[Skip to Navigation]
[Skip to Navigation Landing]
Views 720
Citations 0
Comment & Response
September 2015

Anticholinergic Use With Incident Dementia—Reply

Author Affiliations
  • 1School of Pharmacy, University of Washington, Seattle
  • 2Group Health Research Institute, Seattle, Washington
  • 3Department of Biostatistics and Epidemiology, Perelman School of Medicine, University of Pennsylvania, Philadelphia
JAMA Intern Med. 2015;175(9):1577. doi:10.1001/jamainternmed.2015.2574

In Reply Residual confounding refers to any bias arising from failure to adequately adjust for confounders: either unobserved covariates that are not included in the analysis or observed covariates that are inadequately included in the analysis due to measurement error or model misspecification. In our study, we controlled for confounding through regression adjustment for covariates associated with dementia that differed by exposure status.1,2 All covariates noted by Fried as differing between individuals with varying anticholinergic exposure levels were included in regression models to account for potential confounding due to these observed covariates. In addition to the results reported in our article, we performed several prespecified sensitivity analyses to further investigate this issue and assess the robustness of our primary results. Analyses that additionally adjusted for the Charlson comorbidity index produced similar results to our primary analyses. The hazard ratios for participants in the highest anticholinergic category (>1095 total standardized daily doses) were 1.56 for dementia (95% CI, 1.22-1.99) and 1.68 for Alzheimer disease (95% CI, 1.28-2.20). As discussed in the limitations section of the article,2 we share the concern expressed by Fried that residual confounding due to unobserved covariates may persist despite adjustment for observed confounders. The potential for bias due to unobserved confounding exists in all observational studies. However, we do not agree with the strategy that Fried proposes to address this issue. Restricting the sample to those without comorbidities at study entry (ie, potential confounders) would lead to a highly select group of healthy older adults (56.5% of the sample would be dropped in our study) and would limit the generalizability of the results. Both stratification and regression adjustment account for confounding by conditioning on observed covariates that may be associated with the outcome and the exposure, but neither address confounding due to unobserved covariates.1 Given the limitations to generalizability posed by a stratified analysis, we believe our analysis using regression adjustment provides a more scientifically valid approach.