Generalized Propensity Score Using Baseline Variables Sample Clauses

Generalized Propensity Score Using Baseline Variables. First, a Xxx PH model that includes covariates measured at baseline only (X1i, X2,i(0)) is considered, as shown in equation (3.3). BL,1 hBL(t|X) = h0(t)exp .βT X1 + x X XX,0 X2(0)Σ (3.3) T T ^ ^ ^ ^ We denote the parameter estimates by βBL,1 and βBL,2. The linear predic- tor of this model is the estimated propensity score, GPSBL,i = βBL,1X1,i + βBL,2X2,i(0), and is used to determine the distance between individuals for matching. Since the GPS is a scalar, we can use standard distance metrics such as the squared linear distance, Qm,BL = (GPSBL,i − GPSBL,ij ), where individual i would be treated and individual ij would be in the corresponding risk set. Though this method presents an opportunity for a straightforward calcu- lation of the GPS and a simple metric for matching, the treatment assignment model is only able to reduce or remove potential selection bias between treat- ment groups due to unbalanced baseline variables. If treatment assignment was decided at first clinic visit, this method would be sufficient. However, the decision for treatment is generally made after baseline in this population and likely corresponds to disease progression, implying it is necessary to control for the treatment assignment mechanism that depends on time-varying covariates.