achieved with zero delay. Therefore, I adopt zero delay as my preferred specification.

I transform the raw count of all reactors meeting the inclusion criteria (i.e., having begun construction prior to reactor $i$ and being within the same reactor family) using inverse hyperbolic sine (IHS or $sin{h}^{-1}$) transformation. The IHS of a variable $x$ is approximately equal to $ln(2x)=ln(x)+ln(2)$ for large values of $x$, but for small values of $x$ it differs—chiefly in the fact $arcsin{h}^{-1}(0)=0$, whereas $ln(0)$ is not defined. For several of the observations, it takes on a value of 0 in the measure of cumulative experience. While a more familiar solution is to take the transformation $ln(x+1)$, econometricians recommend IHS (Ref. 108). Bellemare and Wichman¹⁰⁹ provide a brief summary of how to interpret IHS coefficients. When the value of an untransformed variable is greater than 10, IHS coefficients are essentially equivalent in interpretation to the coefficients on log-transformed variables.

B.VI. List of Symbols

Table B.IV provides a list of all symbols for variables used in equations.

TABLE A.I Status of All Reactors in the Database as of April 6, 2021

Status	Count
Planned	76
Planning suspended	55
Planning canceled	172
Under construction	51
Construction suspended	18
Construction canceled	77
Operational	447
Permanent shutdown	191
Total	1087

TABLE B.I Reactors for Which Construction Was Abandoned or Is Presently Suspended

Country	Count	Known or Likely Reasons
Austria	1	National referendum banning nuclear power
Brazil	1	Corruption scandal
Bulgaria	2	Fall of communist regime
Cuba	2	Termination of Soviet aid
Czechoslovakia	2	Fall of communist regime
East Germany	5	German re-unification
Iran	1	Suspended during Islamic Revolution, damaged during Iran-Iraq War
Italy	3	National referendum banning nuclear power
Japan	2	In limbo due to post-Fukushima regulatory environment
North Korea	1	Breakdown of diplomatic agreement
Philippines	1	National executive decision
Poland	2	Fall of communist regime
Romania	3	Fall of communist regime
Russia (post 1991)	1	Energy geopolitics
Spain	4	National legislative decision
Soviet Union	4	Short-term response to Chernobyl disaster (RBMK design or other safety issues)
	14	Fall of communist regime
	1	Suspended after Chernobyl disaster, later restarted; canceled after Fukushima-Daiichi disaster
Taiwan	2	National executive decision
United States	42	Cancellation by utility
West Germany	1	Permission to operate denied by state government
World	95

TABLE B.II Predictors of Construction Suspension and Completion

	(1)^a	(2)^b
GDP per capita	9.0%	−8.3%
One standard deviation increase in $ln(GDPp{c}_{c,y})$	(1.83)	(−1.54)
Democracy	−4.9%	4.8%
One standard deviation increase in $De{m}_{c,y}$	(−1.14)	(1.09)
Decentralization^c	4.2%	−3.6%
One standard deviation increase in $De{c}_{c,y}$	(1.11)	(−0.90)
Investor-owned utility	−7.5%	7.4%
$IO{U}_i$	(−1.74)	(1.66)
TMI accident	−1.2%	1.0%
Under construction on March 28, 1979	(−0.30)	(0.27)
TMI accident $\times$ USA	26.7%	−25.8%
Under construction on March 28, 1979 in the USA	(6.24)	(−6.27)
Chernobyl disaster	14.0%	−12.6%
Under construction on April 26, 1986	(2.14)	(−2.01)
Chernobyl disaster $\times$ USSR	9.4%	−2.3%
Under construction on April 26, 1986 in the USSR	(1.93)	(−0.43)
Fukushima Daiichi disaster	11.2%	−9.1%
Under construction on March 11, 2011	(1.39)	(−1.02)
Regime change	35.2%	−28.8%
Under construction during regime change	(4.80)	(−3.24)
Observations	775	732

^a1 if ever suspended, 0 otherwise.

^b1 if ever completed, 0 otherwise; reactors presently under construction are excluded from column (2).

^cAverage marginal effects on the probability of the outcome in percent; t-statistics in parentheses.

TABLE B.III Relevance of Instruments for Column (1) of Table VIII

	First Stage
	F-Statistic
TMI accident under construction on March 28, 1979	253
TMI accident $\times$ USAUnder construction on March 28, 1979, in the USA	4094
Chernobyl disasterUnder construction on April 26, 1986	65
Chernobyl disaster $\times$ USSRUnder construction on April 26, 1986 in the USSR	479
Fukushima Daiichi disasterUnder construction on March 11, 2011	4
Regime changeUnder construction during regime change	75
Observations	556

TABLE B.IV Abbreviations and Symbols Used in the Econometric Specifications

Abbreviation	Full Description
$L{T}_i$	Lead time of the reactor in months, net of any months of suspended construction
$\widehat{L{T}_i}$	Predicted value of $L{T}_i$, conditional on design characteristics (see Sec. IV.A)
$De{m}_{c,y}$	Index of democracy in country $c$ as of year $y$
$De{c}_{c,y}$	Index of decentralization in country $c$ as of year $y$
$GD{P}_{c,y}$	GDP per capita in country $c$ as of year $y$, in 2011 U.S. dollars
$Spe{c}_{s,i}$	Any of $S$ design characteristics and specification variables for reactor $i$
$M{W}_i$	Net electric capacity (original design rating) of reactor $i$
$OT{C}_i$	Takes on the value 1 if reactor $i$ uses OTC, and 0 otherwise
$P{u}_i$	Takes on the value 1 if reactor $i$ was built primarily for the cogeneration of plutonium for weapons, and 0 otherwise
$IO{U}_i$	Takes on the value 1 if reactor $i$ lead utility was investor owned, 0 if state owned, and 0.5 if of mixed ownership
$NumUti{l}_c$	Number of utilities appearing in the data set for country $c$
$M_i$	A control for measurement error related to multi-unit construction (see Sec. B.II in Appendix B)
${\delta }_t$	Fixed effect for reactor type $t$
${\delta }_f$	Fixed effect for reactor family $f$
${\delta }_m$	Fixed effect for reactor model $m$
${\mu }_c$	Fixed effect for country $c$ (as of the year reactor $i$ began construction)
${\nu }_y$	Fixed effect for year $y$ (year in which reactor $i$ began construction)
${\xi }_y$	Any of several indicator variables that takes on the value 1 if reactor $i$ was under construction during event $x$, and 0 otherwise (see Sec. B.IV of Appendix B)
${\epsilon }_i$	Error term