Project Quit was a Web-based smoking cessation program developed and conducted by the Center for Health Communications Research at the University of Michigan, Ann Arbor, and was funded by the National Cancer Institute (NCI), USA. The study protocol was reviewed and approved by the Institutional Review Board of each collaborating institution and of the University of Michigan in January 2004. The primary aim of the study was to identify and test the effects of 5 psychosocial and communication intervention components influencing smoking cessation [8]. The content of the Web-based intervention was based on cognitive-behavioral methods of smoking cessation, including an appeal to motives for quitting, stimulus control, self-efficacy enhancement, and suggestion for coping with tempting situations and emotions. Hypothetical success stories were employed within this overall paradigm.

Five intervention components (outcome expectations, efficacy expectations, success stories, message source, and message exposure) were studied in Project Quit. To screen multiple components, the study employed a multiphase optimization strategy [19] framework, implemented using a 16-cell (2^5-1) fractional factorial design [20,21] in which each of the 5 intervention components were varied at 2 levels, high vs low. However, only 2 components, namely, success stories (hereafter referred to as story) and message source (hereafter referred to as source), were found to have significant effects on smoking cessation in previous analysis [8]. These 2 components were a priori hypothesized to have the strongest effect on smoking cessation. On the basis of these findings, we considered these 2 intervention components in our analysis. The intervention component story refers to success story received by study subjects from a hypothetical character who succeeded in quitting smoking. The story was varied at 2 levels—high vs low tailoring depth (ie, the degree to which the character in the story was tailored to subject’s baseline characteristics). Similarly, the component source refers to the source of Web-based behavioral counseling message received by subjects and was varied at 2 levels—high versus low level—of personalization. High-personalized source included photograph and supportive text from the health maintenance organization’s (HMO) smoking cessation team. It was written in a friendly language and included a signature from the team. In contrast, the low-personalized version included a photograph of a building, representing the HMO, and was impersonally written without a closing signature. Strecher et al [8] provided detailed description of these components, including examples of actual Web-based messages.

Adult participants were recruited from 2 HMOs—Group Health Cooperative (GHC), Seattle; and Henry Ford Health System (HFHS), Detroit; both these HMOs were affiliated with the NCI’s cancer research network. The study participants had a broad representation of ethnicity, gender, age, health status, and geography. Participants’ eligibility criteria included those who (1) had smoked at least 100 cigarettes in his or her lifetime, currently smoked at least 10 cigarettes per day, and had smoked in the past 7 days; (2) were seriously considering quitting in the next 30 days; (3) were 21 to 70 years; (4) were members of either GHC or HFHS; (5) had home or work internet access and an email account that they used at least twice weekly; (6) were not currently enrolled in other smoking cessation program(s) and not currently using pharmacotherapy for smoking cessation; and (7) had no medical contraindications for nicotine replacement therapy. A total of 1866 subjects participated in Project Quit. All subjects received free Web-based experimental smoking cessation program. Participants were randomized to receive either high or low personalized intervention, as described above. To pharmacologically assist them with smoking cessation, all participants, irrespective of their intervention group, also received a free 10-week supply of nicotine replacement therapy patches. Thus, this study allowed participants to focus on the cognitive-behavioral aspects of smoking cessation through combination of various intervention components.

The primary outcome of the study was the binary 7-day point prevalence in smoking cessation at 6 months following baseline assessment. During the 6-month evaluation survey, each subject was asked if she or he had smoked any cigarettes, even a puff, in the last 7 days. Subject who answered “yes” was marked as smoker and nonsmoker otherwise. In addition, data on the number of quit attempts in the past 6 months were collected as a secondary outcome, which is the focus of this study.

In addition to baseline covariates (age, gender, and race), Project Quit also collected variables deemed relevant for smoking cessation. These included (1) number of cigarettes smoked per day as a measure of baseline addiction; (2) the participant’s level of motivation to quit smoking as a predictor of smoking cessation [8]; (3) the participant’s level of education, which was hypothesized to interact with the intervention component story [8]; and (4) participant’s self-efficacy, a consistent predictor of subsequent health-related behavior change based on the social cognitive theory [22].

Of the 1866 subjects who enrolled in the Project Quit study, 1192 subjects responded to the question on the number of quit attempts. Of these responders, 792 subjects followed the study protocol by not using other smoking cessation aids or programs during the study. As the primary examination in Project Quit [8] utilized per-protocol analysis that only included subjects who did not violate study protocol, we used the same strategy to analyze the number of quit attempts in the 792 subjects.

To assess the potential presence of differential missingness across the intervention arms, we conducted chi-square test with 2 categorical variables—intervention arm (4 levels resulting from 2 intervention components, each varied at 2 levels) and nonresponse (2 levels, yes or no).

Baseline covariates considered in this analysis were age (continuous), gender (binary), race (3 levels, but handled by 2 dummy variables—race white and race black), cigarettes smoked per day (continuous), motivation (binary, high vs low, coded 1 or 0), self-efficacy (binary, high vs low, coded 1 or 0), and education (binary, ≤high school vs >high school, coded 0 or 1). The source and story levels were coded as 1 (high) and 0 (low), respectively.

We used the Poisson regression model to analyze the number of quit attempts. The Poisson regression model can be applied to settings where the outcome is a count-type variable with its expectation (mean) varying as a log-linear function of the covariates and intervention components. The model used in this analysis can be specified as log(E(Y|X₁, X₂, …, X₈, A₁, A₂)) = β₀ + β₁ X₁ + β₂ X₂ +…+ β₈ X₈+ (δ₀ + δ₁ X₇)A₁+ (η₀ + η₁ X₈)A₂, where Y denotes the number of quit attempts during the 6-month study period; X_i, i=1, 2, …, 8, denote the baseline characteristics, viz, age, gender, race white, race black, cigarettes smoked per day, motivation, education, and self-efficacy, respectively; and A₁ and A₂ denote the intervention components story and source, respectively. The notation E(Y|X₁, X₂, …, X₈, A₁, A₂) denotes the conditional expectation (conditional mean) of Y, given all the baseline covariates and intervention components. The unknown parameters (β_i, i=0, 1, …, 8; δ₀, δ₁, η₀, η₁) in Poisson regression are estimated by the maximum likelihood method. We used open-source software R, version 3.2.3 [23] for the analysis.

Regression coefficients β_i, i=1, …, 8, denote the main effects of the covariates X_i, i=1, 2, …, 8; β₀ denotes the model intercept; δ₀ and η₀ denote the main effects of intervention components A₁ and A₂, respectively; and finally, δ₁ and η₁ denote the preconceived interaction effect between X₇ and A₁ and that between X₈ and A₂, respectively. Instead of reporting the estimates of β_i, we reported the corresponding adjusted incidence rate ratios, or simply the rate ratios (RRs). These quantities offer a more interpretable way to report results from a Poisson regression model (analogous to reporting odds ratios from a logistic regression model for binary data). Under the above setup, we defined RR for a covariate X_i as the ratio of the expectation of Y given that X_i=1 and the expectation of Y given that X_i=0 (for binary X_i) or as the ratio of the expectation of Y given that X_i=x+1 and the expectation of Y given that X_i=x for some arbitrary value x (for continuous X_i), given that other variables in the model (both covariates and interventions) are fixed. This RR can then be computed as the exponential transform of the regression coefficient (exp(β_i)). RR measures change in the expected outcome when X_i increases by 1 unit (for continuous X_i), or when X_i moves from 1 category to the other (for categorical X_i) on a multiplicative scale.

Re-expression of intervention effects may further facilitate interpretation. The effect of a particular intervention component, say A₁(story), can be expressed as E(Y|X₁, X₂, …, X₈, A₁=1, A₂) – E(Y|X₁, X₂, …, X₈, A₁=0, A₂) = (exp(δ₀ + δ₁ X₇) – 1) E(Y|X₁, X₂, …, X₈, A₁=0, A₂), which in turn can be interpreted as—given all other covariates are fixed, a highly tailored story (A₁=1) increases the expected number of quit attempts by (exp(δ₀ + δ₁ X₇) – 1)100% compared with the low-tailored story (A₁=0). Similarly, for A₂(source), it can be interpreted that a highly personalized source increases the expected number of quit attempts by (exp(η₀+ η₁ X₈) – 1)100%, compared with the low-personalized source. Furthermore, for any of the baseline covariates, the effect of the i-th covariate can be expressed as (exp(β_i) – 1)100%, i=1, …, 8.

We used the standard 5% alpha level to assess statistical significance in our analyses. Whenever appropriate, we also reported the 95% CIs of various effects. On the basis of the Poisson regression results, we then derived the corresponding TRs for recommending personalized smoking cessation interventions and drew decision trees to visually represent TRs.

We expected smokers’ baseline level of addiction, as measured by the number of cigarettes smoked per day (and found in our analysis results presented below), to influence the number of quit attempts. Therefore, once the Poisson regression analysis on the full data was completed, we divided the participants into 2 subgroups: (1) those who used to smoke less than or equal to the observed median of the number of cigarettes smoked per day and (2) those who used to smoke more than the observed median of the number of cigarettes smoked per day. We then repeated the Poisson regression analysis for each of the subgroups.

The per-protocol participants’ baseline characteristics (n=792) are summarized in Table 1. These subjects had a mean age of 46.32 (SD 10.64) years. Of these, 60.6% (480/792) were female, 77.7% (615/792) were white, 12.3% (97/792) were African Americans, 63.4% (502/792) were highly educated, 53.2% (421/792) had high self-efficacy, and 45.7% (362/792) were highly motivated. On average, the participants used to smoke 21.51 (SD 8.94) cigarettes per day at baseline. With respect to randomized interventions, 51.4% (407/792) subjects received a highly tailored story, and 49.9% (395/792) subjects received a highly personalized source. During the 6-month study period, 93.3% (739/792) participants attempted to quit at least once. The number of quit attempts varied from 0 to 10, and the mean quit attempt was 2.74 (SD 2.50) times.

The estimated Poisson regression coefficients, z-scores, RR values along with their 95% CIs, and corresponding P values are reported in Table 2. After adjusting for relevant covariates and treatment components, only the number of cigarettes smoked per day at baseline (RR=0.994; 95% CI 0.989-0.999; P=.03), education (RR=1.155; 95% CI 1.018-1.311; P=.03), and education-by-story interaction (RR=0.825; 95% CI 0.692-0.985; P=.03) were significant. This means that when all other covariates are fixed in the model, for every extra cigarette smoked per day at baseline, the expected number of quit attempts in the 6-month study period changes by a multiplicative factor of 0.994. In other words, for every 10 extra cigarettes smoked per day at baseline, there is a 5.84% decrease in the expected number of quit attempts over 6 months. On the other hand, there is a 15.49% increment in the expected number of quit attempts for highly educated participants (P=.03), as compared with those with low education.

Interaction between education and story is interpreted differently from the main effects of individual covariates. For this scenario, the main effect and interaction effect should be interpreted jointly. Table 3 shows the effects of both high- and low-tailored stories on the 2 levels of education. Giving a high-tailored story to a highly educated smoker results in a 13.5% decrease in the number of quit attempts compared with a low-tailored story (P=.03). However, the result is completely reverse for the low-educated group, where a high-tailored story given to a low-educated person increases the expected number of quit attempts by 4.8% compared with a low-tailored story. On the basis of this result, we derived a TR that recommends personalized smoking cessation interventions (high-tailored story for low-educated subjects and low-tailored story for those who are highly educated). A decision tree to visualize this TR is shown in Figure 1.

We have shown that smokers’ baseline level of addiction, as measured by the number of cigarettes smoked per day, has a negative impact on the number of quit attempts. Using the observed median of 20 cigarettes smoked/day as a threshold, we further divided the participants into 2 subgroups: (1) those who used to smoke ≤20 cigarettes/day and (2) those who used to smoke >20 cigarettes/day. We found that severe smokers (>20 cigarettes/day at baseline) were not influenced by any intervention components. However, less severe smokers with lower education were more influenced by the highly tailored story, which is similar to the whole group of smokers in the study. Results from the Poisson regression analyses for the less severe subgroup of smokers (≤20 cigarettes/day at baseline) are shown in Tables 4 and 5, and can be interpreted in a similar fashion as above.

In addition to the effects that were significant in the full dataset, the main effects of self-efficacy and the intervention component story also came out significant in this subgroup analysis. Figure 2 shows the decision tree for the associated TR.

This study aimed to stratify smokers who are likely to benefit from tailored smoking cessation intervention programs and those who are not. This will allow us to develop personalized smoking cessation interventions. Outcomes of this study can potentially help policy makers to allocate limited public health resources to target subgroups of smokers who are more likely to be successful from tailored smoking cessation interventions. This study analyzed existing data from a large randomized, Web-based smoking cessation trial (Project Quit) to answer the above research question.

In Project Quit, Strecher et al [8] previously studied the impact of 5 Web-based intervention components on the 7-day point prevalence. Using a multivariable logistic regression model, they found that 2 intervention components, namely, story and source, have significant effect on smoking cessation at 6 months. Furthermore, they showed that participants with lower education were more influenced by highly tailored stories, and a highly personalized source had marginally greater impact on smoking cessation in participants with higher self-efficacy. However, 1 drawback with the 7-day point prevalence is that it does not take into account the quitting process over the entire study period. Instead, it only focuses on smoking behavior in a limited time window at the end of the 6-month follow up. Our investigation was designed to specifically overcome this limitation by considering the number of quit attempts during the whole study period as the outcome of interest and examine whether similar effects still hold. The number of quit attempts quantified participants’ willingness and involvement in the smoking cessation process throughout the study period. Our study incorporated 2 findings from prior analysis of Project Quit [8] into this analysis. First, of the 5 intervention components from the original trial, we considered only 2 (story and source) into our model because of their significant effects (remaining components were insignificant). Second, based on the a priori hypothesis and data analysis from Strecher et al [8], we only included 2 interaction effects—one between story and education, and another between source and self-efficacy.

We found that participants with lower education (high school graduates or less) were positively influenced by a high-tailored story to quit smoking, whereas those with higher education were better off with a low-tailored story. Our findings are consistent with those from Strecher et al [8], who found similar effect modification, albeit in the context of the 7-day point prevalence. Findings on the low education group are not surprising as participants in this category can easily associate themselves with fictitious characters in the story that are tailored to them (socioeconomically or otherwise), as opposed to untailored characters. Such association allows the low-educated subjects to “transport” themselves into the story, thus resulting in successful smoking cessation. Strecher et al [8] suggested that the extent of being “transported” has a strong influence on persuasion, which in particular applies to smokers’ attempt and behavior to quit [24,25]. In contrast to Strecher et al [8], we did not find any significant interaction between source and self-efficacy in influencing the number of quit attempts. We speculate that such discrepancy may be due to different outcome under consideration or the smaller sample size in this investigation.

Here, we summarize the strengths of this study that was designed to address gaps in the extant literature on smoking cessation. First, ours is the first analysis of quit attempts data from Project Quit. This will potentially shed new light on smokers’ quitting process experience while participating in a Web-based smoking cessation program. Second, although analysis of quit attempts data are available in the literature [14-18], they are based on observational cohort studies. Our study is the first to analyze quit attempts data from a randomized trial. Third, we utilize TR as a perspective from the precision medicine literature to help better understand the type of smokers who will benefit from tailored intervention in optimizing their quitting effort. The results suggest that smokers with low education are more likely to benefit from tailored interventions. This is consistent with prior findings based on the 7-day point prevalence data [8,20,26]. Thus, this analysis validates the significance of number of quit attempts as an alternative to the commonly used point prevalence outcome. Finally, from a methodological perspective, number of quit attempts is a count-type variable rather than a continuous measurement or binary indicator. Hence, we employed Poisson regression to analyze the dataset.

There are a few limitations in our study. First, because the number of quit attempts was a secondary outcome in Project Quit, this variable had higher rate of missingness compared with the primary outcome of the 7-day point prevalence. For simplicity and easy interpretation, we only conducted a complete-case analysis. One could potentially employ missing data analysis techniques (eg, multiple imputation) to impute the missing values before conducting the analysis. However, as we did not find any evidence of differential missingness across the intervention arms, we argue that the missingness in the current data is mostly noninformative. Thus, data imputation techniques would not offer much benefit over a complete-case analysis. Second, the intervention components in the original study were designed to influence the 7-day point prevalence. One could conceive other potential intervention components not studied in Project Quit, which may potentially better influence smokers’ involvement in their difficult journey toward quitting and their number of quit attempts in particular. We believe that new studies specifically designed to understand the impact of tailored interventions on quit attempts are necessary to answer such questions. Third, the number of quit attempts is a self-reported outcome over a reasonably long period. As it is unrealistic for the participants to remember their exact number of quit attempts in the past 6 months, this variable may have recall bias. However, this concern can be addressed in the current era of mobile health and sensor technologies. New-generation studies should employ mobile apps and wearable devices to capture quit attempts data more accurately and thus minimize measurement errors.

In this study, we aimed to shed new lights on the impacts of Web-based tailored psychosocial and communication intervention components on smoking cessation. Using data from a randomized Web-based trial, we examined the number of quit attempts during a 6-month study period. We also investigated how these impacts are modified by individual characteristics. Collectively, we aimed to identify subgroups of smokers who would successfully benefit from Web-based tailored interventions. We found that highly individually tailored story is significantly more effective for smokers with low education (high school graduate or less) compared with those with higher education (at least some college exposure). Our findings can provide evidence and potentially help policy makers to utilize limited public health resources to cease smoking in low-educated smokers. Nevertheless, we must cautiously note that the number of quit attempts in this study is self-reported, and thus subjected to recall bias. Future studies that incorporate sensor and/or mobile technologies to collect precise data on quit attempts are clearly warranted.