Due to the non-randomized nature of a retrospective observational study, a propensity score analysis was performed to yield a balanced distribution of baseline characteristics (including the cardiovascular risk profile) and to estimate finasteride effects on patient outcomes between the treatment and control groups. Briefly, for the final study population a propensity score was calculated using a logistic regression model, in which the treatment exposure (finasteride) was regressed as dependent variable on relevant baseline characteristics. To prevent misspecification of the propensity score model and related biases, it is recommended to include baseline variables related to the outcome53 (link), known major risk factors for the outcome54 (link),55 (link) and direct causes of the treatment and outcome56 (link), while inclusion of colliders or mediators should be avoided57 ,58 (link). Hence, the following baseline variables were included in the propensity score to achieve covariate balance of known major cardiovascular risk factors or confounders of cardiovascular treatment effects: age59 (link), diabetes60 (link), history of hypercholesterinaemia52 (link), hypertension61 (link), smoking history62 (link), body mass index63 (link), COPD64 (link), systolic65 (link) and diastolic blood pressure66 (link), heart rate67 (link), ACE inhibitors68 (link) or ARB69 (link), ß-blocker70 (link), MR-antagonists71 (link), aspirin72 (link), statins73 (link). In addition, the underlying prostate disease status was included as it might affect treatment and prognosis of the patients74 (link),75 (link). Variables included in the propensity score to achieve covariate balance are listed in Table 1 .
Medical cases of treatment and control group were matched on the logit of the estimated propensity scores (1:1 propensity score matching) using calipers width equal to 0.02 of the standard deviation of the logit. While in general, higher caliper widths may result in reduced variance and an increased number of matched subjects, this could on the other hand decrease balance between groups and introduce more bias in estimating treatment effects (trade-off between variance and bias). In our study a lower caliper width (0.02) was therefore used in order to maximize correct matching and to reduce bias; This caliper width has been used by others previously in similar studies76 (link)–78 (link). Ongoing research addresses the choice of optimal caliper width during propensity score based matching: one study proposed to use a caliper width equal to 0.2 of the standard deviation of the logit of the propensity score, which may need to be taken into account when interpreting our results79 (link). Absolute standardized difference ≤0.1 for measured covariates suggested appropriate balance between the groups (Table1 and Fig. S1 in the Data Supplement).
Medical cases of treatment and control group were matched on the logit of the estimated propensity scores (1:1 propensity score matching) using calipers width equal to 0.02 of the standard deviation of the logit. While in general, higher caliper widths may result in reduced variance and an increased number of matched subjects, this could on the other hand decrease balance between groups and introduce more bias in estimating treatment effects (trade-off between variance and bias). In our study a lower caliper width (0.02) was therefore used in order to maximize correct matching and to reduce bias; This caliper width has been used by others previously in similar studies76 (link)–78 (link). Ongoing research addresses the choice of optimal caliper width during propensity score based matching: one study proposed to use a caliper width equal to 0.2 of the standard deviation of the logit of the propensity score, which may need to be taken into account when interpreting our results79 (link). Absolute standardized difference ≤0.1 for measured covariates suggested appropriate balance between the groups (Table
Free full text: Click here
