RSS Feeds
-
Response to Thombs and Ziegelstein
Submit responseWe appreciate Dr. Thombs’ and Dr. Ziegelstein’s observation that our study was of great importance in reporting the beneficial effect of a depression intervention on mortality. Thombs and Ziegelstein were concerned about “pre-screening” of covariates, leading to over-fitting and ignoring the issue of confounding, resulting from post hoc selection of covariates for inclusion only if they were associated with the outcome. We had a pre-specified approach to identifying and including potential confounders because we knew that imbalances would be likely and adjustment with patient-level variables would be necessary given the practice- randomized design. Our pre-specified approach did address the concern about confounding by identifying potential confounders for inclusion in the model by their association (p < 0.10) with the interaction variables of interest, randomization assignment and baseline depression status, as well as the dependent variable, time to death. Using this approach, only age, level of educational attainment, baseline smoking status, history of myocardial infarction reported at baseline, and baseline suicidal ideation were identified as potential confounders. The intent-to-treat hazard ratio and corresponding 95% confidence interval for patients with major depression was consistent with the reported result (adjusted hazard ratio was 0.62 with 95% confidence interval [0.42, 0.92]). Additional variables for which we adjusted the point estimates reported in Table 4 were requested by reviewers. We want to emphasize the pre-specified nature of our statistical approach and the care with which we selected variables for inclusion in models. The “surprising and unexpected finding” related to a reduction in cancer deaths was unadjusted and therefore was not influenced by the selection of covariates in multivariate models. We stated “any evidence of a potential association of practice intervention assignment and specific causes of death must be viewed as an opportunity for generation of hypotheses to be tested in future intervention research.” We did not call for research on mechanisms related to the decreased mortality risk from cancer. On the other hand, we would not want to be dismissive of the findings with regard to cancer deaths. We did suggest that mediators of the effect of a depression intervention on mortality do deserve further study to increase our understanding of how depression leads to increased mortality. We believe this is the first publication of a randomized clinical trial to report decreased mortality in association with treatment of depression; replication would be welcome.
Conflict of Interest:
None declared
-
The Effect of a Primary Care Practice-Based Depression Intervention on Mortality in Older Adults: Re
Submit responseRecently, Gallo et al. (1) reported that a depression care management intervention significantly reduced risk of 5-year mortality among older primary care patients with major depression compared to patients with usual care. There were no deaths from suicide among patients with major depression in either group. The results from this study are of potentially great importance. Despite many studies that report prospective relationships between depression and important outcomes like mortality, there is generally little evidence that depression treatment reduces overall mortality rates.
The statistical methods used by Gallo et al. for covariate adjustment, however, are known to result in model overfitting, which raises the question of whether these findings would generalize to other similar patient samples. On an unadjusted basis, patients in the intervention practices with major depression were not at lower risk of mortality. They were at significantly greater risk only after adjusting for 10 “influential covariates” that Gallo et al. identified based on significant univariate associations with time to death. Methods like this, however, that prescreen variables for subsequent entry into multivariate regression analyses are indirect versions of automated variable selection procedures (e.g., stepwise regression) (2). The Statistical Guidelines published online by the Annals of Internal Medicine counsel against prescreening variables and state, “Authors should avoid stepwise methods of model building, except for the narrow application of hypothesis generation for subsequent studies.” It has been amply demonstrated that prescreening and other automated variable selection methods capitalize on variability unique to a given sample, radically underestimate the degrees of freedom used to determine estimates in regression models, often generate substantially inflated Type I error rates and artifactually small p values, and don’t consistently produce replicable findings (3).
In the study by Gallo et al., the combined effect of adding the group of 10 preselected “influential covariates” was to substantially elevate, and possibly exaggerate, the hazard ratio associated with the intervention for patients with major depression. It also produced the surprising and unexpected finding that these results were largely due to a reduction in deaths related to cancer (15 in usual care practices versus 8 in treatment practices). Gallo et al. concluded that further investigation is needed to clarify the mechanisms behind the relationship between the depression intervention and decreased mortality risk from cancer. Given the limitations of their analytical methods, however, investigation of causal mechanisms is not warranted until the basic findings of the study are reproduced.
Conflict of Interest:
None declared