OBJECTIVES: To estimate the impact of varying selection bias correction techniques on the average treatment effect. The structural parameters, heterogeneous (ATE), and homogenous (ATE1) average treatment effects were defined as the impact of treatment on total quarterly expenditure, if patients are randomly assigned to biologic disease modifying anti-rheumatoid drugs (DMARDs). METHODS: Retrospective cohorts were constructed from California Medicaid paid claims from January 1, 1999 to December 31, 2005. Non-overlapping quarters were created from pharmacy claims for biologic (adalimumab and etanercept) and standard (lefluonomide, hydroxychloroquine and sulfasalazine) DMARDs. Final sample included 24,504 episodes on 5,510 patients. In the two-stage estimation, the treatment selection model was estimated by multinomial-logit. The outcome model was fixed-effects correlated random coefficients model (Wooldridge-2005), allowing parameter heterogeneity. A generalized residual function constructed based on four different bias-correction techniques for the multinomial-logit selection model namely, Lee's (1983) (LEE), Dubin and MacFadden's (1984) (DMF) and two variants of Dahl's (2002) approach (squared (DHL1) and quadratic (DHL2) series expansions without interactions); controlled endogeneity in treatment choice. The generalized residual was regressed on exogenous covariates to assess multicollinearity. Hypothesis testing was based on cluster-bootstrapped errors. RESULTS: Multicollinearity was not an issue for LEE (RSQUARE=0.07) and DMF (RSQUARE=0.08) approaches, however, it was strong in DHL1 (RSQUARE=0.48) and DHL2 (RSQUARE=0.62), even with six exclusion restrictions. Time-varying endogeneity was significant under all approaches. Controlling endogeneity significantly increased ATE1 for adalimumab, in LEE and DMF approaches but decreased in magnitude in DHL1 and DHL2 as compared to naïve fixed-effects model. When heterogeneity in parameters was allowed, ATE of adalimumab was significantly higher as compared to standard DMARDs, under all bias-correction techniques. ATE for etanercept under LEE ($46,187, p=0.75) and DMF ($76,393, p=0.59) was not significant. However, ATE for etanercept under DHL1 ($192,813, p<0.001) and DHL2 ($191,463, p<0.013) was significantly higher as compared to standard DMARDs.