Medicaid Utilization:
A Comparison between Pre- and Post-Employment

Introduction

With the onset of Welfare Reform, numerous research questions have arisen. One overall prevailing question that continues to haunt Welfare Reform is whether members of a household (particularly children) have been hurt in the implementation of Welfare Reform and the onset of employment. There are indeed many questions on what happens inside the household now that a parent or head of the household must work. While this paper does not offer conclusive answers to these and other questions, it does however offer a brief examination into a household's utilization of Medicaid medical services before and after employment. While utilization of Medicaid medical services does not ensure good health and the lack of utilization does not necessarily indicate "poor" health, changes in the utilization patterns could serve as a warning signal to policy makers - giving administrators the opportunity to investigate. This paper examines the research question "Has Medicaid utilization changed for any member of a TANF household where at least 1 adult member has become employed?" While the findings of this paper do not answer that question conclusively, it does provide some preliminary information using several key administrative files.

Background

The SC Department of Social Services (SCDSS), with the assistance of the Budget and Control Board's Office of Research (ORS), has developed a statistical data warehouse. In this statistical data warehouse, information from several key administrative files are stripped off (either on a monthly or a quarterly basis) and are linked together to answer key research questions. This statistical data warehouse has in part become available through the result of the CHILD LINK federal grant that supported the development of Child Well-Being indicators. Key administrative files have included information from SCDSS's automated CHIPS systems (TANF and Food Stamps only populations); Child Welfare systems like Title XX, Foster Care Tracking, and Child Protective Services; Employment Security System's (ESC) WAGE match which provides quarterly wages on employed SCDSS clients; and the Department of Health and Human Services Medicaid Payment (and Eligibility) systems. For this study, information on the TANF population off of the SCDSS CHIPS system, from ESC's Wage Match, and from the Medicaid Eligibility and Payment systems was utilized.

Study Population

Because the intent of this study was to examine Medicaid Utilization before and after employment, selection of the study population was critical. The initial population of SCDSS AFDC/TANF clients was selected based upon a number of criteria.

1. First, all cases must have been coded as an AFDC/TANF case, though not every client associated with a particular case had to be coded as an AFDC/TANF participant.
2. Second, the case had to be closed for earned income (closure code of 'EI') within a time frame that allowed for adequate tracking using quarterly ESC Wage Match data. Given the availability of ESC data at the time of the study, these cases closed in June 1996, September 1996, or December 1996. No restriction on subsequent returns and closures was imposed.
3. In order to provide adequate time frames to detect differences in utilization and to control for seasonality, the study population required l-year of pre-employment activity (and 1-year of post employment activity - see next step). Therefore third, each of these cases must have been an active AFDC/TANF case for at least one year prior to the closure date.
4. The resulting population from step 3 was then matched to the quarterly ESC Wage Match information. In order to be retained, the case must have had at lease one adult earning wages in the first and last quarters of the post closure year. Continuous employment however was not a requirement.
5. After controlling for one-year pre-employment and one-year post employment and the other above criteria, the resulting population was 1,635 clients. These clients then underwent cleaning to insure that any possible changes in name, etc. were captured for linking purposes. Date constants were also added to mark the beginning of the "pre-" period, the closure date, and the ending of the "post-" period.
6. The 1,635 starting population was next linked to the Medicaid Eligibility files. The matching criteria used varying combinations of a number of key identifiers: SSN, full name, and date of birth. Numerous quality control checks were performed to ensure that clients were matched correctly. 1,557 of the 1,635 SCDSS clients identified linked to the Medicaid Eligibility files resulting in a 95.2% match rate.
7. Because varying Medicaid Eligibility coverage can affect results, the Medicaid Eligibility coverage was next examined for any biases. This study used a working definition of Medicaid Eligibility coverage of 700 or more days of Medicaid eligibility (2 years less 1 month). One thousand and thirty one clients (1,031) met that definition.
8. One result of using Medicaid coverage definition of 700+ days is that it eliminated most of the newborns. While there, is a great deal of interest, of course, about newborns, it was felt that within the confines of this study that an adequate comparison could not be done and that it would bias the study. All pregnancy related and post-partum claims were also excluded. Again, the purpose was to eliminate any bias introduced by a roughly nine (9) month gestation period within the context of a two-year study. For example, impact of conception in the post-period would fall outside the study's time frame, while pregnancies from the pre-period might result in deliveries in the post period. Differences in the lengths of individual pregnancies confounded our inability to determine dates of conceptions necessary to account for these claims. Therefore all claims of these types were excluded to eliminate their confounding effects. In addition, mothers who are pregnant and newborns require higher numbers of doctor's visits and other types of claims.. Over a two year time period that could again bias the results by inflating the number of claims. Because of the natural "aging" process and the 2-year time period, pregnancy related and post partum claims may not be "evenly distributed" across the period. For example, mothers who had a baby in the first year (pre-employment) period were less likely to have a baby in the next period (or post-employment) period.
9. Those clients who met the 700 or more day requirement were next linked to the Medicaid Services files. Claims essentially were divided into three types: HIC (physician, clinic, and laboratory claims), Outpatient (which includes both emergency and non-emergency room visits) and Inpatient hospitalization. For HIC claims, only-physician visits were included. Outpatient claims were further sub-categorized as emergency and non-emergency visits. In addition, outpatient and HIC claims were summed thereby providing an index for ambulatory care utilization.
10. For each claim type and combination of claims, the number of paid claims for the pre-employment and post-employment periods were summed.

Design

1. One way Analysis of Variance (ANOVA) with repeated-measures factor was the tool elected to use for the analysis. This tool allowed for the analysis of the number of Medicaid claims per individual filed one year prior to employment versus the number of claims filed one year after employment.
2. In an attempt to control for natural changes in medical care utilization as individuals age, the analysis was stratified by age. The design is affected by natural changes in people that may occur over time. Because the same population is included in both periods of time - the population quite literally is aging. This aging could bias the results. For example, a child from two years old to three years old may go to the doctor more often for illnesses than that same child who ages from three years old to four years old. Elderly adults also have much higher rates of utilization. The data allowed for three general age classifications to maintain sufficient sample sizes to detect meaningful differences in Medicaid utilization. Each age group had approximately 30% of the population. These age groups were defined as 0-5 years, 6-17 years, and 18 years and over. The pre-employment. and post-employment time frames covered the same months for any given individual thereby controlling for seasonal variations in medical care utilization.
3. In order to partially test for this aging bias, a similar analysis was repeated on the entire 0-5 year old Medicaid population.

Limitations

1. Analysis of variance with repeated measures is subject to some statistical assumptions involving the sample distributions. Repeated measures ANOVA assumes normal distributions with homogeneous variance in the pre- and post- groups. Unlike two-way and r-way ANOVA and least squares regression, repeated measures ANOVA is not particularly robust when these assumptions are violated. Violations tend to result in inflated F-ratios and greater likelihood of falsely rejecting the null hypotheses. In the case of this study, plots did not uncover dramatic departures from normality and the distributions will converge towards normality as sample size increases (i.e. they are asymmetrically normal). Plots also indicate rough homogeneity of variance though formal tests were not performed given the weakness of the substantive findings.
2. The primary limitation of this design is the lack of a control group. Since all subjects in the study had at least one person in the household who became employed, no comparison can be made that can evaluate whether any potential differences in Medicaid utilization are due to changes in employment of persons in the household. In addition, finally, there may be factors other than employment, which explain any potential differences observed in Medicaid utilization. Examples of potential confounding factors in this study could be age or obtaining private medical insurance with employment.

Conclusions

Our analysis finds virtually no support for employment effects on Medicaid utilization by former TANF recipients. The evidence is weak and inconsistent. The substantive differences among pre- and post - employment means is typically .2 claims per person and never exceeds .45 claims when all claim categories are combined. Not surprisedly, few of these differences are statistically significant at the .05 level. Furthermore these insignificant results occur using a technique where undiagnosed violations of underlying statistical assumptions make finding a significant relationship more, rather than less, likely.

More importantly only one specific finding, a decrease in outpatient claims among 18+ year olds is supportive of the employment hypothesis. The magnitude of this difference is not large enough to effect overall claims. Total claims, as well as HIC claims, are instead attributable to aging and differences in the health care needs of 0-5 year olds. It is clear from our analysis of the Medicaid population as a whole, that Medicaid utilization declines as children move from infancy into childhood. This pattern is also found in the SCDSS linked population. In this case, over-all differences in the number of claims are attributable to decrease among 0-5 year olds, particularly through reductions in the number of HIC claims by this group. Thus overall differences in pre- and post- employment Medicaid utilization are an artifact of aging by 0-5 year olds during the time frame of this study.

While the findings for employment effects on Medicaid Utilization are, at best, very weak, it should not be inferred that these relationships do not merit further investigation. This initial study should be supplanted and refined with additional data, and by alternative research designs. A longer time frame might reveal relationships missed by this study. Continued expansion of the Employment Security Commission data set should allow for 18 month pre- and post- periods by early 1999. The results presented here also suggest that a more detailed comparison of the pre- and post- employment SCDSS populations to the general Medicaid populations be in order if only to determine how the SCDSS population differs.

Likewise other analytic strategies should prove useful. For example, the current study focused on controlling for the individual's state of health pre- and post- by examining the same person at each point in time (i.e. matched pairs). This strength brought certain trade-offs such as the inability to control for continuous variables effectively, the inability to use multiple controls, and a lost of statistical robustness. Future studies might use more powerful multivariate techniques coupled with a proxy to control for the health status of the individual.