Racial and Ethnic Differences in Antipsychotic Medication Use Among Children Enrolled in Medicaid

This study drew on data from several sources. The MAX Person Summary (PS) database is a compilation of various state Medicaid Management Information System data files into a uniform data structure that was created to support research and policy analysis. The file contains one record per Medicaid-eligible person per year and reports demographic, eligibility, enrollment, and summary utilization information. The MAX Prescription Drug (RX) file contains one record for each paid drug claim. Drugs in the RX file are identified by a National Drug Code (NDC). In addition to the MAX PS and RX files, we used information from the FirstDataBank database, which includes a crosswalk from NDC to drug type, to identify antipsychotic medications. Finally, we used the Centers for Medicare and Medicaid Services Environmental Scanning and Program Characteristics (ESPC) Database, which houses state-level information on Medicaid and CHIP program characteristics and selected environmental factors. We gathered data from these sources for calendar years 2005 through 2009 to develop a database that would support the analyses.

Our selection of states (Alabama, Colorado, Illinois, Iowa, Louisiana, New Hampshire, North Carolina, and Oklahoma) reflects our goals of representing different census regions and excluding states suspected of incomplete fee-for-service Medicaid claims or high “churning” levels.

We transformed the MAX RX files (one for each state and each year) from a claim-level file to a child-year–level file. This change resulted in a database containing one record per child for each year that the child was eligible for Medicaid. To accomplish the transformation, we created a set of dichotomous variables indicating whether any prescriptions were filled for either a first-generation or second-generation antipsychotic medication. Antipsychotics included aripiprazole, asenapine maleate, clozapine, iloperidone, lurasidone, olanzapine, olanzapine-fluoxetine, paliperidone, quetiapine, risperidone, and ziprasidone. We also retained the first fill date for each. We identified the drugs in each category by using the FirstDataBank database.

We also prepared a child-year–level file from the MAX PS file containing all individual-level information for the selected study states. The file included age group, race-ethnicity, basis of eligibility, and service utilization summaries. The race-ethnicity variable classifies children into six categories: white, African American, American Indian/Alaska Native, Asian, Hispanic (including the designation Hispanic or Latino and the designation Hispanic or Latino and one or more races), and Hawaiian or multiracial (including native Hawaiian/other Pacific Islander and one or more races). The age group variable classifies children into the following age categories: 2–4, 5–11, 12–14, 15–17, and 18–20 years. The eligibility variable was constructed by grouping eligibility categories in the PS file into seven categories: blind-disabled, foster care, poverty, medically needy, Section 1931 (Section 1931 of the Social Security Act requires states to maintain Medicaid coverage for individuals who would have received benefits under the Aid to Families With Dependent Children program), other, and unknown. The prepared PS file also retained several service utilization variables, including claim counts for inpatient hospital, inpatient psychiatric facility, physician, outpatient hospital, targeted case management, and psychiatric services. These services were categorized by using the MAX files’ type-of-service classification system, which is based on guidelines that states use to satisfy federal reporting requirements. Physician services are those provided by or under the supervision of a physician (M.D. or D.O.) (or a dentist if the state allows dentists to provide dental medicine or dental surgery services). The category of physician services excludes outpatient psychiatric services, which are categorized as psychiatric services.

We constructed a comprehensive analytic file by merging the child-year data from the prepared PS and RX files that resulted from the data preparation steps described above. We added an income eligibility limit variable, which varied across states and age groups; we constructed the variable on the basis of information available in the ESPC Database. The eligibility limit is expressed as a percentage of the federal poverty level.

We excluded records corresponding to individuals under age two or over age 20. We also excluded observations corresponding to “dual-eligibles” (persons dually eligible for Medicare and Medicaid), undocumented immigrants, and other children with limited benefits. Also, to further address the issue of churning in Medicaid (10), we required that a child was eligible for at least eight months in a particular year or that a child with fewer than eight months of eligibility had ineligibility spells lasting no more than two consecutive months. This churning adjustment excluded 5,281,769 child-years. After applying all these criteria, the study sample included 17,259,423 child-years (5,843,711 children).

Medicaid antipsychotic prescription data suffer from incomplete observation or censoring. Censoring occurs when the researcher cannot observe events that take place either before or after a certain point in time (11). In the present context, censoring is important because we did not observe all children over the entire study period. Some children became eligible for Medicaid after the start of the study period (January 1, 2005), whereas others lost their eligibility before the end of the study period (December 31, 2009). Still others may have had multiple spells of eligibility. Censoring-adjusted estimates are critical to assessing whether differences are a genuine issue or arise simply because different race-ethnicity groups have different patterns of gaining and losing Medicaid eligibility. We estimated a proportional hazard model, which allows consistent estimation of the coefficients (representing relative changes in antipsychotic prescriptions) under minimal assumptions (11). We report results on the basis of the Weibull distribution, but we obtained similar results with other commonly used specifications (Gompertz and log-log).

We further addressed our research question by using pooled OLS, which combined all child-year observations and treated them as independent data points (11). Although pooled OLS does not adjust for censoring, it has two advantages. First, it makes minimal distributional assumptions. Second, its coefficients are straightforward to interpret: they represent the marginal effect of the covariate on the probability of antipsychotic prescription. For this reason, the model is often referred to as a linear probability model. Like the proportional hazard model, pooled OLS allowed us to control for several confounders simultaneously. We did not use regressions with child-level fixed effects because race-ethnicity groups are time-invariant characteristics, and thus their effect cannot be estimated with such models.

We report regression standard errors for transparency. However, their relevance is questionable because our analyses included the entire population of children in each state and age group rather than a sample of children. We conducted all analyses using Stata, version 11.0.

Summary statistics for all variables included in the analytic file are provided in Table 1. The mean value of the indicator for any antipsychotic prescription showed that 2.0% of the child-years in the sample had an antipsychotic prescription fill. Physician services were most common, with 3.3 claims per child per year, followed by psychiatric services, with 3.0 claims per child per year. The sample contained similar proportions of African-American and white children (37% and 38%, respectively), followed by Hispanic children (15%), children of unknown race (6%), American Indian/Alaska Native children (2%), Asian children (1%), and children of other races (<1%). Children in the 5–11 age group, the largest single age group, accounted for 41% of the records.

Table 2 presents results from a proportional hazard regression that modeled time to antipsychotic prescription as a function of demographic variables and other covariates. The estimates provide further evidence of the presence of strong racial-ethnic differences. The hazard rate for whites (the reference group with a coefficient of unity) was more than double that for African Americans, Hispanics, and Asians. Specifically, the hazard ratio for whites was more than three times that of Hispanics.

Table 3 presents the results of a pooled OLS regression of antipsychotic prescription status on demographic characteristics and other covariates. The results reinforce the findings produced by the proportional hazard model: white children (reference category with a coefficient set to 0) had a higher antipsychotic fill probability. Holding everything else constant, the probability of having an antipsychotic fill was 1.8 percentage points lower for African Americans, 1.5 percentage points lower for American Indian/Alaska Natives, 2.0 percentage points lower for Asians, 1.8 percentage points lower for Hispanics, and 1.3 percentage points lower for Hawaiian or multiracial children than it was for white children. All race-ethnicity estimates were statistically significant (p<.001).

Table 4 provides the results of a pooled OLS model, with an indicator for any psychiatric service utilization as the dependent variable. The indicator for any psychiatric services is equal to unity if the child received psychiatric services during the year and is zero otherwise. Whites (reference category with coefficient set to zero) were more likely than any other racial-ethnic group to receive psychiatric services, followed by Hawaiian or multiracial children, African Americans, American Indians/Alaska Natives, Hispanics, and Asians. Asians were the least likely to receive psychiatric services. Holding everything else constant, Asians were 6.6 percentage points less likely than whites to receive any psychiatric services.

Cross-tabulations (not shown) of child-years by any antipsychotic drug and any psychiatric service indicators provided evidence that it was not uncommon for a child to receive antipsychotic drugs without any accompanying services during the same year. The proportion of child-years with antipsychotic prescription drugs but no psychiatric services claims was 29.0% among white children and 23.0% among nonwhite children.

Our proportional hazard model showed that the hazard rate for white children (the reference group with a coefficient of unity) was more than double the hazard rates for African-American, Hispanic, and Asian children. Our pooled OLS regressions indicated that compared with whites, the probability of having an antipsychotic prescription fill was lower by 1.8 percentage points for African Americans, by 1.5 percentage points for American Indian/Alaska Natives, by 2.0 percentage points for Asians, by 1.8 percentage points for Hispanics, and by 1.3 percentage points for Hawaiian or multiracial children. The probability of an antipsychotic prescription fill across child-years was only 2.4%, which suggests that the magnitude of these effects is large. Magnitudes produced by the proportional hazard model were similar. However, these marked differences in antipsychotic prescription were smaller in magnitude than those implicit in a recent study of children enrolled in Medicaid in seven states in 2004 (6). That study provided unadjusted rates of antipsychotic prescription by racial-ethnic groups. This difference may result from the effect of confounders on unadjusted rates.

The pooled OLS regression analysis showed that differences persisted after the analysis controlled for a large number of covariates, including state, year, age group, gender, eligibility category, and, especially relevant, number of services received. When psychiatric services (and all other characteristics) were held constant, Asians were 2.0 percentage points less likely than whites and Hispanics were 1.8 percentage points less likely than whites to fill an antipsychotic prescription. This finding allows us to conclude that more prevalent use of antipsychotic drugs among whites was not explained by higher use of mental health services. Similarly, we found that nonwhite children were less likely than white children to receive psychiatric services after the analysis controlled for antipsychotic prescription. Thus differences in use of psychiatric services persisted after the analysis controlled for the use of these powerful psychotropic drugs. Altogether, the results raise the concern that a proportion of children might be receiving these drugs without sufficient oversight by mental health care providers, a possibility that has also raised concerns among a panel of state Medicaid medical directors (12). Indeed, in our data, a cross-tabulation for white children revealed that 29% of child-year observations with any antipsychotic drug use were unaccompanied by psychiatric services; this figure was 23% for nonwhites.

We now focus on differences in antipsychotic use among Hispanics because this group displayed one of the most marked contrasts with whites and also because several studies have explored the use of psychotropic drugs and psychiatric services, and the results of these studies provide context for our results. There is evidence that mental health status and service use vary among Hispanic subpopulations (13). However, nationality is not available in the MAX data. Therefore, our discussion focuses on Hispanics in general (including individuals identified as Hispanic or Latino, or Hispanic or Latino and one or more races), and our results should be interpreted as an overall effect across nationalities.

One possibility that emerges from the literature is that Hispanics are averse to pharmacological treatment but not to other mental health services (14). However, results of this study indicate that Hispanics were less likely to use each type of service, even after the analysis held constant the use of other services. Thus an argument based on higher acceptance of pharmacological treatment among whites does not fully explain low antipsychotic drug use among Hispanics because it fails to explain why Hispanic children with the same antipsychotic drug utilization as white children have lower levels of psychiatric service use than white children.

Previous studies have shown that Hispanics’ mental health may be worse than that of other racial-ethnic groups (15). The literature also indicates that there is very little difference among racial-ethnic groups in the identification of mental disorders and encouragement to seek treatment (16). However, we found that Hispanics use fewer mental health services (both psychiatric services and antipsychotic drugs). This finding suggests that differences in the prevalence of mental health problems do not explain differences in antipsychotic and psychiatric service utilization among children of various racial-ethnic groups. A major limitation of this study is that it did not provide evidence about whether antipsychotic drugs are underprescribed to Hispanic children and those from other minority groups or whether these drugs are overprescribed to white children, because clinical information was not included.

The study found strong racial-ethnic differences in mental health treatment among children enrolled in Medicaid. White children were more likely than nonwhite children to receive both antipsychotic drugs and psychiatric services. Differences in mental health needs or differences in levels of aversion to pharmacological treatment were insufficient to explain the differences.