Table 4.
Mean Rate Ratios, Empirical Variance, and Percent Coverage of 95% Confidence Intervals From 1,000 Simulated Data Sets in the Presence of Unmeasured Confounding Due to “Treatment Withheld” Without or With Asymmetric Trimming of the Propensity Score According to Analytic Strategy (True RR = 2.0)
| PS Implementation and Trimming Range | Treatment Prevalence = 0.2 |
Treatment Prevalence = 0.05 |
||||||
| RR | Variancea | MSEb | Coveragec | RR | Variancea | MSEb | Coveragec | |
| Crude | 2.85 | 0.007 | 0.130 | 0.1 | 2.93 | 0.016 | 0.157 | 4.4 |
| True outcome model | 2.00 | 0.003 | 0.003 | 94.0 | 2.00 | 0.007 | 0.007 | 94.1 |
| Outcome model (X1–X6)d | 1.30 | 0.006 | 0.193 | 0.0 | 1.34 | 0.013 | 0.182 | 1.1 |
| PS analysesd | ||||||||
| PS continuous | ||||||||
| Unrestricted | 1.31 | 0.007 | 0.191 | 0.0 | 1.38 | 0.023 | 0.171 | 5.9 |
| Restricted | ||||||||
| 0–100 | 1.31 | 0.007 | 0.190 | 0.0 | 1.41 | 0.019 | 0.146 | 5.9 |
| 1–99 | 1.53 | 0.006 | 0.078 | 1.7 | 1.50 | 0.014 | 0.099 | 14.0 |
| 2.5–97.5 | 1.75 | 0.006 | 0.004 | 38.5 | 1.63 | 0.014 | 0.059 | 39.3 |
| 5–95 | 1.95 | 0.005 | 0.006 | 91.0 | 1.75 | 0.014 | 0.033 | 71.9 |
| PS quintiles (Mantel-Haenszel) | ||||||||
| Unrestricted | 1.52 | 0.008 | 0.086 | 2.1 | 1.61 | 0.016 | 0.066 | 23.9 |
| Restricted | ||||||||
| 0–100 | 1.50 | 0.007 | 0.092 | 1.5 | 1.58 | 0.015 | 0.073 | 19.6 |
| 1–99 | 1.65 | 0.006 | 0.045 | 11.7 | 1.63 | 0.012 | 0.056 | 33.8 |
| 2.5–97.5 | 1.82 | 0.005 | 0.015 | 59.7 | 1.71 | 0.013 | 0.040 | 56.6 |
| 5–95 | 1.98 | 0.005 | 0.005 | 91.8 | 1.80 | 0.014 | 0.027 | 77.5 |
| PS quintiles (standardized mortality ratio) | ||||||||
| Unrestricted | 1.46 | 0.009 | 0.109 | 1.3 | 1.60 | 0.016 | 0.070 | 22.4 |
| Restricted | ||||||||
| 0–100 | 1.45 | 0.008 | 0.116 | 1.1 | 1.57 | 0.015 | 0.077 | 19.0 |
| 1–99 | 1.59 | 0.006 | 0.060 | 6.7 | 1.62 | 0.013 | 0.059 | 30.9 |
| 2.5–97.5 | 1.78 | 0.006 | 0.020 | 49.4 | 1.70 | 0.013 | 0.042 | 53.6 |
| 5–95 | 1.97 | 0.005 | 0.006 | 91.3 | 1.79 | 0.014 | 0.028 | 76.7 |
| PS matching | ||||||||
| Unrestricted | 1.22 | 0.008 | 0.255 | 0.0 | 1.28 | 0.033 | 0.250 | 9.5 |
| Restricted | ||||||||
| 0–100 | 1.22 | 0.008 | 0.254 | 0.0 | 1.28 | 0.033 | 0.249 | 9.4 |
| 1–99 | 1.45 | 0.008 | 0.113 | 1.4 | 1.49 | 0.030 | 0.127 | 33.3 |
| 2.5–97.5 | 1.69 | 0.008 | 0.037 | 38.5 | 1.63 | 0.031 | 0.080 | 60.7 |
| 5–95 | 1.94 | 0.008 | 0.009 | 89.2 | 1.77 | 0.034 | 0.054 | 81.7 |
| Inverse probability of treatment weighting | ||||||||
| Unrestricted | 1.31 | 0.007 | 0.192 | 0.0 | 1.57 | 0.014 | 0.076 | 50.2 |
| Restricted | ||||||||
| 0–100 | 1.29 | 0.007 | 0.203 | 0.0 | 1.52 | 0.013 | 0.093 | 38.6 |
| 1–99 | 1.62 | 0.005 | 0.052 | 19.3 | 1.69 | 0.013 | 0.045 | 70.4 |
| 2.5–97.5 | 1.82 | 0.005 | 0.014 | 75.2 | 1.77 | 0.013 | 0.029 | 83.2 |
| 5–95 | 1.99 | 0.005 | 0.005 | 94.5 | 1.85 | 0.014 | 0.022 | 91.1 |
Abbreviations: MSE, mean squared error; PS, propensity score; RR, rate ratio.
Variance of treatment effect estimates [log(RR)] over 1,000 simulated data sets.
Mean of [log(RR) – log(2.0)]2 over 1,000 simulated data sets.
Percentage of simulated studies in which the 95% confidence interval includes the true value (RR = 2.0).
Outcome and PS models including all measured covariates X1–X6 but not including unmeasured covariate X7.