Methodologic Implications of Allocating Multiple-Race Data to Single-Race Categories

Abstract

Objective

To illustrate methods for comparing race data collected under the 1977 Federal Office of Management and Budget (OMB) directive, known as OMB-15, with race data collected under the revised 1997 OMB standard.

Data Sources/Study Setting

Secondary data from the 1993–95 National Health Interview Surveys. Multiple-race responses, available on in-house files, were analyzed.

Study Design

Race-specific estimates of employer-sponsored health insurance were calculated using proposed allocation methods from the OMB. Estimates were calculated overall and for three population subgroups: children, those in households below poverty, and Hispanics.

Principal Findings

Although race distributions varied between the different methods, estimates of employer-sponsored health insurance were similar. Health insurance estimates for the American Indian/Alaska Native group varied the most.

Conclusions

Employer-sponsored health insurance estimates for American Indian/Alaska Natives from data collected under the 1977 OMB directive will not be comparable with estimates from data collected under the 1997 standard. The selection of a method to distribute to the race categories used prior to the 1997 revision will likely have little impact on estimates of employer-sponsored health insurance for other groups. Additional research is needed to determine the effects of these methods for other health service measures.

The Office of Management and Budget (OMB) issued revised standards for the collection and reporting of race and ethnicity data within the federal statistical system (Evinger 1997; OMB 1997a, 1997b). The prior standard, known as OMB-Directive 15, had been in place since 1977 (OMB 1977). The revised standard allows individuals to choose one or more race groups when responding to federal surveys and other federal data collections. Although full implementation of the new standard is not mandated until 2003, OMB requires compliance for new data collections. The 2000 decennial census, which collected data under the new standard, will provide benchmarks of both the size and characteristics of the multiple-race groups in the United States.

The change in the collection of race information provides new challenges for health services researchers. Although the implications of including race in studies are often debated (Fullilove 1998; Kaugman 1999; Schulman et al. 1995; White-Means 1995; Williams 1994), reports and studies consistently document disparities between racial groups (e.g., Council of Economic Advisors 1998; Hannan, van Ryn, Bruke, et al. 1999; National Center for Health Statistics [NCHS] 1999; Pamuk, Makuc, Heck, et al. 1998; Shi 1999), some mitigated by other factors and some not (e.g. Makuc, Breen, Freid 1999; Shi 1999).

Although new data collections may provide information about multiple-race populations, preliminary estimates suggest that their numbers will be too small for reliable statistical estimates (OMB 1997a, 1997b; Tucker, McKay, Kojetin, et al. 1996). Furthermore, the effect of multiple-race reporting will have differential effects on single-race groups (Bureau of the Census 1997; OMB 1997a, 1997b; Tucker, McKay, Kojetin, et al. 1996). Consequently, estimates derived for race groups under the 1977 directive will not be comparable to estimates derived for the single-race groups under the 1997 standard. This incompatibility is important because health policies are often based on trends from several years of data and from data collected from different sources.

In 1999, the OMB issued Draft Provisional Guidance on the Implementation of the 1997 Standards for Federal Data on Race and Ethnicity (OMB 1999), available on the Internet, but not in final form. This guidance documentation encourages detailed reporting of all race groups, both single and multiple, as long as requirements for confidentiality and statistical reliability are met. However, this document contains a detailed discussion of what the OMB calls “bridge” methods, which compare or combine data collected under the previous OMB-Directive 15 with data collected under the new standard.

The objectives of this article are to disseminate more widely the OMB's bridge methods (OMB 1999) and to highlight potential differences between these methods for race-specific estimates. For illustration, we examined a commonly reported health policy statistic from the National Health Interview Survey (NHIS), the percentage of persons with employer-sponsored health insurance coverage. As more multiple-race data are collected and reported, methods for comparing data between systems will be of decreasing utility. However, during the transition period, separating real change from variation caused by reporting differences will be critical.

METHODS

The NHIS is a continuous household survey designed to measure the health status of residents of the United States (Benson and Marano 1998). Data from the 1993, 1994, and 1995 surveys were used for this study. Each year there were approximately 45,000 households interviewed, resulting in approximately 100,000 observations. In 1993, the health insurance questions were only asked for half of the year. As a result, our study is based on 251,864 respondents; of these, 4,374 reported more than one race.

To illustrate differences between methods, we chose to examine employer-sponsored health insurance because of its policy importance, as well as the disparate coverage rates by race and other demographic factors (Campbell 1999; NCHS 1999). We assessed differences between the methods for children, Hispanics, and those with household incomes below the poverty level. These factors were chosen for their associations with employer-sponsored health insurance as well as with race or multiple-race reporting (Campbell 1999; Lucas 1999; NCHS 1999).

Since 1976, the NHIS has allowed respondents to choose more than one race (Benson and Marano 1998; OMB 1997). As the respondent is handed a card with numbered race categories, the interviewer asks, “What is the number of the group or groups that represent your race?” If a respondent selects more than one category, the interviewer then asks a follow-up question: “Which of those groups would you say best describes your race?” During 1993–95, the card shown to respondents included 15 separate categories. Consistent with the 1977 OMB directive, the 15 groups were collapsed to four single-race categories: White, Black, American Indian, including Alaska Natives (AIAN), and Asian or Pacific Islander (API). The NHIS added an additional category, Other Race, for respondents who do not choose a race group from the list.

Detailed Race was created from responses to the first question, which allowed respondents to choose more than one race group. This information was not included on public use data files of the NHIS and was not used directly for national estimates from the survey. However, on in-house files, up to two of the race groups mentioned were recorded for each respondent; if more than two groups were selected, only two were recorded on the in-house file. Data from the 1997 NHIS indicate that over 97 percent of multiple-race respondents choose only two groups (Lucas 1999). Detailed race was coded into five single-race groups (White, Black, AIAN, API, Other Race) and four multiple-race groups (White/Black, White/AIAN, White/API, and Other Combinations). The Other Combinations category was created because there were too few observations in other multiple-race groups for specific tabulations.

Responses to the follow-up question were used to create Primary Racial Identification. For single-race responses to the first question, Primary Racial Identification is the same as Detailed Race. For multiple-race responses, Primary Racial Identification is the race response to the follow-up question. We assigned multiple-race respondents who did not report a single-race group in the follow-up question to the Other Race category. Primary Racial Identification is used as the reference standard for comparisons between different methods, under the assumption that the race response to the follow-up question is what would have been the response to a “mark only one” race question (Appendix). The responses to the follow-up question were used, in part, to create the race variable available on public use NHIS data files.

Bridge methods refer to a collection of techniques to reassign multiple-race responses obtained using the 1997 standard into the 1977 directive race categories (OMB 1997). The new standard requires official tabulations to include multiple-race groups that meet criteria for statistical reliability and confidentiality. However, the OMB recognized that approaches to make data comparable would be needed for some analyses. Currently, there are a handful of simple methods being considered by the OMB. Although the best approach for combining multiple and single-race data for any particular study may not be any of the methods presented here, these methods were proposed by the OMB because of their simplicity and their applicability to a wide variety of situations. These methods fall into two broad categories: whole allocation methods, which provide rules for reassigning each multiple-race response into one of the single-race groups selected, and fractional allocation methods, which assign part of each multiple-race response into each single-race group selected (Table 1). Fractional allocation methods can be implemented either deterministicallzy, by modifying observation weights, or probabilistically, by random whole assignment. Although the variances associated with fractional deterministic and fractional probabilistic assignment methods will differ, the expected value of the statistics is the same. For this study, we present results for the deterministic methods. More detailed rationale and explanations of these methods are available in the OMB (1997) report.

Table 1.

Description of Bridge Allocation Methods Being Considered by the OMB

Standard errors for the race-specific statistics classification were estimated using SUDAAN to control for the complex design of the survey (Shah, Barnwell, Bieler, 1997). Standard chi-square statistics were used to assess demographic differences between race groups; differences were considered statistically significant at p < 0.05. For the most part, although not always, the standard errors are similar for estimates calculated for the different bridge methods because nearly all respondents reported one race group and are assigned to the same group under all methods. For clarity, only standard errors for the Primary Racial Identification estimates are presented.

RESULTS

During 1993 through 1995, 1.8 percent of the survey population reported more than one race (Table 2). Of these, approximately half reported AIAN/White, whereas most of the remaining reported Black/White or API/White. Few reported other multiple-race combinations.

Table 2.

Percent Distribution and Percent with Selected Characteristics by Race Group

Age, Hispanic origin, and poverty status differed between multiple-race groups as well as between the multiple-race groups and their single-race counterparts (Table 2). As expected, employer-sponsored health insurance differed between race groups. The estimate for multiple-race AIAN/White was between the corresponding estimates for the single-race AIAN and White groups. The estimate for the API/White group was also between the corresponding single-race estimates, but the range was considerably smaller. Employer-sponsored health coverage was similar among the Black/White and Black groups, but considerably lower than the White group.

Estimated race distributions differed between bridge methods with the race groups being differentially affected (Table 3). The estimated percentage of the population in the AIAN group had the largest difference. For the API, Black, and White groups, the estimated population percentages were within 5 percent of the estimated percent using Primary Racial Identification.

Table 3.

Percent Distribution of Race Groups by Bridge Tabulation Methods

Differences in percent employer-sponsored health insurance across bridge methods (Table 4) were smaller than the differences in race distributions, although the relative impacts for different race groups followed the same pattern. The estimates for API, White, and Black groups were also similar to those reported for the single-race groups using Detailed Race (Table 2), evidence that inferences will be similar for these groups if no bridging were attempted.

Table 4.

Race-Specific Estimates of Percent with Employer-Sponsored Health Insurance by Bridge Tabulation Methods and Selected Characteristics

For the AIAN group, the estimate using Primary Racial Identification was lower than estimates using any of the bridge methods. The small proportion of AIAN/White respondents who chose AIAN as their Primary Racial Identification in the follow-up question (12.4 percent) were less likely to have employer-sponsored health insurance than those who did not choose AIAN, lowering the overall estimate of insurance coverage using Primary Racial Identification.

Estimates for children, those below poverty, and Hispanics followed similar patterns to those observed for the overall population. For example, despite disproportionate numbers of children in the Black/White and the API/White groups, employer-sponsored health insurance estimates for Black and API children under 18 were reasonably close for the different bridge methods, and all were within 1 percent of the corresponding estimate using Primary Racial Identification (Table 4). Predictably, employer-sponsored health insurance estimates for AIAN under 18 had a much wider variation than estimates for children in other race groups. However, the standard errors of estimates for the AIAN populations are relatively large, particularly for this subgroup analysis.

DISCUSSION

Will the choice of bridge allocation method affect future race-specific estimates of health services utilization or outcomes? We found similar estimates from each of the bridge methods using our single example, employer-sponsored health insurance, except for the AIAN group. These similarities were consistent overall as well as for children, those living below poverty, and Hispanics. Estimates of employer-sponsored health insurance for the single-race groups using Detailed Race (Table 2) were also similar to those from the bridge methods (Table 4), suggesting that comparisons between data obtained under the old and new standards may not be unreasonable. From these results, the best general bridge method is not obvious. As with many analytic decisions, there is a trade-off between selecting the best method for the widest number of applications and the best method for a particular situation. Although these findings provide some reassurance that future studies may be made comparable with prior studies, at least for estimates of employer-sponsored health insurance, multiple-race groups in the United States are likely to increase and their demographic compositions are likely to change, not allowing us to make definitive conclusions.

Demographic and socioeconomic differences between single- and multiple-race respondents may affect estimates calculated using different bridge methods. As shown briefly in this study (Table 2) and documented more fully elsewhere (Lucas 1999; Parker and Lucas 2000), multiple-race respondents differ in many ways from their single-race counterparts. These differences are not surprising, given the wide socioeconomic disparities that exist between single-race groups (Williams 1994). Corresponding differences in health services utilization are expected. The choice of bridge method will have a larger impact when the demographics and subsequent health outcomes for a multiple-race group differ from those of its component single-race groups. Accordingly, in this study, the bridge methods had a considerably larger effect for the AIAN estimates than for the API estimates. Although the differences here did not lead to large differences in estimates of employer-sponsored health insurance between bridge methods, it is likely that other indices or statistics will be more affected by demographic and socioeconomic differences between single- and multiple-race groups.

The large variation in social factors that are known within single-race groups will be evident within multiple-race groups as well. For example, multiple-race respondents likely differ by their response to the follow-up question, and by extension, how they responded to race questions under the former directive. Although our data did not support a detailed analysis of these responses, the socioeconomic and demographic characteristics of the approximately 80 percent of AIAN/White respondents with White as their Primary Racial Identity are more economically favorable than those of the remaining AIAN/White respondents (data not shown). The greater the association between socioeconomic profile and measures of health services utilization, or outcomes, the greater the effect of the assignment of multiple-race respondents to a single-race category on the subsequent estimate.

This study highlights future problems analysts and policy makers will have comparing AIAN estimates under the new and old OMB standards. The largest group and plurality bridge methods provided AIAN estimates that were closest to those based on Primary Racial Identification. On the one hand, most federal data sources currently do not have sufficient numbers of AIAN respondents using single data years to meet confidentiality and reliability standards (NCHS 1999). However, even for data sources that currently support AIAN statistics, these results indicate that AIAN data collected under the old OMB standard and AIAN data collected under the new standard will not be comparable or easy to bridge. In addition to the demographic differences that affect these statistics, the single-race AIAN group is small compared with the multiple-race White/AIAN group; as a result, the assignment of the White/AIAN respondents has a large bearing on the single-race AIAN estimates.

The impact of the bridge methods will likely vary between surveys, populations, time, and, most importantly, the question at hand. Allocation methods derived solely from reported race groups will not have nearly the predictive power as methods developed from predictive models that incorporate demographic and geographic information. Although models are being developed using the NHIS data, their implementation will certainly be more complicated than these simple allocation techniques. Furthermore, models derived from an existing single data source may not provide better bridging methods for estimates from other data systems, time points, and geographical locations than the more simple bridge methods.

This study illustrates the future complexities of combining race data from different years or different sources. Although our single example of employer-sponsored health insurance does not show a large impact of bridge methods on statistics for most race groups, it is likely that other health measures or population subgroups may be differentially affected.

Appendix