Baseline Results Empirical Results

The Journal of Human Resources 628 gender. It is important to note that the measure of income we investigate relates to the household, but both the attitude indicators and sociodemographic variables refer to the respondent. We will have more to say later about how that affects the interpreta- tion of our results.

V. Empirical Results

A. Baseline Results

Table 2 reports our baseline results. 6 It presents the results of regressing the logarithm of real household income against variables that are standard in micro earnings equa- tions plus indicators of the individual’s level of trust and a second-degree polynomial of the individual’s level of trustworthiness, sometimes interacted with the mean and squared mean level of these variables in the respondent’s country. Because of scaling 6. The regressions are calculated using observations unweighted within countries and with sums of weights equalized across countries. We have also estimated analogous regressions with weighting within countries and sums of weights equalized across countries. None of the principal results reported in this section are affected by this change. Average Trust = .523−.021Average Trustworthiness Standard Error on Slope = .046; Correlation coefficient = -.084 Average Trust Average Trustworthiness 7 8 9 .2 .4 .6 Mexico Belgium West Germany Portugal Netherlands South Africa Spain Finland Britain Canada Brazil Austria Italy USA Chile Turkey Japan India Figure 1 Average Trust versus Average Trustworthiness Slemrod and Katuák 629 concerns raised by the fact that the trustworthiness measure is a 1 to 10 ordinal vari- able and to allow for a nonlinear impact on income, we include a second-degree polynomial in trustworthiness. All of the regressions include country fixed effects coefficients of which are not reported here, and therefore all of the estimated coef- ficients are identified from within-country variation only. The standard errors are adjusted for clustering at the country level. The specification in Column 1 contains only the standard variables in an earnings equation. The marginal return to the respondent’s education level is always positive within the observed range between 6 and 15 years, although decreasing. Based on the estimated coefficients, going from zero to ten years of education adds 87 percent to income. 7 Furthermore, the marginal return of a year of education is 11.2 percent per year at 0 years, and falls to 6.23 percent at 10 years. These results are in line with the empirical literature, 8 lending credence to the survey-based measures of income and 7. In this and all subsequent result interpretations, we use a first-order approximation around one to changes in the logarithmic transformation of the dependent variable. 8. In the human capital earnings approach standard in labor economics, more recent estimates of the return to education fall anywhere between 0.023 Isacsson 1999 and 0.153 Harmon and Walker 1995 per addi- tional year of schooling, depending on the dataset used, the set of control variables and the econometric tech- nique. Card 1999 provides a good summary of this literature. Our marginal effect estimates lie within this range. GDP Per Capita = 1.023+ 32.098Average Trust Standard Error on Slope = 5.016; Correlation coefficient = .731 GDP Per Capita In 000 1990 Average Trust .2 .4 .6 10 20 Brazil Turkey Portugal Chile South Africa Austria Belgium Mexico Spain India Italy West Germany Japan Britain USA Canada Netherlands Finland Figure 2 GDP Per Capita versus Average Trust The Journal of Human Resources 630 education. The respondent’s age, which is undoubtedly partly a proxy for work expe- rience, initially has a positive impact on income, but its impact peaks at age 38.9, and it has a marginal negative effect thereafter. The marginal return falls from 1.82 per- cent per year at age 20 to −2.99 percent per year at age 70. 9 Households with male respondents have an 8.73 percent higher income. 10 Although this is lower in absolute value than the typical finding in the literature, 11 the difference is unsurprising as our results relate to household income rather than the respondent’s income. In the next subsection, we restrict the sample to include only those households in which the 9. Angrist and Krueger 1999, using 1990 Census and March 1990 CPS samples, report coefficient esti- mates on potential experience around 0.041 per year for the Census data and 0 or 0.013 for the CPS data, depending on whether they do or do not use the allocated CPS values. The coefficient estimates on potential experience squared are from −0.00057 to −0.00055 for the Census data and they are statistically insignifi- cant for the CPS data. Our results are similar to these estimates. 10. All the reported estimates are statistically significant at the 95 percent confidence level unless stated otherwise. 11. Altonji and Blank 1999, using the Current Population Survey CPS data, estimate the coefficient on a female indicator variable to be −0.421 in 1979 when the additional controls are education, experience and region, and −0.348 when occupation, industry, and job characteristics are controlled for as well. In 1995, these estimates are −0.272 and −0.221, respectively. When using the National Longitudinal Survey of Youth data from 1994, the coefficient estimates on the female indicator variable are approximately −0.24 to −0.20. Figure 3 GDP Per Capita versus Average Trustworthiness GDP Per Capita = 22.157−1.213Average Trustworthiness Standard Error on Slope = 2.802; Correlation coefficient = −.111 GDP Per Capita In 000 1990 Average Trustworthiness 7 8 9 10 20 Mexico Belgium West Germany Portugal Netherlands South Africa Spain Finland Britain Canada Brazil Austria Italy USA Chile Turkey Japan India Slemrod and Katuák 631 major earners are the respondents and observe a much larger estimated male-female income differential. Columns 2 through 4 show the estimated return to individual trust and trustworthi- ness, ignoring any country-level interaction effects. The results suggest that trust, but not trustworthiness, is associated with higher income. Trusting as opposed to not trusting increases income by 7.59 percent. The differentially approximated marginal effect of moving up one trustworthiness category at the mean level of trustworthiness of 8.06 12 is −1.54 percent. Column 4 shows that the point estimate for the return to trust is not notably changed by including the trustworthiness variables. Although the individual point estimates for the latter variables do change, there is little effect on the marginal return to moving up one trustworthiness category at the mean, which falls slightly to −1.59 percent. Columns 5 and 6 present our central results that allow for interaction between per- sonal characteristics and country-level means of trust and trustworthiness. Column 5 reveals that at the mean level of trustworthiness of 8.06, the return to trust is 8.59 per- cent of income. This estimate is similar to the estimates from the previous specifica- tions. The estimated marginal effect on this return of a one-category increase in the country average level of trustworthiness at the mean level of trustworthiness is posi- tive at 5.16 percent, increasing to the first order the return to trust from 8.59 percent to 13.75 percent; the interaction term is not, however, statistically significant. 13 Therefore there is some indicative evidence that the payoff to trust increases with the average level of trustworthiness in the country, although not at conventional levels of statistical significance. 14 The return to trust is negative in countries that have the aver- age trustworthiness below 6.98, and is positive in countries with the average trust- worthiness above that. All of the countries in our sample, except for Mexico, have means of trustworthiness greater than 6.98, and hence the return to trust is positive in almost all of the countries in our sample. Column 6 reveals that at the mean level of trust of 0.354 and of trustworthiness of 8.06, the marginal return to a one-category increase in trustworthiness is −1.56 per- cent of income, which is similar to the estimates from the previous specifications. The estimated marginal effect on this return of an increase in the country average level of trust from no to full trust is positive at 6.49 percent. 15 This is a striking finding: The 12. This mean and the analogous measure for trust used later are calculated using the sample weights, with sums of weights equalized across countries. 13. Let , , 1 a a t t and 2 at be the point estimates on trust, trust interacted with the average level of trust- worthiness, and trust interacted with the square of the average level of trustworthiness, respectively. Then the estimated return to trust is given by TW TW 1 2 2 + + a a a t t t , where TW is the average level of trustworthiness in the country. The marginal effect of an increase in TW on this return is then given by TW 2 1 2 + a a t t . We evaluate these measures at the mean level of trustworthiness of 8.06. 14. This estimated marginal effect is statistically significant at the 95 percent level if standard errors are not clustered at country level. This approach would be justified in the case when the only source of correlation in the error term is captured by country fixed effects. 15. Let , , 1 2 b b b t t t , and 3 bt be the point estimates on trustworthiness, trustworthiness squared, trustworthi- ness interacted with the average level of trust, and trustworthiness squared interacted with the average level of trust, respectively. Then the estimated marginal return to trustworthiness is given by TW T T TW 2 2 1 2 3 + + + b b b b t t t t , where T is the average level of individual trust in the country, and TW is the individual level of trustworthiness, respectively. The marginal effect of an increase in T on this return is then given by TW 2 2 3 + b b t t . We evaluate these measures at the mean level of trust of 0.354, and the mean level of trustworthiness of 8.06. The Journal of Human Resources 632 Table 2 Regressions Results: The Impact of Trust and Trustworthiness on Real Household Income Dependent Variable: Log of Real Household Income 1 2 3 4 5 6 7 Education 0.1118 0.1119 0.1094 0.1109 0.1123 0.1065 0.1086 0.0696 0.0700 0.0694 0.0695 0.0699 0.0687 0.0687 Education squared −0.2473 −0.2542 −0.2368 −0.2507 −0.2568 −0.2244 −0.2411 ×10 −2 0.2797 0.2827 0.2783 0.2794 0.2815 0.2745 0.2749 Age 0.0374 0.0369 0.0380 0.0377 0.0368 0.0381 0.0377 0.0076 0.0076 0.0073 0.0073 0.0076 0.0073 0.0072 Age squared −0.0481 −0.0476 − 0.0484 −0.0481 −0.0476 −0.0486 −0.0483 ×10 −2 0.0076 0.0076 0.0074 0.0074 0.0076 0.0073 0.0073 Male 0.0873 0.0899 0.0857 0.0888 0.0903 0.0871 0.0907 0.0212 0.0213 0.0213 0.0212 0.0213 0.0212 0.0212 Trust 0.0759 0.0785 −1.9933 −1.7753 0.0177 0.0174 3.1070 3.0512 Trustworthiness 0.0642 0.0411 0.0246 0.2021 ×10 −1 0.3337 0.3229 0.9424 0.8412 Trustworthiness squared −0.1354 −0.1244 −0.3139 −0.4358 ×10 −2 0.2093 0.2042 0.6517 0.5777 TrustworthinessAv.trust 0.3782 −0.2493 ×10 −1 2.2750 1.9548 Trustworthiness squaredAv.trust 0.3360 0.7515 ×10 −2 1.6046 1.3784 Slemrod and Katu ák 633 TrustAv.trustworthiness 4.6434 4.1512 ×10 −1 8.0593 7.9014 TrustAv.trustworthiness squared −2.5604 −2.2856 ×10 −2 5.2039 5.0943 Country fixed effects Yes Yes Yes Yes Yes Yes Yes Return to trust 0.0759 0.0785 0.0859 0.0858 ∂Return to trust∂Av. 0.0516 0.0467 trustworthiness Marg. return to trustworthiness −0.0154 −0.0159 −0.0156 −0.0160 ∂Marg. return to trustworth.∂Av. 0.0649 0.0356 trust Observations 26,046 24,544 25,675 24,235 24,544 25,675 24,235 R-squared 0.42 0.42 0.42 0.42 0.42 0.42 0.42 Within Country R-squared 0.09 0.09 0.09 0.09 0.09 0.09 0.09 Note 1: Observations unweighted within countries, with the sums of weights equalized across countries. Note 2: Standard errors adjusted for clustering at country level in parentheses. Significance level notation: at 10 percent, at 5 percent, at 1 percent. payoff to trustworthiness increases with the average level of trust in the country. At the mean level of trustworthiness, the marginal return to a one-category increase in trustworthiness is negative in countries that have the average trust below 0.523, and is positive in countries with the average trust level above that. In contrast to Column 5, in all but three countries the mean level of trust falls short of this figure, with the three exceptions being Canada, Finland, and the Netherlands. Thus, our results suggest that trustworthiness is in most countries not rewarded with higher income—dishonesty pays. How much it pays varies widely. In a very low-trust country like Brazil, a one- category increase in trustworthiness at that country’s mean level of trustworthiness is associated with an 4.28 percent decrease in income. Column 7 shows that the principal results from Columns 5 and 6 are unchanged when both attitude variables and both interaction terms are included in the same equa- tion. In particular, at the mean level of trustworthiness of 8.06, the return to trust is 8.58 percent of income, and the marginal effect on this return of a one-category increase in the country average level of trustworthiness at the mean level of trustwor- thiness is positive at 4.67 percent, but it is not statistically significant. 16 The return to trust is negative in countries that have average trustworthiness below 6.89, and is pos- itive in countries above that. On the other hand, at the mean level of trust of 0.354 and of trustworthiness of 8.06, the marginal return to a one-category increase in trustwor- thiness is −1.60 percent of income, and the estimated effect on this return of an increase in the country average level of trust from no to full trust is 3.56 percent. At the mean level of trustworthiness, the marginal return to a one-category increase in trustworthiness is negative in countries that have average trust below 0.52, and is pos- itive in countries with an average trust level above that. As suggested earlier by the gender differential estimates, a potential problem with these results is that we use household real income as a dependent variable and respon- dent attitudes and demographic characteristics as independent variables. However, it has been shown see, for example, Mare 1991 and the references contained therein that most married or cohabiting couples are characterized by assortative matching by edu- cation, age and many other characteristics, thus lending more credibility to our results. Another potential problem is that we restrict the coefficient estimates to be the same across all the countries. It is also possible that trust and trustworthiness are more impor- tant for the income of households with self-employed respondents than other respon- dents, because the former are more directly exposed to transactions with strangers rather than long-term employment relationships. Finally, trust and trustworthiness may be endogenous to income. We address all of these issues in the next subsection.

B. Robustness Analysis