Innovation, firm survival and productivity: the state of the art

ABSTRACT

We review the theoretical underpinnings and the empirical findings of the literature that investigates the effects of innovation on firm survival and firm productivity, which constitute the two main channels through which innovation drives growth. We aim to contribute to the ongoing debate along three paths. First, we discuss the extent to which the theoretical perspectives that inform the empirical models allow for heterogeneity in the effects of R&D/innovation on firm survival and productivity. Secondly, we draw attention to recent modelling and estimation effort that reveals novel sources of heterogeneity, non-linearity and volatility in the gains from R&D/innovation, particularly in terms of its effects on firm survival and productivity. Our third contribution is to link our findings with those from prior reviews to demonstrate how the state of the art is evolving and with what implications for future research.

KEYWORDS:

• Innovation
• R&D
• survival
• productivity

JEL Classification:

• O30
• O33

Disclosure statement

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

Notes

1 In this respect, several empirical studies have confirmed that R&D expenditures and innovation foster aggregate economic growth (Mankiw, Romer, and Weil 1992; Manjón-Antolín and Arauzo-Carod 1992; Nelson 1993; Daveri 2002; Ortega-Argilés, Piva, and Vivarelli 2014).

2 Our work is also consistent with the tradition of studying firm survival and firm productivity as indicators of post-entry performance, where the selection process leads productive firms to survive and grow while others to stagnate and ultimately exit (Audretsch and Mata 1995; see also next section).

3 We expect such a shift in focus epistemologically – i.e., irrespective of whether the data period in the post-2010 studies cover the post-crisis years. This expectation and the periodisation it informs are based on the observation that there has been an increase in the number of studies that incorporate uncertainty and volatility into their models explicitly (Doraszelski and Jaumandreu 2013; Peters, Roberts, and Vuong 2017a; 2017b; and 2018; Andrew 2020).

4 In Audretsch (1991), the source of the innovative advantage also differs between newly established and incumbent firms. The newly established firms have innovative advantage if information outside of the industry is more important for generating innovative activity. By contrast, the incumbent firms have the innovative advantage if information based on non-transferable experience in the market is more important for generating innovative activity. The source of innovative advantage, however, is not an explicit part of the model. Its effect on the probability of innovation is captured through the relative sizes of the routinised and un-routinised regimes.

5 This assumption is in line with existing evidence on various stock markets including the UK, which indicates that the volatility parameter is usually around one-tenth of the drift parameter (Casas and Gao 2008).

6 As we indicate below, more recent research pays greater attention to heterogeneity in in the effects of innovation on firm survival. A pertinent example is Hyytinen, Pajarinen, and Rouvinen (2015), who draw attention to the extent of heterogeneity in the related literature before reporting their own findings on the relationship between innovation and start-up survival among Finnish firms.

7 A similar concern is raised in Hall, Mairesse, and Mohnen (2010) when they discuss the case for and against the inclusion of industry and time dummies in R&D and productivity models.

8 It must be noted that the studies that control for frailty tend to pay attention only to whether the sign of the estimated coefficients on innovation remain the same between models with and without frailty.

9 This issue arises irrespective of whether frailty is modelled as a multiplicative or additive term in the baseline hazard. If exists, such correlation is a cause of endogeneity. Researchers can address the letter through Mundlak (1978) corrections, which involve augmenting the survival model with time averages the covariates correlated with frailty to ensure mean independence.

10 This issue arises because the shape of the hazard function is unknown and economic theory provides information only about the relevant covariates and their expected effects on the likelihood of firm exit. Stated differently, survival studies tend to estimate a reduced-form hazard model where the logarithm of the hazard is a linear function of two arguments: the baseline hazard function and the covariate function that includes innovation and other covariates suggested by the theory. The PH estimators assume that the baseline hazard function depends on time only whereas AFT estimators assume that it depends on time and the function of the covariates. Hence, the baseline hazard is the same in both estimators only if the covariates are assumed to be zero. Furthermore, the interpretation of the coefficient estimates differs. In the PH models, the coefficients are semi elasticities of the hazard with respect to the covariates, whereas they measure the effect of the covariates on the length of the predicted time until the failure event is likely to occur in the AFT models.

11 The most common case being through labour-saving innovation (Freeman and Soete 1987; Simonetti, Taylor, and Vivarelli 2000; Piva and Vivarelli 2018).

12 The metaphor ‘standing on the shoulders of giants’ is attributed to Isaac Newton and used in work on the economics of innovation to refer to spillovers from investment in innovation. Innovation spillovers play a central role in endogenous growth models, where investors in innovation benefit from ideas embedded in the existing stock of knowledge.

13 Elasticity and rate-of-return estimates based on the knowledge capital model have become known as the primal approach, in contrast to the dual approach based on cost or profit functions. This review excludes the dual-approach studies as the latter are small in number and their model specifications are more varied than the primal-approach studies. A review of the dual-approach studies is provided in Hall, Mairesse, and Mohnen (2010).

14 The CDM model has inspired a large volume or empirical research after its publication in the Economics of Innovation and New Technology (EINT) in 1998. A recent special issue of the EINT (vol. 26, no. 1-2, 2017) celebrates the twenty years of research informed by the CDM model. The special issue features bibliometric and epistemological reviews that locate the CDM model in the wider field of research on innovation and productivity as well as research articles reflecting the state-of-the-art in the specification and estimation of the CDM model.

15 Crépon and Mairesse (1998) estimated the model in two steps using asymptotic least squares (ALS) or minimum distance estimators. In the first step, the reduced-form (auxiliary) coefficients in each equation are estimated separately, taking account of error correlations. In the second step, the information about the auxiliary parameters is used to estimate the structural parameters of interest – mainly the effects of innovation outputs on productivity. When the innovation output measure is continuous, the coefficient estimate is the elasticity of productivity (usually labour productivity) with respect to the innovation output. When the innovation output is measured with an indicator variable, the coefficient estimate indicates the productivity difference between innovative and non-innovative firms.

16 It must be indicated that lack of firm-level price data is an issue in both Griliches-type and CDM-type knowledge capital models.

17 Stated differently, the elasticity estimates may be upward biased if the knowledge production does not take account of the knowledge spillovers from external R&D or cooperative R&D (Cassiman and Veugelers 2002; Piga and Vivarelli 2003).

18 One reason is the mismeasurement of the R&D capital, which is exacerbated when the growth rates of the R&D capital or its deviation from the mean are used. Another explanation is potential multicollinearity between the time effects reflecting autonomous technological change and the growth rates of R&D capital. A third explanation relates to missing data on cyclical variables such as capacity utilisation or person-hour worked instead of headcount employment (Hall, Mairesse, and Mohnen 2010). A fourth explanation is that R&D investment is less responsive to business cycle conditions or policy interventions compared to physical capital. Finally, Bloom (2007) demonstrates that the persistence of the R&D investment series increases as uncertainty increases. Therefore, the within-firm variation in R&D capital is smaller than the between-firm variation; and the explanatory power of the R&D stock series is reduced when the elasticity estimates are based on within estimators.

19 A judicious review of the methodological developments until the cut-off year of 2010 is not feasible here due to space limitations. However, we refer the reader to excellent discussions in Griliches and Mairesse (1995) and Hall, Mairesse, and Mohnen (2010) on model specification, identification and estimations issues in the research field until 2010.

20 The issue of organizational innovation will be discussed at more length below, where we review the recent contributions.