Estimation of population attributable fraction (PAF) for disease occurrence in a cohort study design

Abstract

The population attributable fraction (PAF) is a useful measure for describing the expected change in an outcome if its risk factors are modified. Cohort studies allow researchers to assess the predictive value of the risk factor modification on the incidence of the outcome during a certain follow-up. Estimation of PAF for both mortality and morbidity in cohort studies with censored survival data has been developed in the recent years. So far, however, censoring due to death in the estimation of PAF for morbidity has been ignored, resulting in estimation of a quantity which is not relevant in practice as some people are likely to die during the follow-up. The risk factors related to the disease incidence may also be related to mortality, and modification of these risk factors is likely to delay the occurrence of both events. Thus, censoring due to death and the impact of risk factor modification must be considered when estimating PAF for disease incidence. We consider both and introduce two measures of disease burden: PAF for the incidence of disease during lifetime and PAF for the prevalence of disease in the population at a certain time. We demonstrate how consideration of censoring due to death changes the estimated PAF for disease incidence and its confidence interval. This underlines the importance of choosing a correct PAF measure depending on the outcome of interest and the risk factors of interest to obtain accurate and interpretable results.

Barriers to Antiretroviral Therapy Adherence and Plasma HIV RNA Suppression Among AIDS Clinical Trials Group Study Participants

“…40,41 To ensure replication of our results and to help interpret the dominance analysis, we calculated the population attributable risk (PAR) of having detectable plasma HIV RNA for each barrier. 42 In a post-hoc analysis, we examined the relative importance of the 14 barriers at 12 weeks using dominance analysis for those who had been ART-naïve at baseline. We used Stata version 13.1 (StatCorp LP, College Station, TX) for our analyses.…”

SASMacros for Calculation of Population Attributable Fraction in a Cohort Study Design

“…The SAS macros based on the recently developed methods (Laaksonen et al 2010a,b) presented in this paper close the gap between theory and application. Using these macros it is now possible to estimate PAF and its confidence interval in a cohort study design both for total mortality (Laaksonen et al 2010b) and for disease incidence (Laaksonen et al 2010a), taking into account the different sources of censoring. The PAF macros are very flexible in that both categorical and continuous risk factors and confounding factors as well as their interactions can be included in the model, as long as the estimation algorithm still converges.…”

“…If, however, the outcome of interest is incidence of disease, PAF is the proportion of disease cases that could hypothetically be avoided during a time interval (0, t] if its risk factors were modified. In this case, mortality before contracting the disease of interest causes selection in the population during follow-up (Laaksonen et al 2010a). If the risk factors that are related to the incidence of the disease of interest are also related to mortality, the modification of these risk factors is likely to affect both the risk of the disease and the risk of death.…”

“…Therefore, censoring due to death needs to be taken into account in the definition of PAF for disease incidence. The importance of considering censoring due to death in the estimation of PAF for disease incidence has been demonstrated elsewhere (Laaksonen et al 2010a). Let T D denote the time of the occurrence of the disease.…”

“…The estimation of PAF for disease incidence is described in detail elsewhere (Laaksonen et al 2010a) and is only briefly summarized here.…”