Measurement invariance is desirable in population studies because non-invariance may cause statistical bias, which could lead researchers to make misguided recommendations. A new study examined patterns of alcohol use from adolescence to adulthood across key demographic variables. This study analyzed data from three waves of the National Longitudinal Study of Adolescent to Adult Health, which used a three-item measure to assess alcohol use, to determine whether this measure demonstrated invariance across racial/ethnic, sexual identity, and education groups. Three types of measurement invariance (configural, metric, and scalar) were examined. Participants (N = 11,715) were a representative sample of U.S. 7th – 12th graders who completed surveys at baseline, five to six years later (at ages 18-24 years), and 13-14 years (at ages 24-32 years). Within-wave results showed there was metric and scalar invariance in alcohol use measurement across racial/ethnic and sexual identity groups, but not across any of the college education subgroup comparisons. Longitudinally, Black and gay/bisexual males were the only male groups to display metric invariance across all waves, but no male groups displayed scalar invariance across all waves. Similarly, only Black females demonstrated metric invariance across all longitudinal comparisons and no female group displayed scalar invariance across all waves. Non-invariance was greater from adolescence to adulthood (from Wave 1 to 3) compared to across adulthood (Wave 3 to 4). The authors stated the results suggested greater variability in drinking behaviors over time for women than for men. Based on the results of this study, the authors cautioned the multi-item alcohol use measure may introduce biased parameter estimates if models do not account for invariance, especially for comparisons across subgroups.
Take away: Measurement invariance is desirable in population models, especially when subgroups are compared. This study found a three-item alcohol use measure from a nationally representative longitudinal study was non-invariant, which can lead to statistical bias and misguided recommendations.