Murray Finkelstein is appropriately concerned about including postevent outcomes in a regression model. This would result in biased associations, namely, the inability to determine if the predictive factor resulted in the event or if the event resulted in the predictive factor. This has been called “survivor-treatment selection bias”1 or, more generically, “time-dependent bias” and is relatively common even in highly cited medical journals. In a recent systematic review,2 we found that 18.6% (95% confidence interval [CI] 15.8%–21.8%) of studies with a survival analysis contained a time-dependent factor and that 40.9% (95% CI 32.3%–50.0%) of these studies were susceptible to time- dependent bias.
However, we strongly disagree that our Cox model was performed incorrectly, since it was corrected for this bias. As stated in the Methods section, we adjusted for the appropriate time-dependent variables and did have a variable expressing time spent in hospital up to that time.3 Our Results section summarizes the findings. The phrase “within 1 year after discharge” used there refers to the censoring time that we used for all analyses in the study. We did not use any “future information” and our methodology was robust.
In the Methods section, we note that we performed a sensitivity analysis using all outpatient visits rather than cardiovascular visits to define our groups; however, the results of this analysis were omitted by the journal because of space limitations. Using the same variables as in Table 3 but with all visits rather than cardiovascular visits, we found similar results: compared with those who had no outpatient visits, patients seen by a family physician (odds ratio [OR] 0.80, 95% CI 0.64–0.96) or a specialist and family physician (OR 0.48, 95% CI 0.40–0.58) had lower mortality rates. Furthermore, similar results were obtained with the Cox model when all visits instead of cardiovascular visits were used: seeing a specialist was associated with lower mortality (hazard ratio 0.95, 95% CI 0.94–0.96).
References
- 1.
- 2.
- 3.