Incidence Estimation in Studies with Only One Follow-up Visit and an Imperfect Test

By Victor G. Sal y Rosas, Department of Biostatistics

Incidence, the number of new cases of a disease found during a specified time period and in a specified group of people, is one of the most important measures in epidemiologic studies of disease. Traditionally, incidence is estimated using data from a cohort of patients who are followed over time; repeated testing at multiple follow-up visits may be involved. These studies are often very expensive and difficult to implement; as a result, in recent years scientists have looked at whether one can use a cohort study with just one follow-up visit, along with a measure of the disease status at that follow-up visit, to estimate the incidence of the disease of interest (data from such a study is often referred to as “current status data”).

An additional complication, particularly in infectious disease research, is that the test used for determining the disease status at the follow-up visit may not be 100% accurate. This limitation raises the possibility that false positive or false negative results will be given to patients. For example, current screening algorithms for tuberculosis in HIV positive individuals have sensitivity (the probability that the test results are positive among those who actually have TB) around 53% and specificity (the probability that the test results are negative among those who really don’t have TB) around 89%. Similarly, current laboratory tests for gonorrhea and chlamydia detection are about 90% sensitive.

My doctoral dissertation work, under the supervision of professor James P. Hughes, develops methodology to adjust for the additional variation introduced by an imperfect test when analyzing current status data (in our initial work, we assumed that the sensitivity and specificity are known from external sources). We are able to estimate the true incidence rate and the impact of factors that may affect the incidence rate (e.g., age, gender) after correcting for potential misclassification errors from the imperfect test. Ignoring the misclassification would induce bias that we have quantified with simulation results. We have also developed methodology for analyzing data where the sensitivity and specificity vary among individuals or groups of individuals (for example, when different tests are used at different study sites). Our approach tries to make as few assumptions as possible, which is important because having only one follow-up makes evaluating any assumptions quite difficult.

Based on our findings, we intend to extend our work in different directions: a) incorporating additional information about the disease status, such as symptoms, beyond those provided by the laboratory test; b) considering more complex scenarios, in which patients can be observed at more than one follow-up visit but the infection is only known to have been present in the time between two visits; and c) estimating the sensitivity and specificity of the test as part of the scientific question of interest.