Institutional Quality and Generalized Trust: A Nonrecursive Causal Model

See Uslaner (2008) for a similar argument. In this article, Uslaner shows that current levels of trust in the US can, to a large extent, be traced back to the 1930s and 1940s. This suggests that at least a part of the trust observed is prior to contemporary government. Moreover, considering that trust tends to be fairly stable across time (e.g., Bjørnskov 2007), then the trust-institutions relationship identified in prior research is, at most, bidirectional and, at least, unidirectional from trust to political-institutions.

The dataset has individual-level information for 66 countries (when considering missing data and the variables I use).

I use multiple imputation techniques found in Stata 10.1 to maintain statistical power and a sizable country level sample. I imputed data for the legal property rights (Belarus, Bosnia, Moldova, and Saudi Arabia) and income inequality (Bosnia, Malta, Netherlands, Saudi Arabia, Serbia, and Slovakia) measures. Note that none of the imputed variables have greater than 10 percent missing cases. Also note that I created 1,000 complete data sets with the missing values filled in with different imputations. The values for the missing data were the mean of the 1,000 values across these data sets. Unlike traditional multiple imputation techniques, I did not take into account uncertainty as represented by the variation across the multiple imputations for each missing value since EQS does not permit such a procedure. I used this procedure instead of the maximum likelihood procedure found in EQS to reduce model complexity.

I ignore particularized trust in the present analysis since the bulk of research in this area is primarily concerned with investigating the relationship between political institutions and generalized trust. For recent research exploring the determinants of particularized trust see Freitag and Traunmüller (2009), Glanville and Paxton (2007), Gleave et al. (2011), and Radnitz et al. (2009).

I also focus on elements of government dealing with fairness and effectiveness since the results of alternative investigatory and confirmatory factor analyses suggest that measures of government should be treated as three separate dimensions: fairness and effectiveness, power-sharing capacity, and universality. This indicates that including, for instance, (a) the Polity IV measure of democracy and the Freedom House measure of political rights (i.e., power-sharing capacity), (b) the World Bank measure of public health expenditures and income inequality (i.e., universality), and (c) the legal property rights and rule of law measures (i.e., fairness and effectiveness) into one dimension is unwarranted. In fact, these indicators should be used only for their respective dimensions. Results available upon request.

It is often more desirable to use tetrachoric or polychoric correlation matrix estimation techniques instead of maximum-likelihood with dichotomous scaled data (Nunnally and Bernstein 1994). This is especially the case for CFAs and only the case for SEMs if the categorical measures are endogenous. Since the categorical indicators in the present article are exogenous, I conducted an alternative CFA with a polychoric correlation matrix estimation procedure. The alternative results parallel those presented here. As a result, I present the maximum-likelihood CFA estimates; results available upon request.

Although small samples are common in the SEM literature (see MacCallum and Austin 2000), there is little consensus on recommended sample sizes. Kline (2005) notes that “…with less than 100 cases, almost any type of SEM analysis may be untenable unless a very simple model is evaluated” (p. 15). In other words, technical problems, such as non-convergence, and issues of statistical power are more likely to occur with small samples. Note, however, that convergence and maximum likelihood solutions were not an issue in any Table 2 model; all coefficients in our final model were statistically significant (see Table 3); and all evaluated models were simple. This suggests that sample size likely did not bias the present findings.

This is a routine practice in the SEM literature to assume zero measurement error for single indicator factors, especially when there are no prior estimates of measurement error in the literature to abstract and assume a reasonable non-zero measurement error. I did, however, analyze the models with varying levels of assumed measurement error estimates, from 0.01 to 0.3, for both generalized trust and monarchy. As expected, measurement error in generalized trust produced underestimation of the β coefficient. Also, as expected, measurement error in monarchy above 0.05 produced weak instrument effects, resulting in either biased estimates, lack of bidirectional or even unidirectional significance between generalized trust and institutional quality, or maximum likelihood convergence issues. These results suggest that some minor unobserved measurement error in the generalized trust and monarchy indicators will not bias the results presented below. For instance, an assumed measurement error of 0.02 for both indicators yielded results similar to those found in Fig. 1.

For model 1, Table 2, I also controlled for gross domestic product, which did not significantly influence generalized trust, but it did, however, affect other relationships in the model. It was highly correlated with both the information technologies and institutional quality dimensions (r > 0.77), which resulted in discriminant validity issues and difficulties in converging on a solution associated with the small sample size (n = 64). As a result, I left gross domestic product out of the analysis since both information technologies and institutional quality capture a large portion of its variance. Results are available upon request.

I analyzed a number of alternative models to further test the sensitivity of the results. First, some of the variables, specifically legal property rights and income inequality, had less than 10% missing cases. In the original analyses, I used multiple imputation techniques found in Stata 10.1 to overcome this issue. To see if the imputed data may have biased the results, I re-analyzed the models in Table 2 using listwise deletion with the missing cases, which yielded an N of 58. The results indicated that none of the key path coefficients deviated from those presented in Tables 2 and 3. Second, I also explored model-based imputation methods available in EQS 6.1 (Bentler 2003). This method replaces a missing score with an imputed value drawn from a full information maximum-likelihood predictive distribution. Although models 2 through 5 had difficulty converging on a solution, the model in Fig. 1 converged, which paralleled the significant positive feedback effect found therein. Third, and finally, I investigated the pairwise deletion option found EQS 6.1, which did not substantively alter the results presented here.

To explore whether 2SLS or generalized method of moments (GMM) is more appropriate for the following instrument validity tests, I used the ivhettest found with ivreg2 in Stata 10.1. Both tests failed to reject the null hypothesis that the disturbance terms are homoskedastic (information technologies instrument, p = 0.11; monarchy instrument, p = 0.88). This indicates that the use of classic 2SLS is efficient and robust.

I use the following syntax in Stat 10.1 for ivreg2 and ivregress, respectively: ivreg2 generalized_trust monarchy (institutional_quality = information_technologies) ivregress 2sls generalized_trust monarchy (institutional_quality = information_technologies).

I use the following syntax in Stat 10.1 for ivreg2 and ivregress, respectively: ivreg2 institutional_quality information_technologies (generalized_trust = monarchy) ivregress 2sls institutional_quality information_technologies (generalized_trust = monarchy).