Estimating the Causal Effect of a Latent Class Treatment On Binary and Count Outcomes

Introduction. The Surgeon General recently reinforced the importance of preventing youth tobacco use with a report detailing the scope, negative consequences, and risk factors for use (US DHHS, 2012). Smoking behavior is multifaceted and understanding how to prevent it requires consideration of individual behavior in concert with availability and peer influences. In particular, latent class analysis can be used to model multiple aspects of smoking to identify subgroups of individuals with different behavior patterns. Whether these patterns then cause later outcomes is a difficult methodological problem. This talk will (1) introduce a new statistical approach to estimating the causal effect of latent class membership on a later outcome; (2) describe an extension of this approach to binary/count outcomes; (3) demonstrate the technique by estimating the causal effects of adolescent smoking latent class membership on the distal outcome of adulthood smoking.

Method. The sample comprised 720 adolescents from the National Longitudinal Study of Adolescent Health (51.7% female). Potential confounders (e.g., individual/peer characteristics) were assessed at 11^th grade (Wave I). Indicators of the latent class model included past-year any and regular cigarette use, past-month any and regular cigarette use, past-month 3+ cigarettes per day, tried to quit in past 6 months, 1+ friends smoke regularly, and cigarettes easily available in the home, assessed at 12^th grade (Wave II). Outcomes of past-month regular cigarette use (binary) and past-month number of days smoked (count) were assessed during adulthood (Wave IV). We used the new latent class causal analysis method to estimate the causal effect of adolescent smoking latent class membership on adulthood smoking.

Results. The four class model fit the data best; classes included non-smokers (56% of sample), occasional light smokers (12%), quit attempters (12%), and regular heavy smokers (19%). Prior to adjusting for confounding, adolescent smoking patterns were related to both measures of adulthood smoking; importantly, the majority of adolescent regular heavy smokers were regular smokers in adulthood. To estimate the causal effect, we included the potential confounders in both the model for predicting selection into the latent classes and the model for imputing the potential outcomes. Results from this new approach to causal effects of latent class treatments using the R package LCCA will be discussed.

Conclusion. We believe this new approach to causal effects of latent class treatments will be valuable to prevention scientists, and we will demonstrate how to estimate causal effects on binary and count outcomes. Implications for prevention based on causal effects of early smoking behavior patterns will be discussed.