Education and International Trade 477 currently in paid work include a varied set of individuals,such as those that are unemployed,students,and homemakers(and may be seeking paid work or plan to seek to work soon),we have also isolated one particular group-those individuals who are retired-who are highly unlikely to reenter paid work in the future and be concerned about how their(potential)wages might be affected by trade.32 The key results from the estimations are reported in Table 1,which displays the estimated effects of education on individual trade preferences in the full NES sam- ple and in each of the different subsamples.To facilitate comparison across sub- samples,rather than showing estimated probit coefficients,we report estimated marginal effects:that is,the change in the probability of favoring protectionism associated with an infinitesimal change in sCHOOLING(for the specific dummy vari- ables for levels of highest educational attainment,the discrete change in the prob- ability is shown). Comparing the results across the subsamples,we find little difference in the estimated effects of education on attitudes toward trade.In all cases,the estimated effects of SCHOOLING are similar,both in terms of magnitude and level of statisti- cal significance,across all models(none of the coefficients is significantly differ- ent from the others across subsamples at conventional levels).This is true for estimations using both the 1992 and the 1996 NES data.For example,in the case of the 1996 survey,using the extensive set of covariates,a change from zero to seventeen years of sCHOOLING(while holding the other covariates at their respec- tive sample means)is associated with an average decrease in the probability of favoring protection of about 0.59(s.e.0.05)for the full sample,0.51 (s.e.0.09) for those currently in paid work,0.48(s.e.0.11)for those currently not in paid work,and 0.53(s.e.0.11)for those who are retired (Models 9 to 12,Panel B).33 The observed relationship between education and trade preferences becomes even more similar across subsamples once we replace the sCHOOLING measure with the separate education dummies.For example,compared to individuals with less than junior high-level educations,completing a CoLLEGE education decreases the probability of being in favor of protection by about 0.28(s.e.0.08)for the full include those few retired/students/disabled/homemakers who also indicated that they are "currently working more than 20 hours per week"in our"currently in paid work"subsample.When the latter are excluded from the"currently in paid work"subsample,the magnitudes of the schooling effect become, if anything,more similar across the in-and out-of-paid-work subsamples than in the results we show here.Full results of these robustness tests are available on request. 32.While pensions for retired workers in some prominent U.S.industries (for example,steel)have been linked to the financial health of their former employers,this is the exception and not the rule. Recent studies of U.S.retirees indicate that less than 17 percent of retirement income in the median household comes from employer-provided pension plans(see Sass 2003,6;and Social Security Admin- istration 2002).The connection between employer-provided pensions and the financial health of the firm also is attenuated by the standards for funding and fiduciary conduct established by the Employee Retirement Income Security Act in 1974(see Sass 1997).On this issue,we might also note that we get identical results when we perform the same tests comparing retirees with workers using the ISSP data (see below),drawn from a variety of countries with a variety of pension and retirement income systems. 33.Predicted effects here,and below,are calculated using the"Clarify"software developed by King, Tomz,and Wittenberg 2001.For each such calculation,all other covariates are set at the sample mean values.Education and International Trade 477 currently in paid work include a varied set of individuals, such as those that are unemployed, students, and homemakers (and may be seeking paid work or plan to seek to work soon), we have also isolated one particular group-those individuals who are retired-who are highly unlikely to reenter paid work in the future and be concerned about how their (potential) wages might be affected by trade.32 The key results from the estimations are reported in Table 1, which displays the estimated effects of education on individual trade preferences in the full NES sam￾ple and in each of the different subsamples. To facilitate comparison across sub￾samples, rather than showing estimated probit coefficients, we report estimated marginal effects: that is, the change in the probability of favoring protectionism associated with an infinitesimal change in SCHOOLING (for the specific dummy vari￾ables for levels of highest educational attainment, the discrete change in the prob￾ability is shown). Comparing the results across the subsamples, we find little difference in the estimated effects of education on attitudes toward trade. In all cases, the estimated effects of SCHOOLING are similar, both in terms of magnitude and level of statisti￾cal significance, across all models (none of the coefficients is significantly differ￾ent from the others across subsamples at conventional levels). This is true for estimations using both the 1992 and the 1996 NES data. For example, in the case of the 1996 survey, using the extensive set of covariates, a change from zero to seventeen years of SCHOOLING (while holding the other covariates at their respec￾tive sample means) is associated with an average decrease in the probability of favoring protection of about 0.59 (s.e. 0.05) for the full sample, 0.51 (s.e. 0.09) for those currently in paid work, 0.48 (s.e. 0.11) for those currently not in paid work, and 0.53 (s.e. 0.11) for those who are retired (Models 9 to 12, Panel B).33 The observed relationship between education and trade preferences becomes even more similar across subsamples once we replace the SCHOOLING measure with the separate education dummies. For example, compared to individuals with less than junior high-level educations, completing a COLLEGE education decreases the probability of being in favor of protection by about 0.28 (s.e. 0.08) for the full include those few retired/students/disabled/homemakers who also indicated that they are "currently working more than 20 hours per week" in our "currently in paid work" subsample. When the latter are excluded from the "currently in paid work" subsample, the magnitudes of the schooling effect become, if anything, more similar across the in- and out-of-paid-work subsamples than in the results we show here. Full results of these robustness tests are available on request. 32. While pensions for retired workers in some prominent U.S. industries (for example, steel) have been linked to the financial health of their former employers, this is the exception and not the rule. Recent studies of U.S. retirees indicate that less than 17 percent of retirement income in the median household comes from employer-provided pension plans (see Sass 2003, 6; and Social Security Admin￾istration 2002). The connection between employer-provided pensions and the financial health of the firm also is attenuated by the standards for funding and fiduciary conduct established by the Employee Retirement Income Security Act in 1974 (see Sass 1997). On this issue, we might also note that we get identical results when we perform the same tests comparing retirees with workers using the ISSP data (see below), drawn from a variety of countries with a variety of pension and retirement income systems. 33. Predicted effects here, and below, are calculated using the "Clarify" software developed by King, Tomz, and Wittenberg 2001. For each such calculation, all other covariates are set at the sample mean values