A Method to Estimate Off-Schedule Observations in a Longitudinal Study

• Published in: Annals of epidemiology Vol. 21; no. 4; pp. 297 - 303
• Main Authors: Reeves, Katherine W., Stone, Roslyn A., Modugno, Francesmary, Ness, Roberta B., Vogel, Victor G., Weissfeld, Joel L., Habel, Laurel A., Vuga, Marike, Cauley, Jane A.
• Format: Journal Article
• Language: English
• Published: United States Elsevier Inc 01.04.2011
• Subjects: Body Mass Index; Data Collection - methods; Data Collection - statistics & numerical data; Data Interpretation, Statistical; Densitometry - methods; Densitometry - statistics & numerical data; Epidemiologic Methods; Female; Humans; Internal Medicine; Linear Interpolation; Linear Models; Longitudinal Studies - methods; Longitudinal Studies - statistics & numerical data; Mammography; Mammography - methods; Mammography - statistics & numerical data; Middle Aged; Missing Data; Multiple Imputation; Retrospective Studies; Multiple Imputation; Linear Interpolation; Missing Data; Epidemiologic Methods; SWAN; LLI; Mammography; BMI; body mass index; Study of Women’s Health Across the Nation; local linear interpolation
• ISSN: 1047-2797; 1873-2585; 1873-2585
• DOI: 10.1016/j.annepidem.2010.11.013

Data in epidemiological studies sometimes are collected off-schedule from planned study visits. In an ancillary study to the Study of Women’s Health Across the Nation (SWAN), longitudinal breast density data were collected retrospectively from mammograms that were not acquired at the study visits. We propose a method to estimate the off-schedule breast density measurements at the time of study visits. This method uses local linear interpolation, with multiply imputed error terms drawn from assumed subject-specific normal distributions based on the within-subject standard deviations of mammographic density measurements. We evaluate the validity and implications of this approach. Coefficients of random intercept models used to assess the association between annual changes in body mass index and dense breast area estimated with this approach (β = −0.17, p = .46) differed from those obtained when each mammogram was matched to the nearest study visit (β = −0.30, p = .04). The proposed estimation approach had a small average prediction error (0.11 cm 2). Because matching does not incorporate breast density changes over time, our local linear interpolation with multiple imputation approach may provide more accurate results. The proposed approach is applicable to other epidemiologic studies with off-schedule data in which the missing variable changes linearly over relatively short periods of time.