The firm-level data Directory UMM :Journals:Journal of Health Economics:Vol19.Issue1.Jan2000:

A final econometric issue with our two-sample technique pertains to the standard errors. In estimating the standard errors we treat the unobservability of M U as a missing data problem, and use a multiple imputations approach. A single f ˆ X X imputation is done by replacing M with M where M is a draw from the normal ˆ Ž . distribution N M, s . A consistent estimate of the covariance matrix is T ˆ M J where ˜ T s V q B . 3 Ž . J J ˜ Ž . In this equation, V is the original uncorrected estimate of the covariance matrix and B is J J 1 X B s b y b b y b 4 Ž . Ý ž ž J j j J y 1 js1 Ž . where b is the coefficient vector from one of our J J s 100 regressions using j X y1 11 M , a simulated value from the above distribution, and b s J Ý b . j j j

4. The firm-level data

The firm-level data we use in our analysis were collected in five employer surveys conducted in 1989, 1990, 1991, 1993, and 1995. These surveys have been used individually and in various combinations in a number of published studies, and are perhaps the most widely cited source of information on employer health care costs and benefits. While the pooled cross-section data set we construct with Ž . these surveys is not without its weaknesses which we will discuss we are aware of no other available employer data set suitable for examining how employers may have responded to the Medicaid expansions. 12 All of the surveys were administered by telephone to a sample of U.S. employers drawn from Dun and Bradstreet’s nationwide list of firms. The surveys 11 Ž . Schenker and Welsh 1988 show the consistency of the multiple imputations estimator. For an Ž . application in a different context, see Brownstone and Valletta 1996 . 12 Several surveys conducted recently also provide information on firms’ insurance decisions and limited information on employee demographics. One is the Robert Wood Johnson Foundation Employer Health Insurance Survey, which was conducted in 1993 for firms in 10 states and again in 1996 at the national level. The National Center for Health Statistics also conducted an employer survey in 1992. While these surveys have some advantages relative to the data we use — most notably larger samples — for reasons of timing, they are not suitable for analyzing the Medicaid expansions. While the 1987 NMES and the 1996 MEPS are household surveys, both gather fairly detailed information on health insurance from respondents’ employers. The employer portion of the latter survey is not yet available to researchers, though when it is, it may be possible to examine the effect of public insurance programs on employer behavior using these data. consist of two groups: the 1989–1991 surveys, sponsored by the Health Insurance Ž . Association of America HIAA , and the 1993 and 1995 surveys, sponsored by KPMGrPeat Marwick, the Robert Wood Johnson Foundation, and the Kaiser Family Foundation. 13 Each of the HIAA surveys, which were intended to be nationally representative, contains between 2000 and 3300 observations on firms of all sizes. The 1993 and 1995 surveys were divided by firm size, with the large and small firm portions sponsored and conducted separately. Both the large and small firm samples consist of approximately 1000 firms in each year. Unfortu- nately, the large firm surveys from the later years contain data only for firms which offer insurance and do not include important questions pertaining to employee characteristics. In addition, the 1993 and 1995 samples include very few firms with one or two employees. In light of these problems, we restrict our analysis to firms with between 5 and 100 employees. While it would be ideal to have data on firms of all sizes, smaller firms are of particular interest, for three reasons. First, smaller firms are more likely to be making marginal decisions about whether or not to offer health insurance. 14 Second, smaller firms are more likely to employ a homogenous workforce, and hence may be more likely to respond to the Medicaid expansions. Third, larger firms are more likely to operate in more than one state, making it difficult to relate Medicaid eligibility rules to firm behavior. To ensure that our sample is nationally representative of firms in this size range, we obtained a measure of the distribution of firms by size and region from the Census Bureau’s County Business Patterns. To calculate firm-level weights, we divide the observations into categories based on year, Census region and five firm-size categories: 5–9 employees, 10–19, 20–49, and 50–99. The weight assigned to firms in each cell equals P rp , where P is the proportion of the c c c population represented by cell c, and p is the corresponding sample proportion. c We also calculated a second set of weights to make the data nationally representa- tive with respect to employees in small firms. The employee-level weight assigned to an observation is the observation’s firm-level weight multiplied by the number of employees in the firm. Table 2 presents firm-weighted and employee-weighted summary statistics from our firm level data. The top panel lists our dependent variables. The first is an indicator variable for whether or not the firm offers health insurance to its workers. As shown in Appendix A, the question on which this variable is based was essentially the same in each year. In the pooled sample, the firm-weighted 13 Ž . The response rates were higher in the surveys conducted by HIAA between 66 and 70 than in Ž . the surveys of the two later years 44 and 58 . 14 According to our data as well as other sources, such as the Bureau of Labor Statistics’ Employee Benefits Survey, insurance provision is virtually universal among firms with more than 100 employees. mean is 0.65 and the employee-weighted mean is 0.77, the difference between the two reflecting the positive relationship between firm size and insurance provision. For firms that offer insurance we also examine various measures of plan generos- ity: whether or not the firm offers dependent coverage and the employee contribu- tions for single and family coverage, expressed as a percentage of the respective Table 2 Summary statistics for employer survey data Ž . Mean standard deviation Minimum Maximum Firm- Employee- weighted weighted Dependent Õariables Firm offers health insurance 0.65 0.77 1 Firm offers dependent coverage 0.95 0.97 1 Ž . Ž . Employees’ share of single premium 15.63 26.53 15.94 25.12 100 Ž . Ž . Employees’ share of family premium 29.83 31.03 34.23 30.76 100 Ž . Ž . Percentage of take-up among FT workers 78.92 26.09 79.26 25.91 1 100 Medicaid eligibility Õariables Ž . Ž . Percentage of firm’s employees eligible 7.63 6.28 7.36 6.23 y1.69 42.10 Ž . Ž . Percentage with eligible family members 13.37 8.24 13.56 8.09 y1.75 50.77 Firm and worker characteristics Ž . Ž . Percentage of workers earning 19.42 28.17 18.13 26.59 100 -US10,000ryear Ž . Ž . Number of employees 17.63 18.43 36.89 28.71 5 100 Ž . Ž . Ž . Firm age in years 18.21 14.46 19.80 16.96 215 Agriculture, mining 0.02 0.02 1 Construction 0.11 0.09 1 Manufacturing 0.14 0.18 1 Transportation, utilities 0.05 0.05 1 and communication Wholesale trade 0.08 0.09 1 Retail trade 0.22 0.19 1 Finance, insurance, real estate 0.06 0.06 1 Services 0.31 0.30 1 County-leÕel Õariables Ž . Ž . Ž . Per capita income US1000 20.67 0.61 20.67 0.58 8.12 58.10 Ž . Ž . Unemployment rate 5.83 2.22 5.86 2.21 1.6 28.8 Ž . Ž . Percentage of residents in poverty 12.62 5.81 12.70 5.76 2.6 47.5 Ž . Ž . Percentage of residents in urban area 74.51 26.52 75.71 25.69 100 Ž . Ž . Percentage of workers in manufacturing 17.36 7.51 17.50 7.55 1.1 51.7 Entries are summary statistics from 1989–1991 HIAA surveys and 1993 and 1995 KPMGrPeat MarwickrWayne State surveys. Summary statistics for explanatory variables pertain to the full sample Ž . N s 3082 . Sample sizes for analyses limited to firms that offer insurance are smaller, and are reported in the appropriate tables. premiums. As with insurance offers, the questions pertaining to these outcomes are nearly identical across the different surveys. For firms that offer insurance we also examine the effect of Medicaid eligibility on employee take-up, though here there are some minor data problems. The HIAA surveys provide sufficient information to calculate the take-up rate among full-time employees who are offered benefits. While ideally we would like to know the take-up rate among all workers offered coverage, since part-time workers are seldom offered benefits, the distinction between all eligible workers and eligible full-time workers is not great. 15 A second problem is that identical questions pertaining to eligibility and take-up were not asked in the latter two surveys. It is possible, however, to use information on the total number of workers covered and yesrno questions on whether any part-time workers are offered insurance to Ž construct a close proxy for the take-up rate among full-time employees see . Appendix A for more details . Because the definition of this variable is not identical across all years, we also estimate a set of take-up regressions on the Ž 1989–1991 sample only. Table 2 reports summary statistics on the take-up rate . for the entire 5 years. A final limitation of our data with respect to take-up is that since the establishment surveys provide no information on the number of workers who are married or have children, we cannot calculate a meaningful take-up rate for dependent coverage. The next panel of Table 2 contains the Medicaid eligibility variables con- structed by combining coefficients estimated in the CPS with data from the employer surveys. The figures indicate that in the late 1980s to mid 1990s, most firms employed very few workers who either qualified for Medicaid or had family members who did. The firm-weighted sample means for the two variables are 7.63 Ž . and 13.37, and the 75th percentile values not reported are 9.92 and 17.83, respectively. 16 The firm characteristics common to all five employer surveys are firm size, firm age, the percent of workers earning less than US10,000 per year, and industry. Since the data file for each year’s survey identifies the zip codes of responding firms we were able to determine the county in which each firm is located, and then merge in several county level-variables from the Area Resource File to control for local labor market conditions. Two of these variables — the county’s unemployment rate and per capita income — vary across counties and years, while the other three — the percentage of county residents in urban areas, the percent in poverty, and the percent employed in manufacturing — are based on the 1990 census, and hence vary only across county. 15 Ž . According to the US Bureau of Labor Statistics 1994 , only 5 of part-time workers in firms with 100 or fewer employees received health benefits in 1992, compared to 71 of full-time workers. 16 As we used a linear probability model in the first stage the predicted probabilities were not bounded by 0 and 1. Use of a logit model in the first stage made no difference in the results, however.

5. Regression results