Dualism and payroll shares across US states

ABSTRACT

This paper analyses US payroll share and components across states for the period 1977–2017. Findings include that spatial clustering in payroll shares decreased until the year 2000. States are clustered in low and high productivity groups. We relate this phenomenon to dualism. High labour productivity states featured high payroll shares early on, but now feature low payroll shares. We label this phenomenon decoupling of labour productivity from real wages. A Divisia Decomposition documents that Rust Belt states dominated the decline in US payroll share in the 1980s, whereas more recently large (coastal) states dominated. While we do not explain (potentially different) mechanisms, decoupling is apparent throughout.

KEYWORDS:

• labour share
• US states
• Divisia decomposition
• mixture models

JEL:

• D33
• J31
• J42
• O51

ACKNOWLEDGEMENTS

The authors thank Peter Skott, Thomas Michl and the participants at the Eastern Economic Association Conference, Boston, 2020, for their comments and suggestions.

DISCLOSURE STATEMENT

No potential conflict of interest was reported by the authors.

Notes

1. Dualism in labour productivity could also correspond to a lack of income convergence (Magrini et al., 2015; Ganong & Shoag, 2017; Kinfemichael & Morshed, 2019).

2. There are important differences at the firm level where reallocation of value added to firms with low labour shares is an important source of the decline in the aggregate labour share, and the observed decoupling. While Autor et al. (2020) clearly document the importance of such reallocation, they further interpret it as a largely benign facet of technological change. Such claims appear premature.

3. In 1997, the BEA and BLS transitioned from the Standard Industrial Classification (SIC) system to the North American Industry Classification System (NAICS). Thus, our complete panels are constructed by linking the 1977–97 SIC and 1998–2017 NAICS data. To exclude real estate activities, we take the ‘private industry’ observation for the given series and subtract ‘real estate and rental and leasing’ for NAICS data and the corresponding activities for SIC data based on the concordance tables supplied by the BLS. For employment data, the ‘private industry’ observation must first be constructed by summing ‘private nonfarm wage and salary employment’ and ‘farm wage and salary employment’ before subtracting the real estate term. See Appendix A in the supplemental data online for more details. Because we do not consider sectoral data in more detail, we are not concerned with the general change in industry classification.

4. However, and as will be seen further below, changes in state payroll shares do not always reflect states’ contributions to changes in the national payroll share. This is due to the importance of a state’s weight, in terms of wage bill and value added, in the decomposition method – which is the topic of section on the Divisia decomposition.

5. Heterogeneity in this context can be quantified by a first-order moment such as the mean and, consequently, explored further through differences in means of groups of states. Spatial dimension and, importantly, spatial dependence ‘reflects a situation where values observed at one location depend on the values of neighboring observation’ (Basile et al., 2014, p. 229).

6. We exclude Alaska and Washington, DC, from the analyses throughout this section. Wyoming is also excluded from the bivariate mixture model analysis. For details, see the descriptions accompanying each figure.

7. In this regard, our approach differs from other applications of mixture models to – usually – personal income distribution data. The critical question in that literature is to identify shapes, thresholds and potentially mechanisms that define different quantiles of a distribution. For a discussion, see Schneider and Scharfenaker (2020).

8. The states in the rightward distributions in both cycles are Connecticut, Delaware, Louisiana, New Jersey, New York, Texas, Washington and Wyoming. California, Illinois and Massachusetts join the distribution in the 2000–07 cycle.

9. For reference, the average payroll share over the entire period (1977–2017) is shown with the dashed vertical line. Figure F3 in Appendix F in the supplemental data online presents the same exercise but compares the average payroll share relative with average labour productivity growth rates. Relative average labour productivity is calculated as the ratio of state i annual average labour productivity over the annual unweighted average labour productivity for all states. Average relative labour productivity growth is simply calculated as the ratio of state i percentage change in peak-to-peak labour productivity over the unweighted average of peak-to-peak change in labour productivity for all states. The analysis covers 47 states. In addition to Washington, DC, we have excluded the resource-intensive states of Alaska, Louisiana and Wyoming which are outliers and would have distorted the analysis and data visualization significantly.

10. The exception is panel (b) of Figure 8, which shows average relative labour productivity levels vis-à-vis the payroll share in the cycle 1989–2000. In subsequent discussion, recall that the members of each distribution might be different. Note that Figure F3 in Appendix F in the supplemental data online, which does the same exercise but for average labour productivity growth rates, finds two distinct distributions across all four periods.

11. Between the first and the last business cycle, three (five) states have transitioned to quadrant I in Figure 8 (see Figure F3 in Appendix F in the supplemental data online). These were Colorado, Maryland and Massachusetts, and Colorado, Ohio, Oregon, Pennsylvania and Wisconsin, respectively.

12. As previously alluded to, our results appear to be in line with a growing empirical literature regarding the lack of income convergence between US states in recent decades. Note the simple yet pertinent argument that diverging bimodal cross-sectional distributions as in our uni- and bivariate mixture models are inconsistent with the neoclassical model of balanced growth (Quah, 1993).

13. We perform simple correlation analysis of peak-to-peak values of the segregation ratio and the payroll share. We exclude Louisiana and Wyoming due to their strong outlier status. The correlation coefficient was $-0.13$ in the 1979–89 cycle, and –0.05, –0.26 and $-0.49$ in the three subsequent periods. Note that disaggregated industry employment data under the SIC classification (1977–2001) is available only for full- and part-time employment, including proprietors and individual general partners. See Appendix A in the supplemental data online for further data sources and definitions.

14. Because the sample variances are unequal, we use the Welch approximation t-test, with a threshold p-value of 0.05.

15. Appendix D in the supplemental data online provides more detail on the specific decomposition here.

16. The time interval $[t-n,t]$ can be defined on an annual basis, over a specific period or over peak-to-peak business cycles.

17. journal reviewer summarized this to imply that states with growing labour shares can contribute negatively to the total. This is indeed the case: suppose there were two states that are identical in all variables except nominal wages; call them low-wage TX and high-wage NY. Now suppose that nominal wages rise in TX but remain below wage levels in NY, and that employment shifts to TX. Then the labour share rises in TX, but falls in the aggregate, because the shift to the lower level of wages continues to dominate growth of wages until TX wages exceed NY wages.

18. An aggregate real-wage component contribution of 33.6 percentage points implies – ceteris paribus – an increase in the payroll share of 33.6 percentage points. Other components are interpreted analogously.

19. Given its pro-cyclical behaviour, the payroll share increased by about 1 percentage point between 1977 and 1979. Adding up changes in the aggregate payroll shares across the four cycles leads to an overall decline in the payroll share of 4.1 percentage points between 1979 and 2017.

20. These states have sizable information and finance sectors. In the previous section we provide suggestive evidence of a negative relationship between the presence of such progressive sectors and the labour share. In line with this hypothesis, but approaching the topic from a micro-data angle, Dube et al. (2020) find significant wage markdowns in online labour markets for high productivity/high wage sectors; while Macaluso et al. (2019) identify a robust association between concentration in the labour markets and upskilling and wage compression, although they speak against a significant association between labour market concentration and the decline in the labour share.

Additional information

Funding

The authors gratefully acknowledge financial support from the Institute for New Economic Thinking (INET).