Over the last two decades, empirical work has tried to explain why some countries have had rapid long-term growth rates in income while others have not. After a period during which the neoclassical Solow (1956) framework was the workhorse of empirical growth analysis, endogenous growth theory introduced alternative models that allow growth to be generated by factors other than exogenous technical change. Endogenous growth theory provided mechanisms through which economic and social policies could affect growth through their effects on human and physical capital accumulation. Consequently, empirical work on growth that ensued extended the neoclassical model to include a number of determinants that are partially correlated with growth, including proxies for government policies and measures of technology diffusion.

A fundamental problem confronting researchers is the lack of an explicit theory identifying the determinants of growth. Indeed, extensions to the neoclassical and endogenous growth models are what Brock and Durlauf (2001) call “open-ended,” as they admit a vast range of logical and testable additions, and a broad number of possible specifications. In fact, a survey of the empirical growth literature by Durlauf, Johnson, and Temple (2005) identifies over 140 proxies of growth determinants put forward by various empirical studies, highlighting the “open-endedness” of growth theories and implicitly the degree of uncertainty surrounding the validity of the competing theories. As a result, researchers began to investigate how robust empirical relations for economic growth are.

Work on investigating the robustness of growth determinants was initiated by Levine and Renelt (1992) and Sala-i-Martin (1997). The former approach—using a version of the extreme bounds analysis introduced by Leamer (1985)—labeled a few variables as robust but was criticized for its restrictiveness. The latter approach identified a relatively large number of robust variables and was criticized for the simplifying assumptions of a fixed model size and the existence of a set of “fxed regressors” appearing in each specification. While these studies were important initial attempts to shed light on the robustness of growth determinants, they did not fully take model or theory uncertainty into account.

Following the early work on investigating the robustness of growth determinants, Bayesian Model Averaging (BMA) techniques were introduced in the context of growth empirics. The BMA techniques—advanced through the work of Raftery (1995)—provide a conceptually attractive solution to the problem of model uncertainty. These techniques assume that the researcher does not know which model is “true” and thus needs to attach probabilities to different possible models. Inferences are then based on a weighted average of the full model space instead of on one selected model, thus incorporating uncertainty in both predictions and parameter estimates.

Fernández, Ley, and Steel (2001), Brock and Durlauf (2001), and Sala-i-Martin, Doppelhofer, and Miller (2004) formally introduced model averaging to the growth empirics literature. While their methodologies differ—the inference in Fernández, Ley, and Steel (2001) and Brock and Durlauf (2001) is based on BMA, while Sala-i-Martin, Doppelhofer, and Miller (2004) advocate making inferences based on a selected group of variables—their results are similar. All these studies find that initial level of income is important in determining growth along with some measures of human capital, some sectoral variables, and regional dummies.

More recent applications of BMA to investigate growth empirics suggest several modifications of the early BMA framework, such as testing the strength of various growth theories instead of concentrating on the individual explanatory variables. In addition, within the context of addressing model uncertainty, researchers began to address other issues that may plague the study of growth empirics, such as including omitted country-specific effects and incorporating heterogeneity, modeling dynamics, and endogenous variables.

In an attempt to test growth theories rather than particular variables, Durlauf, Kourtellos, and Tan (2008) assign priors to various combinations of empirical proxies for a given theory. They find little evidence for the fundamental growth theories of geography and institutions and strong evidence for macroeconomic policy and regional heterogeneity in explaining aggregate growth. In addition, Ley and Steel (2007) and Doppelhofer and Weeks (2009) develop measures of “jointness” to examine whether explanatory variables in growth regressions act as complements or substitutes. Ley and Steel (2007) find evidence of jointness between some determinants of growth—suggesting that they have a separated role in explaining growth and that they should appear jointly in the regressions—and more frequent situations of “disjointness,” where regressors are substitutes and thus should not appear together. In contrast, using a different measure for jointness, Doppelhofer and Weeks (2009) find an important role for jointness among growth determinants.

Rather than modeling heterogeneity as a fixed effect (e.g., by adding a dummy variable) BMA approaches incorporate parameter heterogeneity in the estimation. Brock and Durlauf (2001) allow African countries to have different growth parameters than the rest of the world, and they find evidence of heterogeneity through different coefficient estimates. In addition, Masanjala and Papageorgiou (2008) investigate growth determinants in Africa using BMA and find that initial conditions such as initial primary education and primary resources and geography can explain a significant portion of the differences in Africa's growth from the rest of the world. Finally, Tsangarides (2005) using a new BMA methodology, finds evidence that what is good for growth around the world is also good for growth in Africa, although the marginal impacts vary.

Most of the work using BMA to address model uncertainty in the context of growth has been in the form of cross-country regressions using static models, with variables of interest essentially averaged over the period of analysis. However, recent work began to model dynamics in the context of BMA by exploring the use of panel data in the context of model uncertainty. In addition to increasing the amount of observations available through the within-country variation, the use of panel data captures the dynamic evolution of the growth process and offers the possibility to account for heterogeneity, and control for (or estimate) country-specific effects. In an attempt to model heterogeneity in the context of a panel BMA, Moral-Benito (2009) considers a panel-data model where the lagged dependent variable is correlated with the individual effects.

A common issue in growth empirics is that many explanatory variables are endogenously determined in an economic sense. This, in turn, implies a strong chance that they are endogenous in the statistical sense, that is, correlated with the disturbance term and hence leading to inconsistent estimates. Tsangarides (2004) and Chen, Mirestean and Tsangarides (2009) address the issue of endogeneity in a panel-data context by proposing a new limited information BMA (LIBMA) methodology based on generalized methods of moment (GMM) estimation that they apply to investigate growth determinants. Durlauf, Kourtellos, and Tan (2008) construct instruments for variables that are endogenously determined in the economic sense and introduce a model-averaged version of two-stage least squares. Eicher, Lenkoski, and Raftery (2009) develop formal statistical foundations for an instrumental variable BMA (IVBMA) methodology to address model uncertainty in the presence of endogeneity. Once endogeneity is taken into account, Durlauf, Kourtellos, and Tan (2008) and Mirestean and Tsangarides (2009) find support for the canonical neoclassical growth theory as well as for some macroeconomic policies.

In the continuing investigation of the empirics of growth, increasing attention is being given to the implications of model uncertainty. A growing number of growth researchers are turning to BMA methods, which provide a solid theoretical foundation for addressing model uncertainty. While there is a growing literature focusing on improving and refining the BMA techniques—particularly the impact of the choice of the priors—the work on BMA and its applications has underscored that failing to properly account for model uncertainty results in overconfident and often fragile inferences. This has important implications for policymakers seeking to use findings of growth analyses to offer policy advice, suggesting that policy analysis and recommendations should not be conditioned on a specific model but rather should reflect model uncertainty.