LATEBREAKER
Methods/Statistics
Utilization of locally estimated scatterplot smoothing (LOESS) regression to estimate missing weights in a longitudinal cohort of breast cancer patients Kristina Johnson* Alexa Zimbalist Kelly Radimer Janise Roh Isaac Ergas Charles Quesenberry Marilyn Kwan Lawrence Kushi
Background: Weight data is vital to longitudinal studies that aim to understand how body size may impact breast cancer prognosis. Incomplete data is common in such studies and can lead to biased study results if ignored. Traditional methods to handle missing data rely on making assumptions about missing data patterns. In this study, locally estimated scatterplot smoothing (LOESS) regression models are explored as a data-driven option to minimize missing weight data in a longitudinal cohort of women with breast cancer.
Methods: We identified 10,782 women diagnosed with invasive breast cancer from 2005-2013 at Kaiser Permanente Northern California. Outpatient weights from 2 years pre-diagnosis to 10 years post-diagnosis were included. LOESS regression models were fit to each individual’s weight measurements and subsequently used to estimate weights at baseline (breast cancer diagnosis) and 6 follow-up time points (6, 12, 24, 48, 72, and 96 months post-baseline). The LOESS-estimated weights were compared to a traditional method of selecting the weight measurement closest to each timepoint within a time window of 90 days.
Results: LOESS models identified fewer weights compared to the traditional method at baseline (85% vs 92%). However, LOESS models identified considerably more weights at later follow-up periods, particularly at follow-ups 48, 72, and 96 months post-baseline (48 mths: 85% vs 75%; 72 mths: 78% vs 62%; 96 mths: 69% vs 51%). At baseline and the 6 follow-up timepoints, the weight measurements identified by both methods were highly correlated and more than 80% of the weights differed by 2.50 kilograms or less.
Conclusions: LOESS regression identified more weights at follow-up times further away from breast cancer diagnosis in comparison to the traditional method of weight identification. LOESS regression makes effective use of available longitudinal data and may be a beneficial tool to minimize missing longitudinal data in future studies.