Advanced statistical methodologies to address inherent study limitations. Author Response to Ayubi and Saeid

To the Editor:

We read with interest the Letter to the Editor by Ayubi and Safiri1 on our recently published article2 exploring the association of blood pressure (BP) components with all‐cause mortality by level of cognitive function.

In our study, we opted to use all‐cause mortality as the sole outcome, in line with previous reports studying the detrimental effect of either high or low BP in the elderly3, 4, 5 and did not further explore cause‐specific mortality because of the low number of events of cognitively impaired vs cognitively nonimpaired individuals in the cohorts, especially among specific quartiles of BP components. Ayubi and Safiri proposed that cause‐specific mortality with insufficient data can be examined using novel statistical approaches1

The BP components, comprising the main exposure variables in our study, impact the risk for atherosclerosis6 renal function7 complications of diabetes mellitus8 and risk of cognitive decline and dementia9 as well as cancer incidence and mortality10 which are the majority of causes of death among the elderly in Western societies. We particularly aimed to investigate whether cognitively impaired individuals could be more vulnerable to extreme BP values, pointing to a need for more rigorous BP control in this group. In this context, the primary focus of the study was to explore public health implications rather than test etiological hypotheses via evaluation of the specific impact of BP components on a broad range of conditions. Therefore, all‐cause mortality seems to be valid and less likely prone to misclassification outcome.

On the other hand, cause‐specific mortality analyses could certainly contribute to attribution of risk by specific diseases and also invoke underlying mechanisms. However, before implementing the statistical methods addressing sparse data bias that were proposed by Ayubi and Safiri11, 12 the following question should be raised: Would such an analysis, given the restricted sample size, fulfill its aim of unravelling the mechanisms of such complex associations?

Regarding the second point raised, using advanced statistical methodologies to address time‐varying and time‐modified confounders can, in principle, improve informativeness of data. In practice, however, collection of accurate data on the lifestyle variables used in our analyses, notably social activity patterns, smoking habits, and alcohol consumption, is rather cumbersome and prone to error of unknown direction regarding the magnitude of the indicated exposures. Of note, analyses with fully adjusted estimates (models 1 and 2, Table 2 in our study) showed similar results as those adjusted only for age and sex, indicating that potential bias from inclusion of lifestyle time‐varying factors could only minimally affect the findings.

The welcome experimentation with theoretical models in addressing research methodology and specific bias in epidemiology should fulfill appropriate prerequisites and take into account the research question of the study as well as inherent limitations of the data before it is implemented in real datasets.