Statistics in Medicine. 2007 Jul 30;26(17):3229-39.
Natarajan S, Lipsitz SR, Rimm E.
VA New York Harbor Healthcare System, New York, NY 10010, USA. sundar.natarajan@med.nyu.edu
Significance:
Population-based health surveys (complex surveys) allow population estimates of measures of frequency (incidence and prevalence) and measures of association (odds ratio, relative risk, hazard ratio and attributable risk). In contrast to the other measures of association, population attributable risk (PAR) combines information on prevalence and a measure of association to provide a quantitative estimate of the proportion of disease in the population that is directly attributable to a particular risk factor. Because of the complex sampling frame utilized in these surveys, design-based analyses incorporate the weighting, stratification and clustering. Though all the measures of association can be estimated from such surveys, odds ratios, relative risks and hazard ratios are most commonly computed because their point estimates and associated confidence intervals (CI) can be easily calculated. PAR estimates the public health impact of a particular exposure and is used to estimate the proportion of disease that can be prevented if that exposure were eliminated. PARs are thus important in judging public health priorities and in linking causality with public health policy.
Despite its value, PARs are infrequently used in research, partly due to confusion regarding its definition and partly due to the difficulty in computing them. Because of this, only a small percentage of articles report PAR, and, of those that report PAR, very few report CIs. While point estimates of PAR are valuable and provide an indication of central tendency, they do not provide insight regarding the precision of the estimates by quantifying uncertainty. This serious limitation restricts its applicability. An application of the method in a general medical publication [Mann DM; Lee J; Liao Y; Natarajan S. "Independent effect and population impact of obesity on fatal coronary heart disease in adults" Preventive medicine, 2006 Jan;42(1):66-72] is shown in figure.
This current paper reviews the theory, provides a straightforward method to obtain CIs for PAR and illustrates the process using simple random and complex survey examples. Based on this uncomplicated and valid method of determining CIs for PAR, future researchers analyzing complex survey data can easily provide a population perspective on their results by estimating the PAR and quantify the uncertainty associated with the PAR estimate by computing the CIs leading to better statistical inferences.
Abstract:
Methods to assess uncertainty in the estimated population attributable risk (PAR) by calculating 95 per cent confidence intervals (CIs) are not readily available in software for complex sample surveys. Using the Bonferroni inequality, a simple method to obtain CIs for the PAR is developed.
The method is demonstrated using a simulation in a (2 x 2) table as well as a cohort study to calculate CIs for PAR of coronary heart disease death (using proportional hazards regression).
This article demonstrates a straightforward, theoretically valid method of determining CIs for the PAR. Using this method, researchers analyzing complex surveys can routinely provide a population perspective and a valid measure of the uncertainty for these estimates.
PMID: 17309113
