Online Appendix A: Data generation process for the simulation study The following steps describe the general data generating process. ο· Choose the sample size n, the total number of study subjects per data set. ο· For subject i (π = 1, β― , π), the observed data {ππ , ππ , πΏπ , ππ , ππ ; (π‘ππ , πΏππ ): π = 1, … , ππ } are simulated as follows: ο§ The maximum follow-up time π is pre-specified by the study design. Simulate the rightcensoring time πΆπ according to the right-censoring hazard function βπ (π ). Then the length of study interval ππ = min(π, πΆπ ). ο§ π Simulate ππ and πΏπ . Assume πΏπ = (π₯π1 , … , π₯ππ ) and ππ = (1, πΏππ )π , with π₯π1 indicating the group allocation. Of the π (an even integer) study subjects, π/2 are allocated to the control group (π₯π1 = 0) and π/2 to the intervention group (π₯π1 = 1). ο§ Simulate ππ : ππ ~Bernoulli(ππ ) with ππ = {1 + exp(−πππ π½ )}−1. o If ππ = 1, simulate the subject-specific frailty ππ from Gamma ( ψ−1 , ψ ). Suppose ∗ ∗ ∗ π’π0 = 0. Given π’π,π−1 ( π = 1, … ), generate π’ππ , time to the π -th event, using the following event-order specific hazard function ∗ ∗ ∗ ∗ βππ (π’ππ − π’π,π−1 | ππ , πΏπ , ππ = 1) = ππ β0π (π’ππ − π’π,π−1 ) exp(πΏππ π· ). ∗ The observed number of gap times is ππ = min{π: π’ππ ≥ ππ }. The observed number ∗ of events is ππ = max{π: π’ππ ≤ ππ }. Thus, the observed gap-time process for (π‘ππ , πΏππ ) ∗ ∗ ∗ ∗ ∗ is {(π’π1 , 1), β― , (π’π,π − π’π,π , 1), (ππ − π’π,π , 0)}. If ππ = π’π,π then the last duplet π π −1 π π ∗ (ππ − π’π,π , 0, ππ ) is omitted. π o If ππ = 0, then ππ = 0 and the observed gap time process is {(ππ , 0)}. 1 Note that under each simulation scenario, the right-censoring hazard functions βπ (π ) and the first-event baseline hazard function β01 (π ) are chosen such that a negligible number of susceptible subjects have their first events being right-censored after the largest observed uncensored event time within the first event stratum. We consider two sets of (βπ (π ), β0π (π ), π) and details are given as follows: (1) For time to the first event, the baseline hazard is a monotonically increasing function in the form of a Weibull distribution: β01 (π ) = π 5 . And after the first event, the baseline hazard is a constant within each stratum: β0π (π ) = 3/π 2 for 2 ≤ π ≤ 4 and β0π (π ) = 3/16 for π > 4. Three right-censoring settings are chosen: βπ (π ) = 4π /9 with π = 4, βπ (π ) = 4π /13 with π = 3 , and βπ (π ) = 4π /7 with π = 3 , which lead to the right-censoring proportions of, respectively, about 10%, 25% and 40% among the susceptible subjects in the first event strata in the control groups. (2) For time to the first event, the baseline hazard first increases to a peak value and then gradually decreases, which can be formulated as a log-logistic distribution: β01 (π ) = (1/8)21/8 π‘ −7/8 /(1 + (2π‘)1/8 ). After the first event, the baseline hazard is a constant within each stratum: β0π (π ) = 1/π 2 for 2 ≤ π ≤ 4 and β0π (π ) = 1/16 for π > 4 . Three rightcensoring settings are chosen: βπ (π ) = 1/(16π ) with π = 16, βπ (π ) = 1/(6π ) with π = 16, and βπ (π ) = 1/(9π 4/5 ) with π = 16 , which lead to the right-censoring proportions of, respectively, about 10%, 25% and 40% among the susceptible subjects in the first event strata in the control groups. 2 Online Appendix B: Simulation scenario settings To evaluate the performance of the estimator with the three tail completion methods (0TAIL, ETAIL and WTAIL), we consider the following two sets of simulation scenarios: (i) A single binary independent variable: πΏπ = π₯π1 , ππ = (1, π₯π1 )π . Set π½1 = π1 = −0.5, with three values considered for π0 : (i.1) π0 = log(9), (i.2) π0 = log(1.5) and (i.3) π0 = −log(9). In total, there are three parameter specifications for (π0 , π1 , π½1 ), resembling the situations where the non-susceptible (or cured) fractions in the control groups are small (10%), moderate (40%) and large (90%), respectively. (ii) Two correlated covariates: πΏπ = (π₯π1 , π₯π2 )π , ππ = (1, π₯π1 , π₯π2 )π . Here π₯π2 is a quantitative variable defined as π₯π2 = 5 × Beta(2, 5) − 2.5 if π₯π1 = 0 and π₯π2 = 5 × Beta(5, 2) − 2.5 if π₯π1 = 1 , such that π₯π1 and π₯π2 are strongly correlated (correlation coefficient=0.8). Furthermore, the respective expectations of π₯π2 within the control (π₯π1 = 0) and intervention (π₯π1 = 1) groups are approximately −1.07 and 1.07. The notable difference between the control and intervention groups in terms of their distributions of π₯π2 is set up deliberately. In this way, the true values of the three versions of summary PE are distinguishable, so are their corresponding estimates, and this helps to avoid any incidental doubt about mismatch between the parameter estimates and their true identities in the simulation results. Three parameter specifications are considered: (ii.1) π0 = log(9), (ii.2) π0 = log(1.5) and (ii.3) π0 = − log(9) , each with (π1 , π2 , π½1 , π½2 ) = (−0.5, 0.25, −0.5, −0.25). The three parameter settings imply that π₯π2 is a factor representing predisposition to susceptibility, but if a subject is susceptible, π₯π2 becomes a protective factor, protecting from event occurrence/recurrence. 3 For each of the parameter specifications in sets (i) and (ii), two values are considered for the frailty parameter ψ (0.5 or 1), representing two different scenarios in which the variance of the frailty are 0.5 and 1, respectively. Moreover, sample sizes of 500 and 1000 are considered for the simulation scenarios with small or moderate non-susceptible fractions, i.e., π0 = log(9) or log(1.5), whereas 1000 and 3000 for those with high non-susceptible fractions, i.e., π0 = −log(9). These sample size settings correspond to realistic situations where a large sample size is used for a rare event outcome. Therefore, by combination, there are twelve (three parameter specifications × two values of ψ × two sample sizes) simulation scenarios for each set as specified in (i) and (ii) above, and altogether there are twenty-four simulation scenarios. To further evaluate the estimator’s performance with 0TAIL method, the following two additional simulation settings are considered: (iii) Sensitivity analysis: we consider two situations where the frailty follows a Log-Normal or Binary distribution. For clarity, o Log-Normal frailty: ππ = π π£π with π£π ~Normal (− log(1+ψ) 2 , √log(1 + ψ)) . Here Normal(μ, σ) refers to a Normal distribution with mean μ and standard deviation σ; ψ(1−π1 ) o Binary frailty: Pr ( ππ = 1 − √ π1 ψπ ) = π1 and Pr ( ππ = 1 + √1−π1 ) = 1 − π1. 1 2 Here π1 = is arbitrarily chosen to ensure ππ > 0 for ψ = 0.5 and 1. 3 Note that the above two frailty distributions both satisfy the assumptions E(ππ ) = 1 and Var( ππ ) = ψ, so as to provide some qualitative insight into the sensitivity of the estimator’s performance to the Gamma distribution assumption for the frailty but with the same mean and variances. Simulations are conducted with the parameter specifications in (i.2) and (ii.2) for each 4 frailty distribution. The considered sample sizes are 500 or 1000. There are a total of sixteen simulation scenarios in the sensitivity analysis set. (iv) “All susceptible”: We examine the estimator’s performance under the simulation scenarios with all subjects being susceptible and the parameter specifications are: (iv.1) πΏπ = π₯π1 with π½1 = −0.5 and (iv.2) πΏπ = (π₯π1 , π₯π2 )π with ( π½1 , π½2 ) = (−0.5, −0.25) . Under either (iv.1) or (iv.2), the three distributions defined earlier (i.e. Gamma, Log-Normal and Binary) are assumed for the frailty. Two values, 0.5 and 1, are used for the frailty parameter ψ. In the interest of space, we use only the sample size 500 for each scenario. This leads to twelve simulation scenarios in total. Online Appendix C: Interpretation of results for the simulation scenarios for the settings (iii) sensitivity analysis and (iv) “all susceptible” using the 0TAIL method. In what follows, the simulation results concern the scenarios where the baseline hazard of time to the first event follows a Weibull distribution, the right-censoring proportions among the susceptible are about 10% in the first event stratum in the control group, the frailty parameter ψ = 0.5. Results for and conclusions from other simulation settings are similar and thus omitted. (All simulation result tables are available from the first author on request). The results of simulation scenarios with one binary independent variable, π0 = log(1.5), a mis-specified frailty distribution (Log-Normal or Binary) and sample size π = 500 or 1000 are summarized in Online Table 1. In the case of mis-specified frailty distributions, the estimates for the frailty parameter ψ are no longer valid as indicated by the large negative bias. Nonetheless, the estimates with respect to the parameters π½, π½1 and πΎ still perform very satisfactorily. The 5 results of the “two correlated covariates” scenarios with Log-Normal and Binary distributed frailty are similar to their analogues in Online Table 1 and can be found in Online Table 2. Online Table 3 concerns with the simulation scenarios where all subjects are susceptible, the frailty distribution follows Gamma, Log-Normal or Binary form and sample size π = 500. In line with all subjects being susceptible, the average parameter estimates for the intercepts of the logistic regression in predicting the probability of being susceptible are all very large. Across the simulation scenarios, there is no fixed pattern for the average parameter estimates for π1 . This is because the detection of a null intervention effect requires an infinite sample size. Moreover, in the context of studying common (or rare) events, fitting the data using the logistic regression with only one binary independent variable would yield a large positive (or negative) intercept estimate reflecting the high (or low) event prevalence observed in the control group. On this point, as the logistic function is subject to “saturation” at both tails, a small difference between groups in the observed prevalences based on finite sample size would translate into very dramatic and irrational estimate for the intervention effect. The estimated hessian matrix for π½ involved multipliers πΜπ (1 − πΜπ ) where πΜπ was the predicted probability of being susceptible and very close to 1, thus πΜπ (1 − πΜπ ) was close to 0 and the matrix inversion produced very large values for the variance estimates. Nonetheless, under all scenarios regardless of whether the frailty distribution is correctly specified, the estimation results for π½1 and πΎ are satisfactory. Online Table 4 shows the results of the “two correlated covariates” scenarios where all subjects are susceptible, the frailty follows the Gamma, Log-Normal or Binary distribution, and sample size π = 500. We observed very large negative values in the estimates for πΎ1 and πΎ2 under these scenarios. This happens for the following reason. The quantitative variable π₯π2 consisted mostly of negative values in the control group and positive values in the intervention 6 group, and the distributions of these sample value π₯π2 ’s for the two groups may overlap very little due to the finite sample size. If this was the case and in the context of common event outcomes, fitting the logistic regression model to the data tends to produce a large negative value of parameter estimate for π2 and large positive values of parameter estimates for π0 and π1 . As a result, the predicted probabilities and the predicted counterfactuals based on the values of π₯π2 in the control group were all very close to 1, thus the summary PE parameter πΎ0 can be reasonably estimated. In contrast, though the predicted probabilities for the subjects in the intervention group were very close to 1, their predicted counterfactuals (after removing the large intervention effect) might be grossly biased downwards. The resulting estimates for πΎ1, hence πΎ2 which was a weighted average of πΎ0 and πΎ1, would become irrationally small. In return, the averages of these estimates will be affected by these outliers and driven downwards. In light of this, all the summary statistics with respect to the three summary PE parameters πΎπ (π = 0, 1, 2) were calculated after omitting those corresponding to the smallest 5% of the parameter estimates for πΎ1. The choice of the trimming proportion was rather arbitrary, depend on the severity of the outliers and the number of replications, and we chose 5%. The parameter estimates for πΎπ after trimming are very close to the truth. Across all the simulation scenarios, the estimator performance is satisfactory. 7 Online Table 1: Simulation results using 0TAIL method for the scenarios with a single binary independent variable, Log-Normal or Binary distributed frailty and parameter specifications in set (i.2) where π0 = log(1.5). Frailty LogNormal Binary Parameter Sample size True value Estimate Bias Empirical SD π0 π1 π½1 ψ πΎ 500 0.4055 -0.5 -0.5 0.5 0.5184 0.4071 -0.5070 -0.4949 0.3230 0.5114 0.0016 -0.0070 0.0051 -0.1770 -0.0070 0.1361 0.1883 0.1399 0.1098 0.0783 0.1376 0.1935 0.1310 0.0651 0.0762 95.0 96.6 92.0 -94.4 ----94.4 π0 π1 π½1 ψ πΎ 1000 0.4055 -0.5 -0.5 0.5 0.5184 0.4087 -0.5105 -0.5010 0.3534 0.5160 0.0032 -0.0105 -0.0010 -0.1466 -0.0024 0.1300 0.1899 0.1298 0.1164 0.0738 0.1377 0.1936 0.1334 0.0675 0.0765 95.2 94.8 95.0 -95.2 ----96.2 π0 π1 π½1 π πΎ 500 0.4055 -0.5 -0.5 0.5 0.5184 0.4023 -0.5013 -0.4974 0.3477 0.5144 -0.0032 -0.0013 0.0026 -0.1523 -0.0040 0.0988 0.1390 0.0985 0.0859 0.0567 0.0971 0.1365 0.0937 0.0469 0.0540 94.6 94.4 95.4 -94.8 ----94.4 π0 π1 π½1 π πΎ 1000 0.4055 -0.5 -0.5 0.5 0.5184 0.4052 -0.4989 -0.4983 0.3672 0.5141 -0.0003 0.0011 0.0017 -0.1328 -0.0043 0.0993 0.1464 0.1000 0.0869 0.0574 0.0972 0.1366 0.0947 0.0480 0.0544 95.6 93.4 93.8 -93.2 ----92.6 8 Average SE CPN (%) CPLog (%) Online Table 2: Simulation results using 0TAIL method for the scenarios with two correlated covariates, Log-Normal or Binary distributed frailty and parameter specifications in set (ii.2) where π0 = log(1.5). Frailty LogNormal Binary True value Estimate Bias Empirical SD Average SE CPN (%) CPLog (%) 500 0.4055 -0.5 0.25 -0.5 -0.25 0.5 0.5340 0.5026 0.5183 0.4069 -0.4995 0.2557 -0.4852 -0.2466 0.3128 0.5119 0.4755 0.4937 0.0014 0.0005 0.0057 0.0148 0.0034 -0.1872 -0.0221 -0.0271 -0.0246 0.1898 0.3288 0.1193 0.2182 0.0850 0.1075 0.1397 0.1323 0.1352 0.1903 0.3261 0.1224 0.2137 0.0802 0.0653 0.1343 0.1266 0.1305 95.6 94.6 96.0 94.0 93.8 -93.6 94.8 94.6 ------94.6 94.6 94.8 π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 1000 0.4055 -0.5 0.25 -0.5 -0.25 0.5 0.5340 0.5026 0.5183 0.4066 -0.4940 0.2506 -0.5000 -0.2490 0.3393 0.5247 0.4918 0.5083 0.0012 0.0060 0.0006 0.0000 0.0010 -0.1607 -0.0093 -0.0107 -0.0100 0.1291 0.2252 0.0812 0.1569 0.0560 0.0791 0.0953 0.0900 0.0921 0.1343 0.2302 0.0862 0.1534 0.0573 0.0473 0.0935 0.0874 0.0905 96.0 94.6 96.0 95.0 95.4 -94.6 95.0 94.6 ------94.6 94.2 94.6 π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 500 0.4055 -0.5 0.25 -0.5 -0.25 0.5 0.5340 0.5026 0.5183 0.4020 -0.4922 0.2486 -0.4941 -0.2411 0.3347 0.5112 0.4755 0.4934 -0.0035 0.0078 -0.0014 0.0059 0.0089 -0.1653 -0.0228 -0.0271 -0.0249 0.2013 0.3493 0.1301 0.2291 0.0902 0.1121 0.1503 0.1422 0.1455 0.1907 0.3271 0.1227 0.2180 0.0815 0.0680 0.1357 0.1288 0.1323 93.6 92.0 93.0 93.4 92.6 -94.2 94.2 93.8 ------92.0 93.4 92.6 π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 1000 0.4055 -0.5 0.25 -0.5 -0.25 0.5 0.5340 0.5026 0.5183 0.4121 -0.5099 0.2506 -0.5037 -0.2413 0.3523 0.5309 0.4971 0.5140 0.0067 -0.0099 0.0006 -0.0037 0.0087 -0.1477 -0.0031 -0.0055 -0.0043 0.1350 0.2325 0.0846 0.1570 0.0570 0.0799 0.0914 0.0864 0.0883 0.1344 0.2307 0.0864 0.1551 0.0580 0.0487 0.0931 0.0873 0.0902 95.0 95.0 96.0 95.2 96.6 -94.0 95.0 94.6 ------96.0 96.0 96.4 Parameter Sample size π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 9 Online Table 3: Simulation results using 0TAIL method for the scenarios with “all susceptible”, a single binary independent variable (π½1 = −0.5), Gamma (or Log-Normal, Binary) distributed frailty and sample size of 500. Estimate Bias Empirical SD Average SE 32.5254 -0.1262 -0.4959 0.4659 0.3877 --0.0041 -0.0341 -0.0058 6.7270 5.3996 0.1065 0.0981 0.0649 2107209 2974994 0.1009 0.0531 0.0617 --93.4 -93.4 ----93.6 3.4084 2.8697 0.1011 0.0849 0.0622 2337499 3307376 0.0941 0.0466 0.0579 --91.6 -93.0 ----91.4 1.2472 2416672 -1.4218 3411796 -0.1017 0.0961 93.0 0.0893 0.0490 -0.0622 0.0588 93.6 Note: as all study subjects are susceptible, the true value for πΎ is 1 − exp(π½1 ) = 0.3935. ----93.0 Frailty Gamma π0 π1 π½1 ψ πΎ True value ---0.5 0.5 0.3935 LogNormal π0 π1 π½1 ψ πΎ ---0.5 0.5 0.3935 34.2354 -0.1412 -0.4887 0.3358 0.3835 --0.0113 -0.1642 -0.0100 ---0.5 0.5 0.3935 34.7197 -0.0759 -0.4951 0.3710 0.3873 --0.0049 -0.1290 -0.0061 Binary Parameter π0 π1 π½1 ψ πΎ 10 CPN (%) CPLog (%) Online Table 4: Simulation results using 0TAIL method for the scenarios with “all susceptible”, two correlated covariates (π½1 = −0.5, π½2 = −0.25), Gamma (or Log-Normal, Binary) distributed frailty where ψ = 0.5, and sample size of 500. π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 True value ----0.5 -0.25 0.5 0.3935 0.3935 0.3935 Estimate 30.3359 0.4524 -0.0855 -0.4910 -0.2499 0.4714 0.3949 0.3810 0.3879 Bias ---0.0090 0.0001 -0.0286 0.0014 -0.0125 -0.0055 Empirical SD 9.0845 7.5923 2.2753 0.1690 0.0627 0.0939 0.1084 0.1012 0.1016 Average SE 2528714 4339710 1620692 0.1686 0.0632 0.0540 0.1087 0.1258 0.1217 CPN (%) ---94.8 95.4 -94.7 96.4 96.0 CPLog (%) ------95.4 95.8 96.0 Log-Normal π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 ----0.5 -0.25 0.5 0.3935 0.3935 0.3935 31.6197 0.5004 -0.4677 -0.5017 -0.2447 0.3319 0.3931 0.3859 0.3895 31.6197 0.5004 -0.4677 -0.0017 0.0053 -0.1681 -0.0003 -0.0076 -0.0040 7.5290 5.8322 2.3212 0.1558 0.0596 0.0799 0.0997 0.0957 0.0958 2741682 4696193 1758908 0.1560 0.0585 0.0470 0.0983 0.1271 0.1191 ---94.8 94.2 -94.3 95.0 94.7 ------95.0 95.2 95.0 Binary π0 π1 π2 π½1 π½2 ψ πΎ0 πΎ1 πΎ2 ----0.5 -0.25 0.5 0.3935 0.3935 0.3935 32.7035 0.3226 -0.4661 -0.4896 -0.2479 0.3446 0.3838 0.3791 0.3815 32.7035 0.3226 -0.4661 0.0104 0.0021 -0.1554 -0.0097 -0.0143 -0.0120 Frailty Parameter Gamma 6.1156 2918981 --4.8995 5010071 --2.0984 1873854 --0.1661 0.1571 93.8 -0.0629 0.0589 93.0 -0.0775 0.0487 --0.1021 0.1003 92.9 94.5 0.1008 0.1386 93.5 94.7 0.1005 0.1278 93.5 94.7 Note: as all study subjects are susceptible, the true values for πΎ0 , πΎ1 and πΎ2 are 1 − exp(π½1 ) = 0.3935. 11