Purpose: Decision-makers require cost-effectiveness analyses (CEA) for patient subgroups. In non-randomised studies a key challenge is that the treatment allocation and response surface models may be unknown. The propensity score (PS) can be used in matching, inverse probability of treatment weighting (IPTW) and Genetic Matching (GenMatch) to reduce selection bias due to observed characteristics. These methods have not been tested for estimating subgroup effects in CEA.
Method: This paper extends previous comparisons of PS methods for CEA, to the new context of subgroup analysis. Here, IPTW has potential appeal as it may be relatively statistically efficient. However, if the PS is misspecified, this method may fail to balance covariates in each subgroup. For GenMatch, we propose including subgroup by covariate interaction terms in the balance matrix. We compare these methods in a CEA of Xigris, a pharmaceutical intervention for critically ill patients with sepsis (n=2000), whose effectiveness is anticipated to vary by number of organ failures. In simulation studies we consider the relative performance (covariate balance, bias, root mean squared error [RMSE]) of the methods across the following scenarios: a) Ideal scenario: PS and response surface model correctly specified. b) PS model misspecified by excluding an interaction term. c) Response surface model misspecified, GenMatch given incorrect, initial weights for a covariate of high prognostic importance.
Result: In the case study, following both PS matching and IPTW, covariate balance was poor on key prognostic variables (e.g. for APACHE II P<0.01). For IPTW the variance on the weights was high for both subgroups. Following GenMatch covariates were balanced in each subgroup (e.g. APACHE II P>0.1). For patients with 2 organ failures, the incremental net benefit (λ=£20,000 per QALY) varied by method: from -£42,422 (95% CI: -47,013 to -37,831) for IPTW to -£17,247 (95% CI: -18,955 to -15,539) for GenMatch. In the simulation study, under the ideal scenario, all three methods performed well with low bias and RMSE. Once the PS was incorrectly specified, the bias for both PS methods was seven times that for GenMatch. When GenMatch was given incorrect weights, it still dominated the other methods on RMSE for both subgroups.
Conclusion: CEA should use methods that achieve covariate balance for each subgroup. GenMatch can help minimise bias across a range of circumstances faced in applied CEA.