Association of a Comprehensive Smoking Cessation Program With Smoking Abstinence Among Patients With Cancer

Key Points Question Are there differences in abstinence rates between patients with and without cancer after treatment in a comprehensive tobacco cessation program delivered in an oncologic setting? Findings In this cohort study of 3245 smokers in a tobacco treatment program, mean smoking abstinence rates were 45.1% at the 3-month follow-up, 45.8% at the 6-month follow-up, and 43.7% at the 9-month follow-up; rates did not differ between patients with and without cancer. Patients with head and neck cancer were among those with the highest abstinence rates. Meaning When exposed to a comprehensive tobacco treatment program, smokers with and without cancer showed sustained high quit rates and did not differ from each other, suggesting that comprehensive treatment in an oncologic setting may be successful.


Statistical Analysis: Supplementary Comparisons
In addition to the 3 major comparisons described in the main paper, we carried out supplementary comparisons to assess the impact on abstinence of: a) either having a smoking related (SMCa+) or unrelated (SMCa-) current cancer, and b) of being a cancer survivor (S). The no cancer history group (CaHx-) served as the reference.

Bivariate Analysis
We conducted simple bivariate comparisons of demographics and baseline variables shown in eTable 1 for the major cancer groupings used in the supplementary comparisons (CaHx-, SMCa+, SMCa-, S) using Chi-square for categorical comparisons and the t-test for continuous (P values are 2-sided).

Regression at Discrete Time Points
To provide covariate-adjusted risk ratios and associated standard errors we used the modified Poisson regression model [17][18][19][20] with a sandwich variance estimator to evaluate differences in abstinence for the comparisons of SMCa+ and SMCa-vs. CaHX-, and S vs. CaHx-using Stata version 14 21 . Separate models were evaluated for each comparison at each of the three time points (3,6, and 9 months after consultation), that were both unadjusted and adjusted for the demographic and baseline covariates in Table S1. Usually, logistic regression is recommended when the probability of the event of interest is low and the baseline risks for subgroups are constant. In that case, the odds ratio (OR) and risk ratio (RR) converge 22 . However, when the probability of the outcome is not rare, and the baseline risks are not constant (as it is the case in our study), the interpretation of the OR as a RR leads to exaggeration of the actual probabilities. In addition, obtaining RRs from logistic regression is not straightforward and can be extremely tedious 23 . Also, there is a consensus to report RRs instead of ORs in clinical studies, because they represent consistent estimates of the actual risk 24 , with valid confidence intervals that are easier to interpret than ORs 25 . For these reasons we chose the modified Poisson regression model over logistic regression to estimate RR point estimates with valid confidence limits. This method has been proposed as a superior alternative to logistic regression when estimating RRs  , and has been extensively used in the analyses and publication of clinical studies 18,20,19 .

Longitudinal Analysis
We used multivariate mixed effects modified Poisson regression models 18

Statistical Software and Reporting
All models were estimated in Stata statistical software version 14 21 using the "poisson" routine for the timespecific models and "xtpoisson" for the longitudinal mixed effects model modified 20,19 to estimate relative risks . For all comparisons in both sets of analyses, we estimated and report relative risks (RR) of abstinence with their corresponding 95% CIs and two-tailed P values. . Bonferroni corrections for multiple comparisons were applied by dividing .05/number comparisons X number of time points, within each model tested.

Supplementary Details for Multiple Imputation Approach
To implement the multiple imputation, we created 10 imputed datasets using Stata's "mi impute chained" command. This method uses a sequential regression multivariate imputation approach 27,28 to impute one variable at a time conditioning on all other variables, and using a Gibbs-like algorithm to obtain imputed values by simulating from the posterior predictive distribution. In addition, all imputations were performed separately for each cancer site category and were subsequently combined. Results from the multiple imputation datasets were analyzed and combined using Stata's "mi estimate" command. . In this study we evaluated abstinence rates using three different scenarios of missingness: a) Intention to Treat (ITT), b) Respondent Only (RO), and, c) Multiple Imputation (MI). This strategy of evaluating abstinence is suggested by Hedeker et al. 29 where they state that "it is important to examine results under a range of plausible values for the association of missing and smoking stratified by past smoking behavior" (p.1572). Although we did not stratify by past smoking behavior, we used past smoking behavior to predict missing smoking status and stratified by cancer group. In addition, our MI approach to imputation of missing smoking status is also recommended by Hedeker et al. 29 who state that "Our approach advocates the use of multiple imputation, because individual, sampling and imputation variation can be accounted for. " (p. 1972). Although, all three missing methods we used provided comparable estimates of abstinence, with ITT being the most conservative, we believe that the MI method is more reliable since it utilizes information from individuals on demographics, psychosocial, and smoking-related characteristics to predict smoking status. In addition, our imputation accounted for cancer group membership by multiply imputing abstinence nested within cancer group. Acknowledging the fact that when missingness is substantial, no imputation method can reliably recover the true abstinence rates 30 , we present in our manuscript all abstinence rates by follow-up time point and by imputation method (Figure 2 in manuscript). Given the relatively large number of covariates we included in the imputation model, and by nesting each imputation model within each cancer group, we believe that the MAR assumption given the covariates is plausible, and the multiple imputed estimates are the least biased from the three methods. This method of nested multiple imputation has been shown to produce reliable estimates and standard errors with nonignorable missing smoking data 31 .

Response Rates
The response rates of the study sample (N=3245) for assessing abstinence were quite high for a clinical sample:   92%, 87%, and 83% for 3-, 6-, and 9-month time points for the participants who had a cancer diagnosis, and 91%, 84%, and 77% for the 3-, 6-, and 9-month time points for the participants without a cancer diagnosis. Overall, data on abstinence were missing only for 7.8% of the observations at the 3 month time point, 13.6% at the 6-month time point and 18.5% at the 9-month time point. The baseline covariates with missing data were FTCD (nicotine dependence) at 12%, years smoke at 9% and the PHQ (presence/absence of psychiatric comorbidities) at 9%. Missingness in baseline covariates was due to nonresponse on specific questionnaire items.

eResults. Demographics and Models Demographics
As shown in eTable 1, compared with no cancer patients (CaHx-), cancer patients (CaHx+, SMCa+, SMCa-) were more likely to be older, male, white, and to smoke more (CPD), be more nicotine dependent (higher FTCD), ); and to have smoked longer. Those with a cancer history or smoking related (CaHx+, SMCa+) also had a higher risk of having depression on the PHQ. Survivors (S) were older, White, smoked longer, and had a higher risk of anxiety on the PHQ. The smoking related cancer group was slightly more likely to be prescribed smoking cessation medication (96% vs. 94%).

Smoking Related Cancer and ITT and Longitudinal Models
No significant differences in abstinence rates were found in the comparisons of no cancer history (CaHx-) smoking-related (SMCa+) or non-smoking-related (SMCa-) cancer vs CaHx-, or survivors (S) at any the 3-, 6-, and 9month follow-ups, nor in the longitudinal models (eTables 5-8). While nominally significant differences were noted in comparisons between SMCa+ and SMCa-groups (see eTables 5 and 6) in the multiply imputed sample, these effects did not survive correction for multiple comparisons (p<.016).
The ITT and respondent-only results were consistent with the multiply imputed results and showed no significant differences for these same comparisons (see supplement eTables [13][14][15][16]