Editor,—In their recent review of methodological issues in sexual behaviour research, Fenton et al¹ provide a comprehensive overview of the major types of sexual behaviour research, the sources of measurement error which may affect such research, and different approaches to measuring various forms of measurement error. We would like to provide an important footnote on the implications of the poor measurement of sexual behaviours for drawing inferences from studies of sexually transmitted infections (STIs) which attempt to adjust for sexual behaviours in their analyses.

The role of systematic measurement errors in study design and analysis, as described by Fenton et al, is widely recognised. Given their impact on inferences of association, great care is taken in most studies to avoid these biases. The effects of random measurement error, or non-differential misclassification, on epidemiological inference typically receive less attention. Most researchers realise that non-differential misclassification of exposure and/or outcome measures will lead to an attenuation of the resulting measure of association.

However, the fact that random measurement error in potential confounding variables may wreak havoc on the inferences which are made from study results is seldom acknowledged. The non-differential misclassification of a dichotomous confounding variable may lead to inadequate statistical adjustment (often referred to as residual confounding) and the false appearance of statistical interaction when none is present.² When confounders are measured as polytomous or continuous variables (for example, condom use never/sometimes/always or number of sexual partners), random measurement error can bias the adjusted measure of association unpredictably—in some instances making the adjusted measure of association less accurate than the crude.^{3, 4} These forms of misclassification are generally of greatest concern when the true exposure-disease association is relatively weak compared with the exposure-confounder and outcome-confounder relation,⁵ as is the case in most research around STIs. Even small random errors can have major effects on adjusted measures of association, and the unpredictability of the effects of misclassification may be compounded in multivariate analyses.⁶

With this in mind, Fenton et al's review of the difficulties involved in the accurate measurement of sexual behaviour has powerful implications for studies attempting to control for covariates associated with risk for STIs. Studies which attempt to adjust during statistical analysis for numbers and types of sexual partners, frequency of sexual contacts, or condom use practices, are likely to encounter some degree of random measurement error. Although perhaps non-differential with respect to exposure or outcome, this mismeasurement may lead to unpredictable biases and/or mis-specified analyses, and in turn, spurious inferences.

In summary, the inadequate measurement of sexual behaviour requires special consideration in any study attempting to adjust for the confounding role of sexual behaviours in associations involving STIs. We hope that Fenton et al's review of the challenges posed by the collection of sexual behaviour data helps to draw attention to this frequently overlooked methodological aspect of the epidemiology of STIs.