Change-score and Oldham's models gave highly consistent results across five cardiometabolic markers, whereas baseline-adjusted change-score models yielded different results for all; thus, when exposure precedes the baseline outcome, baseline adjustment in change-score analyses may substantially alter effect estimates.
Change-score and Oldham's models produced highly consistent estimates across all five cardiometabolic markers.
Baseline-adjusted change-score models yielded different results from the unadjusted change-score model for all five markers.
Statistical tests of whether change in the outcome depends on baseline level did not indicate whether baseline adjustment would materially change results.
Based on change-score models, former drinkers had lower 5-year TG change and higher SBP, DBP, and HDL-C changes than light drinkers.
Heavy drinkers had lower DBP change than light drinkers in change-score models.
The study used a large population-based cohort to compare modeling strategies.
This study asked how the choice of statistical method affects results when researchers look at how an exposure measured at the start of a study relates to later changes in a health measure. Using data from the German National Cohort (about 205,000 people at the start and 140,000 at follow-up roughly 5 years later), the authors looked at alcohol drinking habits and changes in triglycerides, blood pressure, HDL cholesterol, and long-term blood sugar (HbA1c). Two approaches, a simple change-score model and a model that adjusts for the average of baseline and follow-up values (Oldham's), gave very similar answers. A third approach, which also adjusts for the starting value of the outcome, gave different answers for all five markers. Statistical tests that check whether change depends on starting level did not help in deciding whether adjusting for baseline would matter. In terms of alcohol findings, former drinkers had lower triglyceride change and higher blood pressure and HDL cholesterol changes than light drinkers, while heavy drinkers had lower diastolic blood pressure change. This research suggests that when the exposure (like drinking) came before the baseline measurement of the outcome, adjusting for baseline can substantially change conclusions, so the choice of analysis method deserves careful thought.
Bouka M, Nimptsch K, Grittner U, Gonzalez Chiappe S, Maskarinec G, Peters A, et al.. (2026). Comparison of modeling strategies to study the effect of an exposure on changes in continuous outcomes: evaluating the impact of alcohol consumption on changes in cardiometabolic health markers based on the German National Cohort (NAKO).. International journal of epidemiology. https://doi.org/10.1093/ije/dyag245