The validity of self-reported behaviors: methods for estimating underreporting of risk behaviors

Abstract

Purpose: When individuals underreport risk behaviors, data gathered from public health research and practice will underestimate risk. To date, there is little guidance on if or how reports can be adjusted to better reflect true levels of a risk behavior in a given cohort, sample or, by extension, population.

Methods: We develop the underreporting correction factor (UCF), which can be used to correct estimates of the prevalence of a risk behavior using self-report of the behavior and a specific (but not necessarily sensitive) biomarker. The UCF rests on three assumptions: (1) there is no overreporting of the behavior, (2) the biomarker can only be acquired if the person engages in the behavior, and (3) the presence of the biomarker does not affect reporting of the behavior. We investigate the sensitivity of the UCF to violation of these assumptions and develop confidence intervals for the UCF and the corrected prevalence of the behavior.

Results: The UCF is most sensitive to the second assumption (biomarker perfectly specific). We apply the UCF to estimates of sexual risk behaviors in various settings using a variety of biomarkers.

Conclusions: Implementation of the UCF corrects for underreporting and more accurately quantifies risk in cohorts.

Keywords: HIV; Risk behaviors; Self-reported behaviors; Sexual behaviors; Sexually transmitted infections.

FIGURE A.1

Illustration of a lag between time of the risk behavior and the time of a positive test on the biomarker. Assume the behavior and the biomarker are measured at time t. The participant's report at time t reflects behavior over the intervals h and r(i.e. T_t = T_h ∪ T_r). However, if there is a lag of length l, the value of the biomarker at time t (B_t) reflects the behavior over the intervals l and r(i.e. B_t = B_r ∪ B_l). Note that the length of l and h are equal.

FIGURE 1

Effect of assumptions 1 – 3 on bias of UCF where Bias = UCF − P(T = 1|R = 0). A. Assumption 1. Effect of overreporting of the risk behavior on bias in the UCF. B. Assumption 2. Effect of sensitivity and specificity of the biomarker on bias in the UCF. Bias is 0, regardless of sensitivity, when specificity is 1. C. Assumption 3. Bias in the UCF when the presence of the biomarker increases the likelihood of reporting the behavior. Φ represents the ratio of the probability of underreporting among those with and without the biomarker (see text). Assumption 3 requires that Φ = 1. Violations of this assumption are expressed in terms of Φ from 0 to 1.