^{1}

^{2}

^{3}

^{4}

This is an open-access article distributed under the terms of the Creative Commons Attribution License (https://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work, first published in the Journal of Medical Internet Research, is properly cited. The complete bibliographic information, a link to the original publication on http://www.jmir.org/, as well as this copyright and license information must be included.

Project Quit was a randomized Web-based smoking cessation trial designed and conducted by researchers from the University of Michigan, where its primary outcome was the 7-day point prevalence. One drawback of such an outcome is that it only focuses on smoking behavior over a very short duration, rather than the quitting process over the entire study period.

The aim of this study was to consider the number of quit attempts during the 6-month study period as an alternative outcome, which would better reflect the quitting process. We aimed to find out whether tailored interventions (high vs low) are better in reducing the number of quit attempts for specific subgroups of smokers.

To identify interactions between intervention components of smoking cessation and individual smoker characteristics, we employed Poisson regression to analyze the number of quit attempts. This approach allowed us to construct data-driven, personalized interventions.

A negative effect of the number of cigarettes smoked per day (

A highly individually tailored story is significantly more effective for smokers with a low level of education. This is consistent with prior findings from Project Quit based on the 7-day point prevalence.

Smoking is the leading preventable cause of death worldwide [

In modern quantitative precision medicine literature, the idea of personalizing treatments to individual patients is often operationalized as a

Point prevalence is often a popular choice in assessing smoking cessation. In fact, Project Quit study was designed with a 7-day point prevalence as the primary outcome [

This study aims to identify a subgroup of smokers who are most likely to benefit from Web-based tailored behavioral interventions for smoking cessation. We will identify this subgroup based on their willingness and involvement in the quitting process, measured by the number of quit attempts during the study period. Such an approach will potentially allow health care researchers to use the limited public health resources more efficiently in shaping the health care policy.

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 [

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 ^{5-1}) fractional factorial design [

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 [

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 [

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|X_{1}, X_{2}, …, X_{8}, A_{1}, A_{2})) = β_{0} + β_{1} X_{1} + β_{2} X_{2} +…+ β_{8} X_{8}+ (δ_{0} + δ_{1} X_{7})A_{1}+ (η_{0} + η_{1} X_{8})A_{2}, where _{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 A_{1} and A_{2} denote the intervention components story and source, respectively. The notation E(Y|X_{1}, X_{2}, …, X_{8}, A_{1}, A_{2}) denotes the conditional expectation (conditional mean) of _{i}, i=0, 1, …, 8; δ_{0}, δ_{1}, η_{0}, η_{1}) in Poisson regression are estimated by the maximum likelihood method. We used open-source software R, version 3.2.3 [

Regression coefficients β_{i}, i=1, …, 8, denote the main effects of the covariates X_{i}, i=1, 2, …, 8; β_{0} denotes the model intercept; δ_{0} and η_{0} denote the main effects of intervention components A_{1} and A_{2}, respectively; and finally, δ_{1} and η_{1} denote the preconceived interaction effect between X_{7} and A_{1} and that between X_{8} and A_{2}, 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 A_{1}(story), can be expressed as E(Y|X_{1}, X_{2}, …, X_{8}, A_{1}=1, A_{2}) – E(Y|X_{1}, X_{2}, …, X_{8}, A_{1}=0, A_{2}) = (exp(δ_{0} + δ_{1} X_{7}) – 1) E(Y|X_{1}, X_{2}, …, X_{8}, A_{1}=0, A_{2}), which in turn can be interpreted as—given all other covariates are fixed, a highly tailored story (A_{1}=1) increases the expected number of quit attempts by (exp(δ_{0} + δ_{1} X_{7}) – 1)100% compared with the low-tailored story (A_{1}=0). Similarly, for A_{2}(source), it can be interpreted that a highly personalized source increases the expected number of quit attempts by (exp(η_{0}+ η_{1} X_{8}) – 1)100%, compared with the low-personalized source. Furthermore, for any of the baseline covariates, the effect of the _{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.

Before the primary data analyses, we examined potential differential missingness across the intervention arms and found no significant difference (

The per-protocol participants’ baseline characteristics (n=792) are summarized in

Participant characteristics. Descriptive summary refers to mean (SD) for continuous characteristics and frequency (percentage) for categorical variables.

Participant characteristics | Descriptive summary (n=792) | |

Age in years, mean (SD) | 46.32 (10.64) | |

Female | 480 (60.6) | |

African American | 97 (12.3) | |

White | 615 (77.7) | |

Other | 80 (10.1) | |

>High school | 502 (63.4) | |

≤High school | 290 (36.6) | |

Number of cigarettes smoked per day, mean (SD) | 21.51 (8.94) | |

High | 362 (45.7) | |

Low | 430 (54.3) | |

High | 421 (53.2) | |

Low | 371 (46.8) | |

Deeply tailored | 407 (51.4) | |

Low-tailored | 385 (48.6) | |

Highly personalized | 395 (49.9) | |

Low-personalized | 397 (50.1) |

The estimated Poisson regression coefficients, z-scores, RR values along with their 95% CIs, and corresponding

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.

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

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.

Summary results of the Poisson regression model for the number of quit attempts outcome (n=792).

Variable | Regression parameter estimate | Z-score | Adjusted rate ratio (95% CI) | |

Age (years) | 0.001 | 0.662 | 1.001 (0.997-1.005) | .51 |

Gender (male) | 0.070 | 1.572 | 1.073 (0.983-1.171) | .12 |

Race (dummy for white) | −0.041 | −0.577 | 0.959 (0.833-1.105) | .56 |

Race (dummy for Black) | 0.134 | 1.504 | 1.143 (0.960-1.361) | .13 |

Number of cigarettes smoked per day (NCigs per day) | −0.006 | −2.185 | 0.994 (0.989-0.999) | .03^{a} |

Motivation | 0.055 | 1.148 | 1.056 (0.962-1.160) | .25 |

Education | 0.144 | 2.230 | 1.155 (1.018-1.311) | .03^{a} |

Self-efficacy | −0.054 | −0.840 | 0.948 (0.836-1.074) | .40 |

Story | 0.047 | 0.643 | 1.048 (0.908-1.209) | .52 |

Source | −0.082 | −1.285 | 0.921 (0.813-1.044) | .20 |

Story × education | −0.192 | −2.125 | 0.825 (0.692-0.985) | .03^{a} |

Source × self-efficacy | 0.080 | 0.926 | 1.084 (0.914-1.284) | .36 |

^{a}Denotes

Estimated intervention effect of story, expressed as a percentage change in the expected number of quit attempts, stratified by education level, mathematically expressed as (exp(δ_{0} + δ_{1} Edu) – 1)100% (n=792).

Education | Estimate (95% CI) |

>High school (Edu=1) | −13.50 (−22.58 to −4.42) |

≤High school (Edu=0) | 4.798 (−10.17 to 19.76) |

An estimated treatment regimen to recommend personalized smoking cessation intervention for the whole population. Edu: education; HS: high school.

Summary results of the Poisson regression model for participants who smoked less than or equal to 20 cigarettes per day (n=546).

Variable | Estimate | Z value | Adjusted rate ratio (95% CI) | |

Age (years) | 0.002 | 0.676 | 1.002 (0.997-1.006) | .50 |

Gender (male) | 0.083 | 1.539 | 1.087 (0.972-1.202) | .12 |

Race white | −0.148 | −1.813 | 0.862 (0.724-1) | .07 |

Race black | −0.051 | −0.504 | 0.950 (0.762-1.139) | .61 |

Number of cigarettes smoked per day (NCigs per day) | −0.014 | −2.108 | 0.986 (0.973-0.999) | .03^{a} |

Motivation | 0.095 | 1.687 | 1.099 (0.978-1.221) | .09 |

Education | 0.256 | 3.142 | 1.291 (1.085-1.497) | .002^{a} |

Self-efficacy | −0.159 | −2.131 | 0.853 (0.729-0.978) | .03^{a} |

Story | 0.189 | 2.022 | 1.208 (0.987-1.428) | .04^{a} |

Source | −0.077 | −1.028 | 0.926 (0.790-1.062) | .30 |

Story × education | −0.414 | −3.700 | 0.661 (0.516-0.806) | <.001^{a} |

Source × self-efficacy | 0.052 | 0.499 | 1.053 (0.839-1.267) | .62 |

^{a}Denotes

Estimated intervention effect of story, expressed as a percentage change in the expected number of quit attempts, stratified by the education level for the persons who smoked less than or equal to 20 cigarettes per day (n=546).

Education | Estimate (95% CI) |

>High school (Edu=1) | −20.20 (−29.93 to −10.47) |

≤High school (Edu=0) | 20.76 (−1.32 to 42.83) |

An estimated treatment regimen to recommend personalized smoking cessation intervention for smokers who smoked less than or equal to 20 cigarettes per day at baseline. Edu: education; HS: high school; NCigs: number of cigarettes smoked.

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 [

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 [

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 [

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.

Group Health Cooperative

Henry Ford Health System

health maintenance organization

National Cancer Institute

number of cigarettes smoked per day

rate ratio

treatment regimen

The original Project Quit trial was funded by the US NCI grants P50 CA101451 and R01 CA101843. This work (secondary analysis of Project Quit data) was supported by BC’s start-up grant from Duke-NUS as well as MOE2015-T2-2-056 grant from Singapore’s Ministry of Education.

None declared.