Table 1.
Empirical type I error rates of different propensity score methods for comparing survival functions between treatment groups.
| Statistical method | Prevalence of treatment | ||
|---|---|---|---|
| 0.05 | 0.10 | 0.25 | |
| Effect in overall population of all subjects | |||
| Stratification (Cox regression stratifying on PS strata) | 0.439 | 0.571 | 0.629 |
| Stratification (stratified log-rank test) | 0.439 | 0.572 | 0.629 |
| IPTW (Cole and Hernán) | 0.070 | 0.072 | 0.051 |
| IPTW (Xie and Liu) | 0.722 | 0.604 | 0.350 |
| Effect in population of treated subjects | |||
| Caliper matching (naïve Cox regression) | 0.013 | 0.006 | 0.010 |
| Caliper matching (Cox regression with robust standard errors) | 0.034 | 0.030 | 0.029 |
| Caliper matching (log-rank test) | 0.013 | 0.006 | 0.010 |
| Caliper matching (stratified log-rank test) | 0.035 | 0.033 | 0.039 |
| Nearest neighbour matching (naïve Cox regression) | 0.012 | 0.008 | 0.073 |
| Nearest neighbour matching (Cox regression with robust standard errors) | 0.033 | 0.030 | 0.144 |
| Nearest neighbour matching (log-rank test) | 0.012 | 0.008 | 0.073 |
| Nearest neighbour matching (stratified log-rank test) | 0.033 | 0.033 | 0.283 |
| IPTW (Cole and Hernán) | 0.009 | 0.006 | 0.006 |
| IPTW (Xie and Liu) | 0.000 | 0.000 | 0.000 |
Note: The cells contain empirical estimates of the type I error rate. These were the proportion of 1000 simulated datasets in which the null hypothesis of no difference in survival functions was rejected at the P < 0.05 level.