Racial-Ethnic Differences in Incident Olanzapine Use After an FDA Advisory for Patients With Schizophrenia

We used inpatient, outpatient, and pharmacy claims from the Florida Medicaid program for fiscal years 2001–2006. This study received a waiver from the Harvard Medical School Institutional Review Board.

We included Medicaid enrollees with at least one second-generation antipsychotic fill (for olanzapine, quetiapine, risperidone, ziprasidone, or aripiprazole) between January 1, 2001, through December 31, 2006. We defined the first fill date during this period as the enrollee’s index prescription-fill date. We measured incident antipsychotic use during six time segments based on the timing of relevant publications, media reports, FDA regulatory activities, and Florida Medicaid policy changes. [A table outlining the time frames is available online as a data supplement to this report.]

We restricted the cohort to adults ages 18–64 not dually enrolled in Medicare or a managed care plan because such enrollment would mean we would lack access to their complete service utilization records. We identified enrollees as having schizophrenia if they had at least one inpatient or two outpatient claims indicating schizophrenia (ICD-9-CM code 295.x). To assess prior diagnoses that might influence an enrollee’s likelihood of receiving olanzapine, we required enrollees to have continuous Medicaid enrollment for six months before their index date. Finally, we restricted the sample to incident antipsychotic users, excluding those with an antipsychotic fill in the 180 days prior to their index fill date.

We assessed the likelihood of initiating olanzapine versus any other second-generation antipsychotic agent among new second-generation antipsychotic users. We coded race-ethnicity as white, black, or Hispanic using Florida Medicaid enrollment files. We excluded individuals identified as “American Indian” or “Asian” because of small samples (<1%), and we excluded the 14.7% of enrollees classified as “other race” to avoid racial-ethnic misclassification.

We assessed racial-ethnic differences by implementing the Institute of Medicine (IOM) definition of health care disparity, which states that service use comparisons by race and ethnicity should adjust for racial-ethnic differences in need for services (9–11). The implementation followed a three-step process: first, using all available covariates, we estimated a regression model that best fit the data; second, we balanced racial-ethnic groups on characteristics that would influence clinical decision making for prescribing olanzapine rather than an alternative second-generation antipsychotic (hereafter called “need” variables); and third, we estimated predictions by using coefficients from the first step to adjust values from the second step.

To complete the first step, we used modified Poisson regression models to estimate changes in olanzapine use among incident users after the advisory, estimating generalized linear models within Stata with a Poisson distribution and log link. Covariates were split into need and non-need variables. Need variables, measured in the six months before the enrollee’s index prescription-fill date, were age, inpatient mental health services use, and prior diagnoses of metabolic-related conditions and comorbid mental health conditions that could affect antipsychotic selection, complicate the treatment of schizophrenia, or indicate a specific treatment. [Programmatic definitions of covariates are provided online in a data supplement to this report.] Non-need variables were Medicaid eligibility category (Supplemental Security Income [SSI] or “other”), Medicaid district (districts 1–11, representing Medicaid-defined geographical areas within Florida), sex, and whether the enrollee was receiving care from a psychiatrist. Second, we used a rank-and-replace method (11) to equalize need across racial-ethnic groups, allowing us to match the health care needs and preferences of black and Hispanic enrollees to those of white enrollees. Third, we calculated, using the coefficients and non-need values from the original regression model and the adjusted need covariate values, the predicted risk of receiving olanzapine for each individual from a racial-ethnic minority group.

In sensitivity analyses we compared the IOM-concordant racial-ethnic differences described above with estimates of differences both with and without adjustment for all available covariates. First, we compared the unadjusted means between racial-ethnic groups. Second, we estimated the residual direct effect (RDE) of race-ethnicity. The RDE model implicitly defines a disparity as the residual racial-ethnic difference remaining after adjustment for all observable factors. Next, to test the robustness of the IOM model estimates with our selection of need-based covariates, we estimated separate models. In the first, age was excluded, and in the second, the enrollee’s sex and receipt of care from a psychiatrist before the incident drug fill were included as need-based predictors. Finally, we restricted the IOM analyses to individuals receiving Medicaid through SSI to account for differences in severity by eligibility category. We also separately restricted analyses to individuals who were not previously hospitalized to avoid misclassification of the incident prescription because medications initiated within the hospital are not observable in Medicaid claims.

Our sample included 7,901 new users of a second-generation antipsychotic. Of the incident users, 45% (N=3,532) were white, 32% (N=2,518) were black, and 23% (N=1,851) were Hispanic. From the preadvisory period through the most recent study period, we observed large differences in incident olanzapine use among Hispanics compared with whites (Table 1). [Differences are also shown in a figure available as a data supplement to this report.] Before the advisory and consensus statement publication, approximately 57% (N=757) of incident second-generation antipsychotic fills among Hispanics were for olanzapine, compared with only 40% for whites (N=1,050) or blacks (N=737). The absolute difference in the risk of receiving olanzapine between white and Hispanic second-generation antipsychotic initiators was 17% during the preadvisory period (adjusted risk difference [ARD]=.17, 95% confidence interval [CI]=.13–.20). Olanzapine use declined among all groups over each subsequent time point. By the initial postadvisory period, olanzapine was initiated for only 17% of black (N=32) or white (N=38) and for 24% of Hispanic (N=32) users of second-generation antipsychotic medication.

Although the likelihood of receiving olanzapine declined significantly among all second-generation antipsychotic initiators, the relative risk of receiving olanzapine declined only slightly over the study period. For example, Hispanics were 41% more likely than whites to receive olanzapine during the preadvisory period, and an increased relative risk was observed over the study period. However, by the latest study period, the absolute difference in risk was only 3% between white and Hispanic enrollees. There were no differences between whites and blacks in the likelihood of receiving olanzapine over any period studied.

Unadjusted model estimates were similar to those generated with the IOM-concordant method, which adjusted for need variables. In the unadjusted model, Hispanics had a 16% absolute difference in the risk of receiving olanzapine in the preadvisory period compared with whites (risk difference=.16, CI=.13–.19). In the initial postadvisory period, the risk difference was 7% (risk difference=.07, CI=.03–.11).

Estimates of the residual direct effect of race show more modest differences between whites and Hispanics. Using this model, we found that the risk of receiving olanzapine was only 49% for Hispanics and 40% for whites during the preadvisory period. This resulted in an absolute risk difference of approximately 8% (ARD=.08, CI=.05–.11). As in the IOM-concordant method, olanzapine use declined among each racial-ethnic group over each subsequent period. However, the magnitude of the difference in risk between whites and Hispanics was smaller and was not statistically significant in the postadvisory periods.

Results from models using the two-alternative specification of clinical need, described previously, were largely unchanged, suggesting that the IOM-concordant method was robust to potential misspecification. In addition, sensitivity analyses restricted to individuals with SSI coverage and analyses restricted to individuals without a history of inpatient treatment were consistent with our primary analysis, although differences between groups without prior hospitalizations were more modest.

In this sample of new second-generation antipsychotic users diagnosed as having schizophrenia, Hispanics were more likely than whites to receive olanzapine. Absolute differences were greatest in the preadvisory period and decreased over time, resulting in small absolute differences in the risk of receiving olanzapine between groups during the most recent time points. However, the relative risk of olanzapine use was similar for Hispanics and whites at each time point, suggesting that recommendations were applied uniformly across racial-ethnic groups. Although similarity across racial-ethnic groups is encouraging, this finding also suggests that targeted efforts to further reduce use among persons from racial-ethnic minority groups may help to close the gap between white and nonwhite medication users when differences in preadvisory medication use exist.

There are several potential explanations for our findings that Hispanics were more likely than whites to receive olanzapine. First, some evidence suggests that Hispanic patients may be less likely than whites to adhere to antipsychotic treatments (12). Because olanzapine has a lower rate of discontinuation compared with other antipsychotic agents (13), it might be preferred as a first-line treatment among patients considered at higher risk of discontinuation and nonadherence. Moreover, the physician’s assessment of illness severity and likelihood of adherence may be complicated by language or cultural barriers, potentially resulting in physicians relying on olanzapine as a primary treatment among nonnative English speakers.

Although we observed large declines in olanzapine use among all racial-ethnic groups, the metabolic risk advisory is likely only responsible for a portion of this reduction in use. Several major marketing, policy, and clinical information changes occurred after the advisory period (specifically, a classwide black-box warning, publication of the Clinical Antipsychotic Trials of Intervention Effectiveness [CATIE] trial [13], and a temporary change in Florida Medicaid policy), each of which may have affected drug utilization. In addition, although the decrease in olanzapine use in the early postadvisory period would suggest improved quality of care, this interpretation is less clear after the publication of the CATIE trial (where, compared with other treatments, olanzapine was associated with a longer period until discontinuation).

We found that the magnitude, and in some cases the significance, of the differences in olanzapine use varied by the method used to define racial-ethnic disparities. The primary difference between these models is that the IOM-concordant method allows for disparities estimates to include geographic differences or differences mediated by socioeconomic status, whereas the RDE model explicitly removed the influences of these factors from the racial-ethnic disparity estimate. Differences in effects estimated between these two methods suggest that socioeconomic and geographic factors may play an important role.

There were large imbalances in geographic distribution of racial-ethnic minority group enrollees within the Florida Medicaid districts. For example, approximately 85% of the Hispanic enrollees lived within one district, whereas black and white enrollees were distributed more evenly across districts. [A table showing distribution by district is available online as a data supplement to this report.] The observed differences between the risk differences and IOM-concordant estimates may be due to differences in practice patterns by district. For example, if Hispanic patients are more likely to receive care from only a few providers within a region, the differences observed may be concentrated within clinics or by physician and not widespread. If this is the case, targeting clinics serving a high percentage of minority groups may help to eliminate differences in prescribing patterns where they exist. Debate continues about whether geographic differences in prescribing or health services use should be considered in analyses of racial-ethnic disparities (9,14); nevertheless, we found that racial-ethnic prescribing differences existed even after analyses controlled for geographic variation.

Our study had several limitations. First, our sample consisted of fee-for-service Medicaid enrollees from a single state. However, Florida Medicaid provides services to one of the largest Hispanic populations in the United States, and our results are consistent with other work examining patterns of second-generation antipsychotic use among Hispanics, blacks, and whites (15).

Second, our selection of need-based variables was largely based on judgments regarding patient need. In the case of antipsychotics, metabolic conditions and disease severity can guide clinicians’ agent selection within the class. However, some variables may influence prescribing decisions even where evidence of relative benefits is lacking. Because this process is subjective, we modified the need-based variables definition in sensitivity analyses and found that our estimates were robust to changes in model specification.

Third, an ascertainment bias (for example, if persons from racial-ethnic minority groups are less likely to receive treatment for metabolic-related conditions) might influence our model estimates. However, we found little evidence of such a bias, given that racial-ethnic minority groups in our sample had higher rates of metabolic conditions than whites. Fourth, categorization of Hispanic ethnicity was based on Medicaid enrollment files, and details regarding culture of origin were not available. It will be important for future work in this area to identify Hispanic cultural differences in prescribing and the role of patient preferences in medication use decisions.

Fifth, although patient preferences should be measured and included in the IOM disparities estimate, we are unable to measure these in Medicaid claims. Although this remains a limitation of our study, patient preferences are more likely to influence initiation of pharmacotherapy as opposed to agent selection, which we studied in this analysis. Finally, this study focused on second-generation antipsychotic agent switching in response to the metabolic risk messages, but physicians might have used other strategies for managing metabolic risks. Future research is needed to explore the fuller array of management strategies that providers may use to address metabolic risks for patients who are taking antipsychotic medications.

Our findings suggest that even among individuals receiving the same class of therapies, there may be clinically important differences in individual agent use. An important finding was that although there were large declines in olanzapine use among all racial-ethnic subgroups, statistically significant differences in the use of olanzapine were observed even after the FDA advisory and consensus statement were published. This study highlights the importance of communication and dissemination of drug safety information, specifically to underserved and vulnerable populations. As the nation moves toward evaluating and making treatment recommendations based on the relative benefits or harms of available treatments, it will be important to monitor the adoption of these recommendations across diverse patient populations.

This work was supported by grant RO1 HS0189960 from the Agency for Healthcare Research and Quality. Dr. Dusetzina received funding through a Ruth L. Kirschstein–National Service Research Award Post-Doctoral Traineeship sponsored by the National Institute of Mental Health (NIMH) and through grant T-32MH019733-17 from Harvard Medical School. Dr. Cook receives support from NIMH grant R01 MH091042. Dr. Busch received support from NIMH grant K01MH071714. Dr. Huskamp receives support from a Robert Wood Johnson Foundation Investigator Award in Health Policy Research. The funding sources had no role in the design and conduct of the study; collection, management, analysis, or interpretation of the data; and preparation, review, or approval of the manuscript for publication.

Dr. Alexander is a consultant for IMS Health and has served as an ad hoc member of the FDA Drug Safety and Risk Management Advisory Committee. The other authors report no competing interests.