Estimating exposure in cross sectional research: Measuring the impact of telescoping bias

Background: A great deal of public health policy relies on cross sectional measures of prevalence. When these measures are defined within a specific time frame, such as past 6-month experience, they become subject to a potential telescoping bias resulting in an overestimation of exposure. While the impact of telescoping bias has been discussed, little research exists measuring this potential bias directly. Methods: Utilizing a linked National Crime Victimization Survey (NCVS) dataset which surveys households for three and a half years, unbounded interviews included the baseline household interview and interviews with individuals who missed one or more prior bounding interview (due to a move or a missed interview). Logistic regression solved with GEE (Generalized Estimating Equations) was applied to fit the data from bounded interviews. To estimate the "expected violence proportion" based on bounded interviews, we extrapolated the model six months prior to the first bounded interview. These estimates were compared with observed unbounded estimates. Results: Overall, an overestimation of reported violence exposure occurred in the unbounded interview (expected proportion 0.19; observed proportion: 0.26). This telescoping violence was greater among younger respondents who report more victimization than older respondents. Telescoping bias was also most pronounced when an individual recently moved or missed the prior household interview. Discussion: While understanding current risk through the use of cross sectional data is efficient and vital in public health, particularly for making policy decisions about prevention, telescoping bias must be considered when reporting prevalence estimates.