Determinants of institutional quality: an empirical exploration

ABSTRACT

Recent literature has underlined the role that institutions play in the process of development, making it essential to understand why differences exist in the quality of institutions across countries. The goal of this study is to investigate the determinants of institutional quality. Our results confirm that institutional quality is conditioned by variables that can be modulated by public policy, such as income per capita, international openness, education, taxation, and patterns of income (re)distribution. Our conclusions differ from the pessimistic outlooks of works highlighting deterministic factors, such as colonial or geographical factors, as determinants of institutional quality.

KEYWORDS:

• Institutional Quality
• development
• inequality
• tax Revenue
• fiscal Covenant
• international Openness

JEL CLASSIFICATION:

• D02
• O4
• O10

Acknowledgements

Not applicable

Disclosure statement

No potential conflict of interest was reported by the authors.

Supplementary Material

Supplemented data of this article can be accessed here.

Notes

1. Here IV 2SLS stands for the Instrumental Variables 2-Stage Least Squares.

2. These structural approaches are compatible with the role that in some cases purposively motivated actors can play in promoting institutional change. See, for example, Teorell and Rothstein (2015) and Rothstein and Teorell (2015).

3. The pooled model (with equal constants for different countries) is statistically rejected by the F-test (as well as by the Breusch-Pagan Lagrange multiplier test, if random effects were relevant), as well as the model with random effects by the Hausman test. These results are available upon request.

4. At the same time, in the panel framework like ours that relies on the asymptotic increase of the number of cross-sections (large N) with fixed number of periods (T), a consistent estimation of $\{a_j\}$ cannot be obtained from the estimated residuals in the FE framework, thus precluding the implementation of the iterative Cochran-Orcutt-like procedure to get more efficient estimates.

5. Typically, we fix the number of instruments to 70 constituting around 70% of the total number of cross-sections (that ranges from 94 to 114 in specifications with different variables). In sensitivity analyses, we furthermore vary this share from 20% to almost 90%.

6. These two principal components explain about 80% of total variation of explanatory variables in both cases. Furthermore, as will be seen shortly, these two components turn out to be significant, whereas the remaining components are not.

7. Namely, these principal components are formed only from the explanatory variables without including the lags of the dependent variable.

8. Formally, such inference is somewhat optimistic as instead of unknown true PCs their estimates were used, but the estimated loadings in PC1 are of fairly similar size, thus either all variables are significant, or all are insignificant, which was seen not to be the case already in Table 1.

9. Since the Gini index (of net income) is employed.

10. It must be noted that we also tested for the presence of decreasing returns and cross-effects among the determinants of institutional quality, but we found no evidence supporting them.

11. Using the specification as in column (3) of Table 4, Table A6 presents additional results with the number of instruments ranging in columns (1)–(5) between (approximately) 50% and 90% of the number of countries. Further reduction of instruments to barely 20% can be achieved without large changes of estimated coefficients (see columns (6)–(9) in Table A6) provided one accepts the usage of clustered period effects (into a few groups with most similar estimated values) instead of the unrestricted 20.

12. It should be noticed that the number of GMM instruments is accordingly reduced by the increased number of regular instruments in order to avoid the increase in the total number of instruments.

13. Namely from the World Income Inequality Database (WIID) that does not have arbitrary imputations (see Jenkins 2015) instead of the Standardized World Income Inequality Database (SWIID, see Solt 2019).

14. We thank the reviewers for this suggestion.

15. The specification with random effects was also investigated (rejecting its adequacy) with the results similar to those reported for the pooled model and not included here (available from the authors).

16. Given a not so small effective average T ≈ 16, one could expect––we thank a reviewer for this observation––the FE bias in dynamic panels to be rather moderate, although Judson and Owen (1999) warn to be careful even with T = 20. However, in our case the errors of both the static and dynamic models are (highly) serially correlated as revealed by the Wooldridge (2002) test. In such a situation, the static model still can be consistently (although highly inefficiently, in our case) estimated with fixed T as N → ∞ using the proper OLS/FE choice, whereas the dynamic model cannot be consistently estimated (even under T → ∞). Nevertheless, the FD-based estimation has somewhat smaller (relative to FE) serial correlations of errors with p-values just above 0.01 (compare columns (6)-(7) and (9)-(10) in Table A11).

17. The (system) GMM estimator has many advantages, but the results can be sensitive to the set of alternative instruments, even if all of them were admissible. Additional sensitivity results with varying sets of instruments are available from the authors upon request.

18. Although this can induce some pre-estimation bias of statistical inference, despite that the estimated unconstrained period effects are only used for the grouping and, therefore, such effect is likely to be small.