Nicotine and Tobacco Research 2019 Discussion Paper
Analyses Baseline measures were summarized using descriptive statistics, and between-group differences in baseline characteristics were tested.30 Potential differences between cohorts in baseline characteristics were explored using parametric (analysis of variance) and nonparametric (chi-square) tests as appropriate.
Using a series of longitudinal models implemented with gener- alized estimating equations with robust standard errors, we tested the effect of randomization on binary smoking outcomes (7PPA and continuous abstinence) over time (smoking status at week 5 through follow-up) controlling for potential confounders including contam- ination risk (defined as any Wellness participant who reported par- ticipation in Yoga during the 8-week treatment).
Models included effects of intervention, time, time × intervention, as well as potential confounders. Data were clustered within participant within cohort, and standard errors were adjusted accordingly. We explored poten- tial moderating effects of smoking rate at baseline by including main effects of smoking rate, and all two- and three-way interactions between smoking rate, time, and treatment group.
We tested the effects of randomization on mean smoking rate over time using a longitudinal mixed effects model in which smok- ing rates postquit date (week 5) through follow-up were regressed
on time, group, and time × group. Models adjusted for confound- ers including contamination risk and baseline smoking rate, and included a subject-specific intercept to adjust for repeated measures of the outcome within participant. Potential cohort effects were also explored. As with our primary outcomes, we explored potential moderating effects of baseline smoking rate on treatment effects.
Using Latent Class Models (LCMs), we sought to identify smok- ing patterns during the treatment period (exploratory analysis). The outcome of interest was self-reported 7PPA from quit date through EOT, with cotinine validation at week 8. LCMs assume the popula- tion is made up of a finite number of patterns (smoking patterns in this case).
This technique reduces participant-level data from a vec- tor of up to 8 weeks of data to a single class, corresponding to their most likely pattern of smoking behavior. Class (pattern) provides an objective grouping that can be used as a predictor or outcome in subsequent analyses. To identify the number of classes supported by the data, we fit a series of LCMs ranging from 2 to 6 classes and identified the model that minimized the Bayesian information criteria value (which maximizes fit). The optimal solution was a 4-class model (significant model fit and significantly lowest Bayesian information criteria), which is presented later. Classes were com- pared based on baseline characteristics, randomized group, smoking
Figure 1. Consort diagram.
Nicotine & Tobacco Research, 2019, Vol. 21, No. 11 1519
rates at 3- and 6-month follow-up, and baseline motivational vari- ables (ie, motivation, readiness, and confidence) using chi-square tests and analysis of variance as appropriate. Finally, we explored potential dose effects both within and across groups using a similar modeling strategy.
All analyses were conducted on the intent-to-treat sample (N = 227). Models used likelihood-based approaches to estimation and thus made use of all available data without directly imputing missing data. Results were compared to the conservative assumption that missing equals smoking (in the case of binary outcomes) and did not differ substan- tially from the maximum likelihood estimation. All analyses were run using SAS v. 9.3 and R, and significance value was set at α = .05 a priori.
Sample Among participants (N = 227) randomized at baseline, 55.5% were women, 72.3% had attended at least some college, 42.3% were mar- ried or partnered, and 55.9% were employed full time. Mean age was 46.2 (SD = 12.0) years. Mean smoking rate at baseline was 17.0 (SD = 7.8) cigarettes/day. There were no significant between-group differences in baseline characteristics (Table 1), and no differences between study cohorts. Overall, the study retention rate was 94.7% through final follow-up with no difference between groups (p > .05).
Smoking Outcomes Longitudinal adjusted models indicate significant group effects favoring Yoga with respect to 7PPA at EOT (odds ratio [OR] = 1.37, 95% confidence interval [CI] = 1.07% to 2.79%). The odds of 7PPA at EOT were 37% higher for Yoga versus Wellness participants. Effects were no longer significant at 3- and 6- month follow-up (ps > .05). There were no significant effects of group on continuous abstin- ence (p = .52). Overall, 11.2% of participants had prolonged abstin- ence at 6 months, with no significant group effect (p = .92).
Moderating Effects Exploration of the moderating effects of baseline smoking rate sug- gests there were significant effects of group on 7PPA at EOT among those with low smoking rates at baseline (≤10 cigarettes/day = 25th percentile; OR = 2.43, 95% CI = 1.09% to 6.30%). Among light smokers at baseline, the odds of 7PPA at EOT for Yoga were 2.43 times that of Wellness. There were no moderating effects of baseline smoking rate on continuous abstinence (p > .05).
Adjusted models indicate that Yoga participants were smoking significantly fewer cigarettes per day at EOT compared to Wellness (adjusting for baseline). Specifically, there was a 1.54 (standard error = 0.59, p = .01) difference in cigarettes per day favoring Yoga at EOT. This effect was most pronounced among those with higher smoking rates at baseline (≥20 cigarettes/day = 75th percentile). Among these individuals, those in Yoga smoked 2.66 fewer ciga- rettes per day at EOT compared to Wellness (standard error = 1.33, p = .04).
Patterns of Quitting Behavior LCM analysis suggests a 4-class model is best supported by the data. Patterns show that 16% of participants quit by week 4 and had high probability of remaining quit through week 8 (Class 1: Quit Date Quitters). Most participants (71%) were not able to achieve abstinence (Class 2: Non-quitters), 5% of participants were slow and steady quitters such that the slope increased over time with
high probability of being quit by EOT (Class 3: Slow Quitters), and 8% of participants had high probability of quitting on or before week 4 with a decline in odds of remaining quit thereafter (Class 4: Relapsers). These patterns are depicted in Figure 2.
Participants randomized to Yoga were significantly more likely to be Quit Date Quitters or Slow Quitters compared to Wellness partic- ipants (p = .04). Both Quit Date Quitters and Slow Quitters (67.6% and 60%, respectively) were more likely to report 7PPA at 3 months than Non-quitters and Relapsers (2.2% and 14.3%, respectively). Although there were no significant between-class differences in base- line demographics (Table 2), there was a significant effect of baseline smoking rate on the distribution of classes. Specifically, light smok- ers (≤10 cigarettes/day) were significantly more likely to be Slow Quitters compared to other classes (p = .04).
There were significant between-class differences in readiness and confidence in quitting (and a trend for motivation to quit) at
Table 1. Participant Characteristics at Baseline by Intervention Group (N = 227)
Variable Overall, N = 227
Yoga, n = 113
Wellness, n = 114
Age, mean (SD), y 46.2 (12.0) 46.1 (12.0) 46.4 (12.0) Gender (female), No. (%) 126 (55.5) 67 (59.3) 59 (51.8) Race (white), No. (%) 195 (85.9) 102 (90.3) 93 (81.6) Hispanic/Latino, No. (%) 8 (3.5) 4 (3.5) 4 (3.5) Education level, No. (%) High school graduate
or less 63 (27.8) 31 (27.4) 32 (28.1)