The primary analytic strategy involved a mixed model approach to examine the effects of the dependent variables (e.g., craving, mood) across assessment types. The mixed model approach is ideally suited for analysis of repeated-measures data because it allows for more specific estimation of the correlation structure of the residuals, and more efficiently handles unbalanced designs and missing data, without excluding participants or imputing values (Gibbons, Hedeker, & Waternaux 1988; Gibbons et al., 1993). Because of the the repeated nature of EMA data, the residuals may be both homoscedastic and autocorrelated (Schwartz & Stone 1998). Thus, to allow for heteroscedastic variance over time, a first-order autoregressive error structure was added to the model. Changes in model deviance were compared with a chi-square distribution, between the model with and without an autoregressive error structure, to examine whether adding an autoregressive error structure yielded significant improvement in model fit. An initial analysis was conducted without covariates in the model. Baseline levels of cotinine, FTND, age, cigarettes per day, and sex were then entered into the model for a second analysis. Because there was
