Changes in Elderly Immigrants' Health and Healthcare Service Use by Race/Ethnicity in Response to PRWORA: An Analysis Using Multilevel and Population-Average Models

Purpose: PRWORA's requirements of citizenship for public assistance might have different effects on minority elderly immigrants (EIs) who have different levels of acculturation to life in the U.S. To correctly estimate changes in patterns of EIs' health and healthcare service use in response to PRWORA, the selection of accurate analytical methods is critical. Although immigrants within the same race/ethnic group likely share similar culture/belief systems and their acculturation to the U.S. is differently influenced by policies, few studies have incorporated multilevel modeling approaches (MLM) to correct for the correlation within-cluster . Differences between individuals within the same race/ethnic groups should not be interpreted in the same way as differences between individuals from different race/ethnic groups . When analyzing complex survey data incorporating multiple-stratification, clustering and unequal-weighting of samples of subpopulations, analysts only address design-based bias through sampling weights for population-average models. MLM, however, requires different sampling weights. As part of a research study designed to determine effects of PRWORA on EIs' health and healthcare service use, I contrasted the two approaches.

Methods: The aggregated NHIS data (1993-1996 and 2002-2008 except 2004) included 4,909 Hispanic, 4276 White, 587 Black, 597 Chinese, 552 Filipino, and 959 other API elderly immigrants. I contrasted the results from a logistic model adjusting for design-based bias in population-average models to those from MLM correcting for the model-based correlation with two binary response variables, DV2 (2-week doctor visit) and LATOTAL (limitation of activities). For the population-average model, I used SURVEYLOGISTIC in SAS, using Taylor-series linearization and logit-link function, while using MLwiN for model-based sampling weights, which is a designated program for MLM. I used a three-level random intercept logit model and selected MQL to approximate linear transformation for each level and Taylor-series linearization.

Results: Overall, DV2 among EIs significantly increased compared to pre-PRWORA levels. Before PRWORA, both approaches reported no significant differences among race/ethnic groups. After PRWORA, differences among groups were significant. DV2 was reported at a lower level for all minority groups compared to Whites. In particular, all three Asian groups reported significantly lowest levels. As for LATOTAL, on average, EIs' self-reported LATOTAL scores improved. Compared to pre-PRWORA, however, point estimates in MLM reported Chinese and API became significantly worse (from exp(-0.94) and exp(-0.807) to exp(-0.125) and exp(-0.252) respectively). As for the better model selection, MLM had estimates for intercept β0= exp(-3.951) (S.E. = 0.527), while SURVEYLOGISTIC reported β0= exp(-4.57) (S.E.= 0.6163). In addition, much smaller standard errors were observed across almost all parameter estimates in MLM.

Implications: Results revealed MLM's superiority in estimating population parameters in this study of diverse race/ethnic groups and their behaviors. On average, post-PRWORA populations reported an increase in DV2 and decreased LATOTAL, but findings clearly revealed the existence of groups disadvantaged by PRWORA. The results suggest policy makers consider revoking requirements for citizenship for destitute EIs' access to public assistance, or provision of programs for their achievement of citizenship.