Fitting marginalized two-part models to semicontinuous survey data arising from complex samples.

OBJECTIVE: To accurately model semicontinuous data from complex surveys, we extend marginalized two-part models to a design-based inferential framework and provide guidance on incorporating complex sample designs. DATA SOURCES: 2014 Medical Expenditure Panel Survey (MEPS). STUDY DESIGN: We describe the use of pseudo-Maximum Likelihood Estimation and Jackknife Repeated Replication for estimating model parameters and sampling variance, respectively. We illustrate our approach using MEPS, modeling total healthcare expenditures in 2014 as a function of respondents' age and family income. We provide SAS and R code for implementing the extension, assessing model-fit indices, and evaluating the need to incorporate complex sampling features. DATA EXTRACTION METHODS: Data obtained from www.meps.ahrq.gov. PRINCIPLE FINDINGS: A 100 percentage-point increase in family income as a percent of the federal poverty level was associated with a 5%-6% increase in healthcare spending. People over 65 had an increase of 4-5 times compared to those younger. Accounting for complex sampling in the models led to different parameter estimates and wider confidence intervals than the unweighted models. Ignoring complex sampling could lead to inaccurate finite population inference. CONCLUSION: Researchers should account for complex sampling features when analyzing semicontinuous data from surveys.

Duke Scholars

Author Valerie A. Smith Population Health Sciences