Table 1.
Combinations of Confounder Summary Scores, Confounding Adjustment Methods, Data-Sharing Approaches, and Outcome Types Evaluated in a Study of Privacy-Protecting Analytical Methods
| Confounding Adjustment Method and Data-Sharing Approach | Statistical Analysis Performed at the Analysis Center | |
|---|---|---|
| Binary Outcomea | Time-to-Event Outcomeb | |
| Propensity Score | ||
| Stratification | ||
| Pooled individual-level | PS- and site-stratified (reference analysisc) | PS- and site-stratified (reference analysis) |
| Risk-set | PS- and site-stratified | Case-centered logistic regressionc |
| Summary-tabled | PS- and site-stratified | PS- and site-stratified CPR |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Matching | ||
| Pooled individual-level | PS-matched, site-stratified (reference analysis) | PS-matched, site-stratified (reference analysis) |
| Risk-set | PS-matched, site-stratified | Case-centered logistic regression |
| Summary-table | PS-matched, site-stratified | PS-matched, site-stratified CPR |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Inverse probability weighting | ||
| Pooled individual-level | IPW, site-stratified (reference analysis) | IPW, site-stratified (reference analysis) |
| Risk-set | IPW, site-stratified | IPW, site-stratified |
| Summary-table | Not established | Not established |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Matching weighting | ||
| Pooled individual-level | Matching-weighted, site-stratified (reference analysis) | Matching-weighted, site-stratified (reference analysis) |
| Risk-set | Matching-weighted, site-stratified | Matching-weighted, site-stratified |
| Summary-table | Not established | Not established |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Disease Risk Score | ||
| Stratification | ||
| Pooled individual-level | DRS- and site-stratified (reference analysis) | DRS- and site-stratified (reference analysis) |
| Risk-set | DRS- and site-stratified | Case-centered logistic regression |
| Summary-table | DRS- and site-stratified | DRS- and site-stratified CPR |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Matching | ||
| Pooled individual-level | DRS-matched, site-stratified (reference analysis) | DRS-matched, site-stratified (reference analysis) |
| Risk-set | DRS-matched, site-stratified | Case-centered logistic regression |
| Summary-table | DRS-matched, site-stratified | DRS-matched, site-stratified CPR |
| Effect-estimate | IVW meta-analysis | IVW meta-analysis |
| Inverse probability weighting | ||
| Pooled individual-level | Not established | Not established |
| Risk-set | Not established | Not established |
| Summary-table | Not established | Not established |
| Effect-estimate | Not established | Not established |
| Matching weighting | ||
| Pooled individual-level | Not established | Not established |
| Risk-set | Not established | Not established |
| Summary-table | Not established | Not established |
| Effect-estimate | Not established | Not established |
Abbreviations: CPR, conditional Poisson regression; DRS, disease risk score; IPW, inverse-probability–weighted; IVW, inverse-variance–weighted; PS, propensity score.
a Unless otherwise specified, logistic regression was used to obtain estimates of odds ratios and their 95% confidence intervals for binary outcomes.
b Unless otherwise specified, Cox proportional hazards regression was used to obtain estimates of hazard ratios and their 95% confidence intervals for time-to-event outcomes.
c Case-centered logistic regression is a logistic regression model with the proportion of exposed outcome events among all events used as the dependent variable and the log odds of having the study exposure in the risk set used as the independent variable, specified as an offset (9). Each risk set, anchored by a unique outcome event time, comprises patients who experienced the outcome and patients who were still at risk of developing the outcome at that time point. When combined with confounder summary scores, the risk set is created within a matched cohort or stratum defined by the confounder summary score within a site. In this particular analysis, each risk set comprised the patient or patients who developed the outcome plus all other at-risk patients belonging to the same PS stratum at the time of the event within each site.
d In situations where the regression-based analysis was not feasible for the summary-table data-sharing approach, we used the Mantel-Haenszel method to compute a weighted estimate for the desired effect estimate.