Throwing out the baby with the bathwater?: Comparing 2 approaches to implausible values of change in body size

In childhood obesity research, the appearance of height loss, or "shrinkage," indicates measurement error. It is unclear whether a common response--excluding "shrinkers" from analysis--reduces bias.

Using data from the National Longitudinal Study of Adolescent Health, we sampled 816 female adolescents (≥17 years) who had attained adult height by 1996 and for whom adult height was consistently measured in 2001 and 2008 ("gold-standard" height). We estimated adolescent obesity prevalence and the association of maternal education with adolescent obesity under 3 conditions: excluding shrinkers (for whom gold-standard height was less than recorded height in 1996), retaining shrinkers, and retaining shrinkers but substituting their gold-standard height.

When we estimated obesity prevalence, excluding shrinkers decreased precision without improving validity. When we regressed obesity on maternal education, excluding shrinkers produced less valid and less precise estimates.

In some circumstances, ignoring shrinkage is a better strategy than excluding shrinkers.