Analysis of Smoking Trends with Incomplete Longitudinal Binary Responses

Abstract

The generalized estimating equations procedure (GEE) widely applied in the analysis of correlated binary data requires that missing data depend only on remote covariates or that they be missing completely at random (MCAR); otherwise GEE regression parameter estimates are biased. A weighted generalized estimating equations (WGEE) approach that accounts for dropouts under the less stringent assumption of missing at random (MAR) through dependence on observed responses gives unbiased estimation of parameters in the model for the marginal means if the dropout mechanism is specified correctly. WGEEs are applied in the estimation of 7-year trends in cigarette smoking in the United States from a cohort of 5,078 black and white young adults. Analysis using WGEE suggests that there was a general decline in cigarette smoking only among white females, whereas the only other subgroup for which smoking declined was white males of the older birth cohort (1955–1962) with college degrees. The results of WGEE are compared to a likelihood-based method valid under MAR that does not require specification of a missing data model.

Key Words:

• Dropouts
• Maximum likelihood
• Missing at random
• Weighted generalized estimating equations