Regression results Directory UMM :Journals:Journal of Health Economics:Vol19.Issue1.Jan2000:

5. Regression results

5.1. The firm’s decision to offer insurance Table 3 presents results from several linear probability models of the firm’s decision to offer health insurance benefits. 17 We report models using both Medicaid eligibility variables and using firm- and employee-level weights. In a fully structural model, a firm’s decision to offer insurance will depend on the premiums it faces in the market. The estimation of such a model is compli- cated by the fact that premiums are observed only for those firms offering Ž . coverage see Feldman et al., 1997 . Since our interest is not in the effect of premiums per se, we estimate the model in the reduced form. As a result, the coefficients estimated for some variables combine an indirect effect working through their effect on premiums, as well as more direct effects. For example, our finding of a strong positive effect of firm size is at least partly explained by the fact that, ceteris paribus, smaller firms face higher premiums due to higher per-employee administrative costs. Our results indicate that the probability of offering insurance increases by one percentage point for every 3 to 5 years that a firm has been in business. As with the firm size effect, this finding is consistent with results from other studies using Ž . establishment data Levy, 1997; Feldman et al., 1997 . The estimated coefficients on the county-level variables also conform with our expectations. Insurance provision is more common in counties with higher average income and less common in counties with higher poverty rates, though the latter effect is less precisely estimated. We find that health insurance provision is more common in Ž urban areas, a result that has been noted previously Markowitz et al., 1991; . Frenzen, 1993; Feldman et al., 1997; Coburn et al., 1998 . These significant county effects increase our confidence that our measures of Medicaid availability are not picking up unobserved factors related to local economic conditions. Our results imply that firms that make greater use of low wage workers are significantly less likely to offer insurance. The firm-weighted results imply that a 10 percentage point increase in the percentage of workers earning less than US10,000 per year lowers the offer probability by roughly 5 percentage points. Evaluated at the sample means, this corresponds to an elasticity of y0.15. The employee-weighted results are quite similar, translating to an elasticity of y0.12. Because Medicaid is a means-tested program, this implies that firms with more Medicaid-eligible workers are less likely to offer insurance. However, this rela- tionship, in and of itself, does not bear directly on the question of crowd-out due 17 We report linear probability models both to ease interpretation and because the multiple imputa- tions method we use to estimate the standard errors is complicated by the use of a nonlinear model. However, probit models yield the same qualitative results aside from the standard errors. L. Shore-Sheppard et al. r Journal of Health Economics 19 2000 61 – 91 77 Table 3 Determinants of the firm’s decision to offer health insurance Firm-weighted Employee-weighted Ž . Ž . Ž . Ž . 1 2 3 4 Ž . Ž . Percentage of workers eligible for Medicaid y0.00006 0.00573 – y0.00080 0.00468 – Ž . Ž . Percentage in families with eligible members – 0.00301 0.00431 – 0.00194 0.00349 UU UU UU UU Ž . Ž . Ž . Ž . Percentage of workers earning -US10,000ryear y0.005 0.001 y0.006 0.001 y0.005 0.001 y0.006 0.001 UU UU UU UU Ž . Ž . Ž . Ž . Number of employees 0.014 0.001 0.014 0.001 0.010 0.001 0.010 0.001 UU UU UU UU Ž . Ž . Ž . Ž . Number of employees squared y0.0001 0.0000 y0.0001 0.0000 y0.0001 0.0000 y0.0001 0.0000 UU UU UU UU Ž . Ž . Ž . Ž . Ž . Firm age in years 0.003 0.001 0.003 0.001 0.002 0.0004 0.002 0.0004 Ž . Ž . Ž . Ž . Agriculture, mining 0.008 0.104 0.004 0.105 0.013 0.072 0.000 0.076 UU UU Ž . Ž . Ž . Ž . Construction y0.068 0.049 y0.072 0.050 y0.087 0.034 y0.088 0.034 U Ž . Ž . Ž . Ž . Transportation, utilities, communication 0.088 0.054 0.093 0.055 0.036 0.036 0.042 0.038 UU Ž . UU Ž . UU Ž . UU Ž . Wholesale trade 0.116 0.044 0.127 0.045 0.067 0.028 0.077 0.029 Ž . Ž . Ž . Ž . Retail trade y0.031 0.038 y0.031 0.038 y0.039 0.029 y0.036 0.029 Ž . Ž . Ž . Ž . Finance, insurance, real estate y0.047 0.054 y0.036 0.054 y0.018 0.036 y0.007 0.036 Ž . Ž . Ž . Ž . Services 0.053 0.035 0.059 0.037 0.017 0.022 0.022 0.023 UU UU Ž . Ž . Ž . Ž . Ž . Per capita income US1000 0.006 0.002 0.006 0.002 0.003 0.002 0.003 0.002 Ž . Ž . Ž . Ž . Unemployment rate 0.006 0.006 0.006 0.006 0.006 0.004 0.006 0.004 Ž . Ž . Ž . Ž . Percentage of residents in poverty y0.003 0.003 y0.003 0.003 y0.002 0.002 y0.002 0.002 UU UU UU UU Ž . Ž . Ž . Ž . Percentage of residents in urban area 0.001 0.0005 0.001 0.0005 0.001 0.0004 0.001 0.0004 Ž . Ž . Ž . Ž . Percentage of workers in manufacturing y0.001 0.002 y0.001 0.002 y0.001 0.001 y0.001 0.001 2 R 0.269 0.269 0.273 0.273 Sample size 3062 3062 3062 3062 Ž . Standard errors, corrected for two-step estimation method see text are in parentheses. All models include year and state dummy variables. UU Statistically significant at the 0.05 level. U Statistically significant at the 0.10 level. to the expansions. Rather, it is the coefficients on our fitted Medicaid eligibility variables that provide tests of the hypothesis that the Medicaid expansions caused a reduction in insurance offers. 18 In each of the four models, the estimated coefficient on the Medicaid variable is Ž . small and not significantly different from zero. In column 1 , the point estimate of y0.00006 implies that a 10 percentage point increase in the percentage of a firm’s workers eligible for Medicaid will reduce the firm’s probability of offering insurance by less than one-tenth of one percentage point. At the employer-weighted sample means, this represents an elasticity of y0.0007. However, because the confidence interval around this estimate is fairly wide, we cannot rule out substantially larger effects. With a bootstrapped standard error of 0.0057, the Ž . lower bound of our 95 confidence interval i.e., the most negative effect implies that a 10 percentage point increase in Medicaid eligibility would lead to an 11 Ž percentage point decline in the offer rate and the upper bound is a positive effect . of roughly the same magnitude . When we use the family-based measure of Ž Ž .. Medicaid eligibility and employer weights column 2 , the point estimate is positive, though the 95 confidence interval includes effects as negative as y0.0059. The employee-weighted results are slightly more precise, though still statistically insignificant. The finding that the expansions did not appear to affect the provision of insurance by small firms, although imprecisely estimated, is robust to changes in the way the Medicaid variables enter the regression. For instance, to examine whether there were effects of eligibility levels which were limited to firms with high levels of eligibility, we estimated models in which the percent eligible for Medicaid entered in categorical form. 19 We tried a variety of categorizations; none indicated a significant relationship between the availability of Medicaid for a firm’s employees and the decision to offer insurance. Models including the percent eligible squared also implied that any crowding-out that occurred as a result of the expansions was not caused by employers dropping coverage. This finding is also robust to the inclusion or exclusion of key control variables. One possible criticism of the specifications reported in Table 3 is that by including a full set of state dummies in addition to the county variables, we may be ‘‘over-controlling’’ for local economic conditions and leaving too little residual variation to be explained by our Medicaid variable. To examine whether this was the case, we tried replacing the state dummies with dummies for the census region or division, which would increase the cross-sectional variation for identifying the crowd-out effect. Again, the results from such models were not qualitatively different from those reported in Table 3. 18 This point is illustrated by the fact that when we drop the percent under US10,000 variable from the regression we obtain significant negative Medicaid effects. 19 The control variables are the same as in Table 3. Full results are available from the authors. Finally, an argument can be made that any negative effect of Medicaid eligibility on firm offer decisions may be limited to very small firms. To test this, we divided the sample into four equal-sized groups by firm size: firms with 10 or fewer employees, 11 to 25 employees, 26 to 50 employees, and more than 50 employees. We then re-estimated the models separately for each of the four groups, and found similar small, insignificant effects in all of them. 5.2. Measures of health plan generosity Although we found no evidence that firms reduced their likelihood of offering coverage, some small insurance-providing firms may have responded to Medicaid Ž expansions by encouraging workers to enroll their children and perhaps them- . selves in the public program. The most direct way to do this is to drop family coverage altogether. A more subtle approach is to increase employee premium Ž contributions to encourage workers with Medicaid-eligible dependents and those . who are eligible themselves to voluntarily drop coverage. Table 4 reports the key results from regressions that attempt to investigate these potential effects. The column layout is similar to Table 3, though for the sake of brevity we do not report any control variable coefficients. 20 The first row of Table 4 pertains to models in which the dependent variable equals one if the firm offers dependent coverage, and zero if insurance is offered to workers only. It is here that we find our strongest evidence of crowd-out. All four specifications reported yield a negative relationship between the fraction of a Ž firm’s employees estimated to be eligible for Medicaid or in families with eligible . members and the decision to offer family coverage. The effects are slightly larger and more precisely estimated when we weigh by the number of employees: the coefficients of y0.006 and y0.004 in the third and fourth columns have absolute t-statistics of 2.34 and 2.43, respectively. These effects are substantially larger than those from the offer equations of Table 3. The point estimate from column 3 Ž . column 4 implies that a 10 percentage point increase in the Medicaid eligibility Ž . variable decreases the probability of offering family coverage by 6 4 percentage Ž . points. At the lower most negative bound of our 95 confidence interval, a 10 point increase in Medicaid eligibility would correspond to a 12 percentage point decline in family offers. The coefficients on most of the other variables in the row 1 model are statistically insignificant. In particular, the percent of low-wage workers in the firm does not affect the likelihood of a firm offering family coverage, conditional on offering coverage at all. The exceptions are that larger firms and ones that have been in business longer are more likely, and firms in urban areas are less likely, to 20 Sample sizes differ across the outcomes because of missing data on premiums and employee contributions for some firms. Table 4 The effect of employee Medicaid eligibility on the generosity of employer health benefit programs Firm-weighted Employee-weighted Percent of workers . . . Percent of workers . . . Eligible for Ineligible Eligible for Ineligible Medicaid families Medicaid families Dependent Õariable UU UU Ž . Ž . Ž . Ž . Ž . 1 Firm offers y0.003 0.004 y0.002 0.003 y0.006 0.003 y0.004 0.002 family coverage Ž . N s 2290 2 R 0.106 0.106 0.065 0.065 Ž . Ž . Ž . Ž . Ž . 2 Employees’ 0.425 0.498 0.188 0.312 y0.003 0.402 0.075 0.268 share of premium, single coverage Ž . N s 2017 2 R 0.104 0.105 0.104 0.105 Ž . Ž . Ž . Ž . Ž . 3 Employees’ 0.030 0.618 y0.108 0.424 y0.456 0.518 y0.279 0.354 share of premium, family coverage Ž . N s1988 2 R 0.183 0.183 0.177 0.177 UU Statistically significant at the 0.05 level. Ž . Standard errors, corrected for two-step estimation method see text are in parentheses. Table entries are estimated coefficients of imputed Medicaid eligibility levels in regressions on the dependent variables noted above. All models include the same set of controls as in Table 3. offer dependent coverage. Firms in mining and agriculture and finance, insurance, and real estate are somewhat more likely than manufacturing firms to offer family coverage, while firms in transportation and utilities are somewhat less likely. Differences across states are small and generally not statistically significant. In rows 2 and 3, the dependent variables are the employee contributions required for single and family coverage, both expressed as a percentage of the relevant plan premium. The hypothesis that firms altered their premium contribu- tion policies to encourage workers to substitute Medicaid for employer-sponsored coverage predicts positive coefficients on our Medicaid variables. For single coverage, the firm-weighted coefficients are positive and fairly large. The point estimate in the first column implies that a 10 percentage point increase in Medicaid eligibility increases the percentage of single premiums paid directly by employees by over 4 percentage points. The estimated coefficient in column 2 implies that a similar increase in the percentage of workers with eligible family members increases the employee share by just under 2 percentage points. How- ever, because of large standard errors, both of these estimated effects are statisti- cally insignificant, with t-statistics less than one. 21 When we use employee weights, the point estimates are much smaller, while the standard errors decline only slightly. To the extent that employers did increase contributions in response to the availability of Medicaid, we would expect the effect to be greatest for family Ž . coverage. However, when we use percent family coverage contributions as our dependent variable, the Medicaid coefficients are substantially smaller than in the corresponding single coverage regressions. In fact, the employee-weighted coeffi- cient estimates are quite large and negative, though, again, insignificant due to large standard errors. Thus, while the imprecision of our results makes any definitive conclusion impossible, it is hard to rationalize the pattern of the point estimates with the hypothesis that employers’ premium contribution policies were influenced by the availability of Medicaid. 22 In addition to these models, we also estimated regressions in which the dependent variable is the single or family contribution measured in dollars or as an indicator variable for whether any contribution is required at all. These additional regressions also offer no support for the hypothesis of a positive effect of Medicaid eligibility on the existence or level of employee premium contributions. However, as with the results reported in the table, the standard errors are quite large, which means we cannot rule out the possibility of fairly substantial effects. 5.3. Employee take-up Regardless of how employers responded to the Medicaid expansions, private insurance may have fallen as a result of newly eligible workers declining coverage Ž . offered to them. As noted, the analysis by Cutler and Gruber 1996 using the CPS Employee Benefit Supplement suggests that this was the primary way that the expansions reduced private coverage. To address this possible effect, the final outcome we examine is the take-up rate among full-time employees who were offered insurance by their employers. Because the take-up rate reflects decisions made by individual employees rather than employers, we estimate employment-weighted regressions only. The first two columns of Table 5 contain results for all firms in our sample that offer insurance. The next two columns contain results for a subsample of 966 firms that have 25 or more employees. There are two arguments for restricting the sample in this way. First, the fact that the take-up rate is observed only for firms that offer insurance 21 It is important to note that the dependent variable in row 1 is an indicator variable, whereas those Ž . in rows 2 and 3 range theoretically from 0 to 100. As a result, the coefficients are not directly comparable. 22 Since Medicaid eligibility increased most rapidly in the early part of our sample, we estimated the Ž . regressions in Tables 3 and 4 using only the HIAA 1989–1991 data. These results were qualitatively the same, except the standard errors were even larger when the whole sample is used. L. Shore-Sheppard et al. r Journal of Health Economics 19 2000 61 – 91 82 Table 5 Take-up of private insurance by full-time employees, all years All firms Firms with 25 employees Employee contribution required Ž . Ž . Ž . Ž . Ž . Ž . 1 2 3 4 5 6 Ž . Ž . Ž . Percentage of workers y0.099 0.459 – y0.784 0.651 – y0.582 0.562 – eligible for Medicaid Ž . Ž . Ž . Percentage in eligible – y0.218 0.286 – y0.555 0.380 – y0.512 0.339 families UU UU UU UU UU UU Ž . Ž . Ž . Ž . Ž . Ž . Employee premium y0.149 0.021 y0.149 0.021 y0.187 0.037 y0.187 0.037 y0.145 0.024 y0.145 0.024 Ž . contribution single coverage UU UU UU U UU Ž . Ž . Ž . Ž . Ž . Ž . Percentage of workers y0.268 0.088 y0.242 0.068 y0.198 0.129 y0.223 0.100 y0.202 0.104 y0.201 0.080 earning - US10,000ryear Ž . Ž . Ž . Ž . Ž . Ž . Number of employees 0.062 0.081 0.063 0.081 0.242 0.237 0.250 0.238 0.064 0.100 0.053 0.099 Ž . Ž . Ž . Ž . Ž . Ž . Number of employees squared y0.0004 0.001 y0.0004 0.001 y0.002 0.002 y0.002 0.002 y0.001 0.001 y0.001 0.001 Ž . UU Ž . UU Ž . Ž . Ž . UU Ž . UU Ž . Firm age in years 0.082 0.032 0.081 0.032 y0.033 0.037 y0.031 0.037 0.104 0.037 0.105 0.037 UU UU Ž . Ž . Ž . Ž . Ž . Ž . Agriculture, mining 2.879 3.727 4.059 3.979 3.700 5.065 5.243 5.436 10.188 4.676 11.991 4.859 UU UU UU UU Ž . Ž . Ž . Ž . Ž . Ž . Construction y5.899 2.657 y5.773 2.673 y8.074 3.462 y7.991 3.458 y4.294 3.177 y4.038 3.181 Ž . Ž . Ž . Ž . Ž . Ž . Transportation, utilities, y1.098 3.121 y1.547 3.120 1.498 4.173 0.952 4.190 0.297 3.826 y0.410 3.852 communication Ž . Ž . Ž . Ž . Ž . Ž . Wholesale trade y0.841 1.988 y1.534 2.048 y0.777 2.562 y1.499 2.712 0.636 2.356 y0.347 2.449 UU UU UU UU Ž . Ž . Ž . Ž . Ž . Ž . Retail trade y5.308 2.755 y5.447 2.748 y7.706 5.339 y7.006 4.988 y7.594 3.514 y7.480 3.487 Ž . Ž . Ž . Ž . Ž . Ž . Finance, insurance, 1.040 2.457 0.414 2.393 y2.220 3.623 y2.044 3.224 y1.114 3.033 y1.607 3.002 real estate UU Ž . UU Ž . Ž . Ž . UU Ž . UU Ž . Services y3.391 1.553 y3.862 1.623 y2.319 2.374 y3.190 2.540 y4.083 2.031 y5.032 2.158 Ž . Ž . Ž . Ž . Ž . Ž . Per capita income 0.280 0.179 0.285 0.178 0.226 0.264 0.232 0.259 0.443 0.295 0.456 0.292 Ž . US1000 U U Ž . Ž . Ž . Ž . Ž . Ž . Unemployment rate 0.177 0.387 0.194 0.383 0.474 0.616 0.468 0.603 0.750 0.448 0.756 0.443 U U Ž . Ž . Ž . Ž . Ž . Ž . Percentage of y0.100 0.190 y0.104 0.190 y0.406 0.244 y0.404 0.243 y0.192 0.239 y0.186 0.237 residents in poverty Ž . Ž . Ž . Ž . Ž . Ž . Percentage of y0.002 0.029 y0.003 0.029 y0.008 0.039 y0.007 0.039 0.015 0.040 0.015 0.040 residents in urban area Ž . Ž . Ž . Ž . Ž . Ž . Percentage of 0.026 0.095 0.024 0.095 y0.200 0.141 y0.200 0.141 0.017 0.119 0.016 0.120 workers in manufacturing 2 R 0.200 0.200 0.279 0.280 0.232 0.233 Sample size 2050 2050 966 966 1421 1421 Ž . Standard errors, corrected for two-step estimation method see text are in parentheses. All models include year and state dummy variables and are run using employee weights, as noted in the text. UU Statistically significant at the 0.05 level. U Statistically significant at the 0.10 level. raises the issue of selectivity bias. Since as firm size increases it becomes less and less common for firms not to offer insurance, selectivity bias should be less of a problem for larger firms. 23 The second argument is that the percent take-up rate will vary little, and is likely to be measured with substantial error, in very small firms. 24 Since few workers will decline coverage when no premium contribution Ž . is required even if alternative coverage is available , we also estimate the take-up equation on a sample of 1421 firms that require a monthly premium contribution Ž . for either single or family coverage columns 5 and 6 . The explanatory variables include those used in the other regressions plus the employee premium contribution required for the single coverage plan. While this variable is arguably endogenous, the results from Table 4 suggest that employer contribution policies are orthogonal to the availability of Medicaid. Because of this and the fact that it is an important determinant of employee take-up, we include Ž the employee contribution on the right hand side though dropping it has no . material effect on any of the other coefficients . As expected, the estimated effect of the premium contribution is negative and significant. Since the dependent variable is scaled from 0 to 100, the column 1 results imply that a US10 increase in the monthly premium contribution reduces the employee take-up rate by 1.5 percentage points. Evaluated at the sample means for our data, the premium contribution coefficient from the full sample regression corresponds to an elastic- ity of y0.045, which is quite comparable to the take-up elasticity of y0.036 Ž . calculated by Chernew et al. 1997 using a combined employer–employee data set. Ž Consistent with other studies using different types of data Chernew et al., . 1997; Cooper and Schone, 1997 the results in Table 5 indicate that the take-up rate declines with the percentage of low wage workers in the firm. Again, while workers earning less than US10,000 are substantially more likely to qualify for Medicaid, this is not necessarily evidence of a negative impact of Medicaid, since it is difficult to disentangle the effects of Medicaid availability on demand for employer-provided insurance from other factors affecting insurance coverage for low-wage workers. It is, however, relevant to more general policy questions concerning the insurance coverage of low wage workers. For example, this result indicates that requiring firms to offer coverage to low-wage workers may not 23 An alternative approach to the issue of selection is to combine offering and non-offering firms to estimate models of ‘‘percent covered,’’ coding all non-offering firms as zeros. We estimated such models and they generated insignificant Medicaid effects. This result is not surprising given our results for the offer equations. 24 In the small group market, insurance carriers typically require a minimum employee participation rate before they sell insurance to a small firm. Consider a firm of five employees facing a minimum participation rate of 75. If this firm offers insurance, the take-up rate can only have two values: 100 or 80. L. Shore-Sheppard et al. r Journal of Health Economics 19 2000 61 – 91 84 Table 6 Ž . Employee take-up results for HIAA data 1989 – 1991 All firms Firms with 25 employees Employee contribution required Ž . Ž . Ž . Ž . Ž . Ž . 1 2 3 4 5 6 Ž . Ž . Ž . Percentage of workers y0.511 0.692 – y1.636 1.346 – y1.217 0.845 – eligible for Medicaid U Ž . Ž . Ž . Percentage in – y0.565 0.488 – y0.851 0.786 – y1.030 0.556 eligible families UU UU UU UU UU UU Ž . Ž . Ž . Ž . Ž . Ž . Employee premium y0.223 0.046 y0.223 0.046 y0.248 0.079 y0.247 0.080 y0.205 0.054 y0.205 0.054 Ž . contribution single coverage Ž . Ž . Ž . Ž . Ž . Ž . Percentage of workers y0.172 0.119 y0.147 0.101 y0.062 0.232 y0.178 0.155 y0.066 0.148 y0.071 0.115 earning - US10,000ryear Ž . Ž . Ž . Ž . Ž . Ž . Number of employees 0.065 0.101 0.056 0.101 0.227 0.297 0.239 0.299 0.071 0.148 0.046 0.149 Ž . Ž . Ž . Ž . Ž . Ž . Number of employees squared y0.0004 0.001 y0.0004 0.001 y0.002 0.003 y0.002 0.003 y0.001 0.001 y0.001 0.001 U U UU UU Ž . Ž . Ž . Ž . Ž . Ž . Ž . Firm age in years 0.074 0.043 0.073 0.043 y0.051 0.053 y0.050 0.053 0.125 0.053 0.125 0.052 Ž . Ž . Ž . Ž . U Ž . UU Ž . Agriculture, mining 3.875 4.475 6.089 5.245 4.971 8.128 5.848 9.126 10.409 5.589 13.848 6.211 UU UU Ž . Ž . Ž . Ž . Ž . Ž . Construction y3.464 3.136 y3.466 3.128 y6.670 3.193 y7.247 3.171 y3.745 3.887 y3.933 3.949 Ž . Ž . Ž . Ž . Ž . Ž . Transportation, utilities, y1.202 3.710 0.183 3.879 y1.104 5.267 y1.114 5.538 0.589 5.542 y0.899 5.664 communication Ž . Ž . Ž . Ž . Ž . Ž . Wholesale trade y2.186 2.701 y3.347 2.963 y4.328 3.803 y4.003 4.026 y0.462 2.976 y2.028 3.152 U U Ž . Ž . Ž . Ž . Ž . Ž . Retail trade y5.043 3.634 y5.027 3.643 y12.465 7.771 y9.754 6.560 y8.077 4.917 y7.513 4.869 Ž . Ž . Ž . Ž . Ž . Ž . Finance, insurance, y1.859 3.616 y2.537 3.518 y8.884 6.352 y6.540 5.195 y5.128 4.641 y5.434 4.469 real estate UU UU U UU Ž . Ž . Ž . Ž . Ž . Ž . Services y4.727 2.286 y5.578 2.490 y5.454 3.444 y5.909 3.825 y5.392 3.113 y6.809 3.381 Ž . Ž . Ž . Ž . Ž . Ž . Per capita income 0.183 0.368 0.186 0.367 0.341 0.476 0.343 0.478 0.084 0.476 0.098 0.476 Ž . US1000 Ž . Ž . Ž . Ž . Ž . Ž . Unemployment rate 0.313 0.492 0.293 0.493 0.041 0.767 0.043 0.760 0.427 0.744 0.398 0.742 U U Ž . Ž . Ž . Ž . Ž . Ž . Percentage of y0.179 0.215 y0.177 0.216 y0.330 0.237 y0.331 0.240 y0.456 0.257 y0.443 0.256 residents in poverty Ž . Ž . Ž . Ž . Ž . Ž . Percentage of y0.010 0.040 y0.010 0.040 y0.041 0.057 y0.045 0.057 y0.010 0.059 y0.008 0.060 residents in urban area Ž . Ž . Ž . Ž . Ž . Ž . Percentage of 0.057 0.141 0.055 0.141 y0.023 0.244 y0.020 0.223 y0.049 0.187 y0.054 0.187 workers in manufacturing 2 R 0.200 0.246 0.279 0.374 0.232 0.277 Sample size 1313 1313 646 646 874 874 Ž . Standard errors, corrected for two-step estimation method see text are in parentheses. All models include year and state dummy variables and are run using employee weights, as noted in the text. UU Statistically significant at the 0.05 level. U Statistically significant at the 0.10 level. increase coverage as much as desired, since low-wage workers appear less likely to accept coverage that is offered. As with the other outcomes analyzed, our estimates of the effect of Medicaid eligibility on take-up are plagued by relatively large standard errors. However, for all models estimated, our point estimates imply a negative effect of Medicaid eligibility on the percentage of workers accepting offers of insurance, and the pattern of the results across the various subsamples is consistent with expectations. For example, the Medicaid eligibility coefficients are substantially larger in the sample that excludes very small firms than in the full sample. The column 4 results imply that a 10 percentage point increase in the percentage of workers in Medicaid-eligible families causes take-up to fall by 5.5 percentage points. With an absolute t-statistic of 1.47, this effect is significant at the 15 level. When we restrict the sample to firms that require employees to contribute directly for their Ž . coverage column 6 the coefficient on the family-based Medicaid variable is Ž . Ž approximately the same size y0.512 , with a similar significance level p-value . s 0.13 . As noted above, our measure of take-up is clearly defined for the 1989, 1990, and 1991 surveys, and less clearly defined in the two later years. Consequently, we estimated the same models reported in Table 5 using only data from the three Ž . HIAA surveys 1989 to 1991 . The results from these regressions, which are reported in Table 6, also suggest a negative relationship between the percentage of workers in a firm who are eligible for Medicaid and the percentage of workers in the same firm who accept an offer of private coverage. In all cases, the Medicaid coefficients from these restricted samples are larger than their full-sample counter- parts. For example, when we drop the 1993 and 1995 data, the Medicaid coefficient for the column 2 specification goes from y0.217 to y0.565. However, because the sample size falls from 2050 to 1313, the standard error on the Ž . coefficient also increases from 0.261 to 0.488 , yielding an absolute t-statistic of 1.16. Similarly, when we use only the 1989–1991 data and focus on firms that require premium contributions, we obtain a coefficient of y1.030 on the family- Ž . based Medicaid coefficient standard error s 0.556; absolute t-statistic s 1.85 . Ž . Taken together, these results provide some support albeit weak for the hypothesis that the Medicaid expansions crowded out private insurance by inducing some workers to decline offers of employer-sponsored insurance.

6. Conclusions