Table 2.
Approaches to reducing residual confounding by unmeasured factors
| Crossover studies (e.g., case-crossover design) |
External adjustment (e.g., survey information with clinical details in a subsample) |
Proxy measures (e.g., high-dimensional propensity scores) |
IV methods (e.g., two-stage regression) |
|
|---|---|---|---|---|
| Approach | Different time periods within the same patient serve as control time periods |
Additional information on clinical risk factors will be collected on a subsample of patients and used to adjust finding in main study |
Many measured covariates and their interactions adjusted by propensity score methods may be proxies for unmeasured confounders |
A correlate of the study exposure not related to measured and unmeasured confounders serves as an un-confounded substitute for the drug exposure |
| Advantage of approach | Adjusts all measured and unmeasured time-invariant confounders, including genetic predispositions |
The study investigator will define the detail an quantity of additional clinical information that will be gathered in a subsample |
Well-established propensity score methods can be applied to improve efficiency when adjusting hundreds of covariates |
Provides consistent effect estimates even in the presence of unmeasured confounders |
| Assumptions that need to be made |
No within-person confounding over time. Case–time control studies can help reduce within- person confounding if time trends in controls are representative of time trends in cases |
All relevant confounders must be defined; subsample must be representative |
A high-dimensional matrix of measured covariates also represents unmeasured confounders |
No association between instrument and confounders; no direct effect of the instrument on the study outcome other than through the actual drug exposure |
| Testability of assumptions | Time trends of measured confounders can be examined and extrapolated to unmeasured confounders |
Completeness of the list of confounders observed in the subsample must be argued; representativeness can be tested |
The degree of representation of unmeasured confounders is not knowable from the data. Effect measure modification by propensity score may suggest residual confounding |
Validity of assumptions must be argued. Improvement in the balance of measured covariates between treatment groups can be demonstrated and extrapolated to unmeasured confounders |
IV, instrumental variable.