Time-Varying Mediation of Pharmacological Smoking Cessation Treatments on Smoking Lapse via Craving, Cessation Fatigue, and Negative Mood

Abstract

Introduction

The addictive nature of nicotine makes smoking cessation an extremely challenging process. With prolonged exposure, tobacco smoking transforms from being a positive reinforcer to a negative one, as smoking is used to mitigate aversive withdrawal symptoms. Studying the variations in withdrawal symptoms, especially during their peak in the first week of a quit attempt, could help improve cessation treatment for the future. The time-varying mediation model effectively studies whether altering withdrawal symptoms act as mediators in the pathway between treatment and cessation.

Aims and Methods

This secondary data analysis of a randomized clinical smoking cessation trial of three pharmacotherapy regimens (nicotine patch, varenicline, and nicotine patch + mini-lozenge) analyzes ecological momentary assessment (EMA) data from the first 4 weeks post-target quit day (TQD). We assess whether withdrawal symptoms (eg, negative mood, cessation fatigue, and craving) mediate the pathway between pharmacotherapy and daily smoking status and whether this effect varies over time.

Results

We found a statistically significant time-varying mediation effect of varenicline on smoking status through craving, which shows decreasing risk of lapse via reduction in craving. We did not find significant time-varying mediation effects through negative mood and cessation fatigue.

Conclusions

This study supports the importance of craving suppression in the smoking cessation process. It also helped identify specific timepoints when withdrawal symptoms increased that would likely benefit from targeted cessation intervention strategies.

Implications

This study aimed to understand the underlying dynamic mechanisms of the smoking cessation process using a new analytical approach that capitalizes on the intensive longitudinal data collected via EMAs. The findings from this study further elucidate the smoking cessation process and provide insight into behavioral intervention targets and the timing of such interventions through the estimation of time-varying mediation effects.

Introduction

The prevalence of tobacco smoking has been in a steady decline since the 1960s. Nevertheless, it persists as the leading cause of preventable morbidity and mortality in the United States.¹ Much of the progress can be attributed to the continued development of effective pharmaco-behavioral treatments and strong tobacco control programs and the increased motivation to quit (eg, 68% of current cigarette smokers² want to quit smoking entirely). However, the long-term abstinence rates for a given quit attempt rarely improve beyond 30%, even with evidence-based treatment (2008 PHS Guideline). Tobacco dependence is a chronic relapsing disease (2008 PHS Guideline). Chronic exposure to nicotine, the addictive element in tobacco, results in neuroadaptation so that abstinence produces withdrawal symptoms (eg, craving, negative affect), thereby increasing the likelihood of relapse.³

Tobacco withdrawal syndrome “is a collection of characteristic symptoms or signs typically instigated by reduced intake of a dependence-producing drug following a period of sustained exposure.”⁴ The impact of tobacco withdrawal syndrome on smoking cessation is well established.^3,5–10 Extant literature documents two contradictory theories, one that negates withdrawal symptoms as a key “explanatory construct” of relapse (Marlatt’s Cognitive-Behavioral Theory)^11,12 and the other which posits a strong association between severity of tobacco withdrawal and the likelihood of relapse (Wikler’s Pharmacologic Theory).¹³ Research challenging withdrawal symptoms as vital predictors of relapse has been attributed to the inaccurate measurement of withdrawal symptoms, and since the early 2000s, evidence supporting the association between different dimensions of withdrawal syndrome and relapse has grown.^4–8

Based on these correlational inferences, withdrawal-based nicotine dependence and addiction theories have been extended to suggest that one of the potential mechanisms of cessation pharmacotherapy on abstinence among smokers is via withdrawal reduction.³ However, using mediation techniques to evaluate the mechanisms underlying the treatment effect on relapse has been limited primarily due to inadequate longitudinal outcome data and a lack of analytical strategy to build support for a causal relationship. To address some of these concerns, Piper et al.³ examined whether the overall effect of pharmacotherapy on successful smoking cessation was mediated via withdrawal suppression. The analysis was executed using structural equation models with a composite latent mediator defined by the growth trajectory (quit-day to 1-week post-quit change) coefficients of potential mediators. However, this study did not address the indirect (mediated) effect of the pharmacotherapy via the trajectory of the withdrawal symptoms beyond week 1 post-quit. Also, the growth trajectory coefficient approach summarized the withdrawal symptoms experienced over a week, thus failing to capture the granularity of the cessation process, ie, the effect of pharmacotherapy-induced changes in momentary withdrawal on the likelihood of future abstinence.

The focus on assessing week 1 post-quit withdrawal and its cumulative impact on long-term cessation has precedence in smoking cessation literature. It is consistent with withdrawal theories that posit that the smoking withdrawal syndrome tends to peak in the first week or two of abstinence and usually resolves over a month.¹⁴ Nevertheless, research also shows that the different components of withdrawal syndrome have different profiles and predict cessation outcomes uniquely.⁹ For example, Piper et al.⁹ report that compared with negative affect, craving profiles were more heterogenous and intense over time. In addition, researchers have explored how the interrelationships of momentary withdrawal symptoms predict initial lapses,^15–19 since initial lapses are strongly associated with future relapses. Findings highlight the complexities of the smoking cessation process—the effect of pharmacological agents is more pronounced in the initial phase of a quit attempt, beyond which craving, negative affect, and cessation fatigue continue increasing, and additional support is required to suppress these withdrawal symptoms.¹⁵ These observations lead us to hypothesize that not only does each withdrawal symptom represent a different pathway through which the mechanism of abstinence/relapse is set in motion, but that these mechanisms change in their strengths of effect with time.

The primary public health goal is to increase sustained abstinence from tobacco smoking. One key to achieving this goal is to continue to develop a more fine-grained understanding of the impact of withdrawal symptoms on subsequent smoking and identify critical timepoints of relapse risk to enhance targeted treatment. Therefore, even when there are no statistically significant differences in abstinence rates between treatment conditions, it is still important to understand the mechanisms via which treatments exert their effects (how), for whom they work (subgroups), at what time they exert these effects (when), and to identify the etiology of abstinence and relapse (why) among individuals attempting to quit.

With the rise of the availability of ecological momentary assessment (EMA) data in behavioral research, ie, the repeated sampling of subjects’ current experiences and behaviors in their natural, real-time environment, monitoring complex dynamic processes such as smoking cessation without recall bias has been facilitated.²⁰ However, statistical methods that fully capitalize on intensive longitudinal data collected via EMA are limited. Existing methods to analyze EMA data include multilevel modeling and time-varying effects models. These techniques are suitable for evaluating person-level effects on the heterogeneity of daily smoking behavior and how the association between withdrawal symptoms and cessation outcomes changes over time, respectively. Thus, a time-varying mediation model is needed to provide better insight into the time-varying treatment effects (direct and indirect effects) on cessation outcomes via time-varying withdrawal symptoms, and research suggests that treatment effects on withdrawal symptoms also varies with time.¹⁵

From previous studies,^15–19 we understand that craving, negative affect, and cessation fatigue are strongly associated with relapse risk, making these constructs potential targets for cessation interventions. Additionally, the effect of pharmacological treatments on these constructs and relapse risk has also been investigated,^15,18,21,22 showing better effectiveness earlier in the quit attempt. However, the role of these constructs as mediators in the effect of pharmacological treatments on daily smoking status during a quit attempt and the time-varying nature of this mechanism has not been studied. To address this gap, we hypothesize that these symptoms mediate the effect of pharmacotherapy on smoking status, and this effect varies as a function of time. To evaluate this overarching hypothesis, we will examine whether the time-varying indirect effects of pharmacological treatments on smoking status (binary outcome) are mediated by continuous time-varying mediators—craving, negative mood, and cessation fatigue, using EMA²⁰ data from a smoking cessation trial. Specifically, we address three hypotheses:

1. The time-varying mediated (indirect) effect of pharmacological treatments on smoking status through momentary craving is present earlier in a quit attempt and this effect is distinct for different treatment arms (H1).

2. The time-varying mediated effect of pharmacological treatments on smoking status through cessation fatigue is present later in a quit attempt (when self-regulation starts waning) and this effect is distinct for different treatment arms (H2).

3. The time-varying mediated effect of pharmacological treatments on smoking status through negative mood is present throughout a quit attempt and this effect is distinct for different treatment arms (H3).

Through these hypotheses we aim to understand two facets of the smoking cessation process—first, whether the time-specific patterns of indirect effects identify potential symptom(s) that could be targeted by interventions to improve the abstinence rate; second, whether there are critical timepoints at which such targeted interventions should be delivered.

Methods

We used data from a randomized open-label clinical trial, the Wisconsin Smokers’ Health Study II (WSHS2—clinicaltrials.gov Identifier: NCT01553084), which compared the efficacies of three pharmacotherapies (nicotine patches, varenicline, and nicotine patch + nicotine mini-lozenges, ie, combination nicotine replacement therapy—cNRT) in improving 26-week cessation rates. Study participants consisted of 1086 smokers from Madison and Milwaukee, Wisconsin, who were motivated to quit smoking. Specifics on screening and recruitment are detailed in Baker et al.²³

Eligible participants who met the inclusion criteria were randomized to one of the three 12-week treatment arms. All participants received smoking cessation counseling.²³ Participants provided information on smoking history, nicotine dependence, sociodemographics, and economic status via questionnaire assessments at baseline. Participants also received EMA prompts each day for 5 weeks, 1-week pre-target quit day (TQD) through week 4 post-TQD. The prompts were administered thrice daily (morning, afternoon, and evening) from week 1 pre-TQD through week 2 post-TQD and every other day in the last 2 weeks of these 5 weeks. Therefore, there are 21 EMAs pre-TQD and 63 EMAs post-TQD. In the morning prompt, smoking status was recorded via two questions, “Have you smoked any cigarettes since you woke up?” and, if appropriate, “How many cigarettes have you smoked since you woke up?” In the evening prompt, smoking status was determined from the question, “Over the whole day today, how many cigarettes did you smoke?” The randomly timed afternoon prompt only assessed smoking in the last 2 hours. Thus, we used the morning and evening prompts to assess any smoking since waking (morning smoking) and post-morning prompt smoking, respectively, creating 14 EMAs pre-TQD and 42 EMAs post-TQD corresponding to smoking status. For example, an individual who reported smoking five cigarettes over the whole day and smoking five cigarettes at the morning prompt (ie, 0 cigarettes smoked since the morning prompt) would be assigned the smoking status “Yes” for morning and “No” for the post-morning datapoints.

The EMA prompts also assessed craving, negative mood, and cessation fatigue using items that were consistent with prior research.²⁴ Participants were asked to rate their cravings and negative affect in the last 15 minutes on 7-point Likert scales (ie, 1 = not at all and 7 = extremely) via the questions, “Have you felt like wanting to smoke in the last 15 minutes?” and “Have you felt any negative emotion in the last 15 minutes?,” respectively. Cessation fatigue was assessed using agreement with the item, “I am tired of trying to quit smoking?” with 1 indicating “Strongly disagree” and 7 meaning “Strongly agree.”

Statistical Analysis

Mediation analysis decomposes the total effect, τ, of an exposure, X, on an outcome, Y, into direct and indirect effects. The direct effect, denoted γ, is defined as the effect of the exposure on the outcome (X → Y) that is not mediated via a third (intermediate) variable (M). The indirect effect (X → M → Y), is defined as the effect of the exposure on the outcome that is due to the mediator. Traditional mediation analysis estimates the direct and indirect effects through regression coefficients and defines the mediated effect either by the difference-in-coefficients method (τ − γ) or the product-of-coefficients method (α × β), where α and β are the regression coefficients from (X → M) and the effect of the mediator on the outcome adjusted for the exposure (M → Y), respectively. When the outcome and mediator are continuous, the two methods coincide.²⁵ However, for a binary outcome, the difference method results in biased indirect effects due to the noncollapsibility of logistic regression, and therefore the product method is preferable.^26,27 Furthermore, with a binary outcome, the traditional approach results in effect estimates that correspond to certain causal estimands under the assumptions of no unmeasured confounding, no interaction, and that the outcome is rare.^26,28 The extension of this correspondence between traditional and causal estimation in the time-varying mediation setting is still developing.^26,28,29

Recently Cai et al.³⁰ extended the traditional mediation approach³¹ to a time-varying mediation approach using a varying-coefficient model³² and local smoothing³³ for the continuous outcome setting. Our study applies the same estimation strategy as Cai et al.³⁰ to a binary outcome setting and explores the dynamic pathways of the smoking cessation process with this new analytical method. Based on the method described by Cai et al.,³⁰ we estimate the time-varying direct and indirect effects of pharmacotherapies on daily smoking status using the following two equations:

$${\displaystyle \begin{array}{l}{\displaystyle M_i(t_{i,j})={\alpha }_0(t_{i,j})+\alpha (t_{i,j})X_i+{\epsilon }_i^M(t_{i,j})i=1,\dots ,1031;j=0,\dots ,42}\end{array}}$$

(1)

$${\displaystyle \begin{array}{l}{\displaystyle Y^{\ast }(t_{i,j})={\beta }_0(t_{i,j})+\gamma (t_{i,j})X_i+\beta (t_{i,j})M_i(t_{i,j-1})}\end{array}}$$

(2)

where X_i is a binary indicator of the exposure group subject i was randomized to, M_i(t_i,j) is the value of the mediator for subject i at timepoint j and $Y^{\ast }(t_{i,j})=\mathrm{l}\mathrm{n}\left(\frac{P(Y(t_{i,j})=1)}{1-P(Y(t_{i,j})=1)}\right)$ is the log-odds of the observed outcome Y(t_i,j) for subject i at timepoint j. α(t_i,j) corresponds to the unit change in the mediator produced by the comparator group, ie, X_i = 1 at timepoint j. β(t_i,j) corresponds to the log-odds of smoking for a unit increase in M_i(t_i,j−1) at timepoint j. γ(t_i,j) is the direct effect of exposure on the log-odds of smoking. Thus, the time-varying mediation effect on the log-odds scale, using product-of-coefficients approach is given by,

$${\displaystyle \begin{array}{l}{\displaystyle I{E}_j(t_{i,j})=\alpha (t_{i,j-1})\times \beta (t_{i,j})}\end{array}}$$

(3)

The intervention of interest in this study is the pharmacotherapy treatment arm subjects are randomized to at the beginning of the study and is time invariant. The outcome is time-varying smoking status (binary—Yes/No). We consider three continuous time-varying mediators, “craving,” “cessation fatigue,” and “negative mood,” to test our hypotheses of time-varying mediation effects of the pharmacotherapies on smoking status.

In our empirical data, the responses to the mediators and outcome were recorded simultaneously. Thus, it is difficult to assert whether the outcome occurred after the mediator or vice versa at the time of a prompt. To preserve the temporal sequence of the variables of interest and draw inferences accordingly, we model the effect of the mediator recorded at t_i,j−1 on the outcome at t_i,j and estimate the indirect effect of symptoms experienced at a past timepoint on smoking status at the immediate future timepoint. We selected the nicotine patch treatment arm as the reference group and estimated the time-varying mediation effect of varenicline compared with nicotine patch and cNRT compared with nicotine patch on smoking status. Three separate sets of models were fitted to estimate the time-varying effects of the three mediators of interest. The estimation strategy follows the two-step approach^30,33—first, we estimate α(t_i,j), standardized β(t_i,j), standardized γ(t_i,j) and IE_j(t_i,j) from equations (1)–(3), followed by smoothing the resulting estimates using the locally weighted smoothing (LOESS) technique and exponentiating³⁴ the standardized β(t_i,j), and IE_j(t_i,j) estimates. Since our mediator(s) of interest are continuous, α(t_i,j) was not standardized.²⁶ However, to equalize the coefficient scales across the logistic regression models, β(t_i,j) and γ(t_i,j) were standardized using y-standardization.²⁶ We also test the statistical significance of the estimated time-varying mediation effect, using the percentile bootstrapping approach³⁰ using 1000 bootstrapped samples to construct 95% point-wise confidence bands. All data cleaning and analyses were conducted using R version 4.0.5 (2021-03-31) and the R package “tvmediation—v 1.0.0.”³⁵

Results

Fifty-five out of the 1086 participants (5%) were nonresponsive on all EMA prompts and were excluded from all analyses. Data from 1031 subjects (52.6% women, 67.8% White) were analyzed, of which 21.8% received nicotine patch (n = 225), 38.6% received varenicline (n = 398), and 39.6% received cNRT (n = 408). The proportion of participants in the three treatment arms mimic the sampling strategy implemented in the parent study. Additionally, the three groups are consistent with respect to the baseline characteristics, including nicotine dependence assessed via the Fagerstrom Test of Cigarette Dependence³⁶ (Table 1). Thus, the possibility of exposure–mediator or exposure–outcome confounding due to measured pretreatment covariates is negligible due to randomization. The abstinence rates among the three treatment groups at 4 weeks post-TQD were recomputed for this study sample and were similar across the treatment groups (Table 1) and slightly higher than the parent study estimates (nicotine patch: 32.8%, varenicline: 35.9%, cNRT: 35.6%).²³

Table 1.

Participant Demographic Distribution

Variable	Nicotine patch	Varenicline	Combination nicotine replacement therapy
	N = 225 (21.8 %)	N = 398 (38.6 %)	N = 408 (39.6%)
Gender, N (%)
Female	118 (52.4%)	212 (53.3%)	212 (52%)
Male	107 (47.6%)	184 (46.2%)	195 (47.8%)
Missing	—	2 (0.5%)	1 (0.2%)
Race, N (%)
White or Caucasian	153 (68%)	267 (67.1%)	279 (68.4%)
Native American/Alaska Native	2 (0.8%)	1 (0.3%)	3 (0.7%)
Black or African American	62 (27.6%)	112 (28.1%)	112 (27.5%)
Asian	1 (0.4%)	—	2 (0.5%)
Other	1 (0.4%)	8 (2%)	7 (1.7%)
Multiracial	6 (2.7%)	10 (2.5%)	5 (1.2%)
Ethnicity, N (%)
Hispanic or Latino	4 (1.7%)	8 (2%)	6 (1.5%)
Not Hispanic/Latino	206 (91.6%)	360 (90.5%)	374 (91.7%)
Missing	15 (6.7%)	30 (7.5%)	28 (6.9%)
Biochemically confirmed abstinence (CO <6 ppm) rate, N (%)
At 4 weeks post-quit	77 (34.2%)	144 (36.4%)	148 (36.3%)
Missing	8 (3.6%)	12 (3%)	11 (2.7%)
Age
Mean (SD)	49.19 (10.97)	46.98 (11.48)	48.43 (11.95)
Median (first, third quartiles)	51 (43, 57)	49 (38, 55)	51 (40, 57)
FTNDa score
Mean (SD)	4.89 (2.24)	4.82 (2.10)	4.77 (2.01)
Median (first, third quartiles)	5 (3, 7)	5 (3, 6)	5 (4, 6)
Missing	—	3	2

Fagerstrom Test for Nicotine Dependence.

On close examination of the smoking patterns across the three treatment arms (see Supplementary Figures 1a–3c), we found no increasing or decreasing pattern in the proportion of individuals responding as having smoked since waking up from days 0 to 28 post-TQD, in the nicotine patch arm. However, there is a decreasing trend in the proportion of individuals who smoked since waking up in the varenicline and cNRT arms. The proportion of individuals who reported smoking at the evening prompt shows no particular trend from days 0 to 28 post-TQD and is relatively higher than the observed values for the morning prompt. Overall, the outcome prevalences at most of the timepoints for each treatment arm and across the different mediator responses were <10%, and thus the rare outcome assumption is met. Therefore, our odds ratio estimates are collapsible (marginal estimates approximately equal the weighted average of stratum-specific estimates) and can be interpreted as relative risks. For mediation analysis, the rare outcome assumption must be evaluated across strata defined by both the exposure and the mediator.³⁷ Although we evaluated the rare outcome assumption, the no confounding and the no interaction assumptions were not assessed but rather assumed to be true. Therefore, as a measure of caution, instead of causally interpreting the indirect (mediation) effect estimates, the bootstrapped confidence intervals were used to test the presence of the mediation effect.²⁶

Mediation via Craving

Figure 1 shows the effects of varenicline and cNRT on daily smoking status, mediated via craving. The first column of Figure 1 shows that compared with the reference group (nicotine patch), the odds of smoking during the first 4 weeks post-quit in the varenicline group mediated via craving are lower and ranges from 0.96 to 0.93. At day 0 post-TQD, the mediated effect is approximately 0.96 (95% CI: 0.92, 0.99), which means that the odds of reported smoking at the evening prompt, decreases by 4% among individuals in the varenicline arm compared with nicotine patch, via reduced craving. This decrease in the risk of smoking via reduced craving declines steadily in the first week and continues to decrease with intermediate increasing trends in early weeks 2 and 3. The time-varying mediated effect is statistically significant throughout the 4 weeks, with the 95% point-wise confidence bands not including 1. Focusing on the individual paths of this mediated effect, we see from the plot representing α(t_i,j−1) denoted by “α-path” that compared with the reference nicotine patch group, craving decreases steeply in the varenicline group over the first week and stabilizes post mid-week 2 after an increasing trend early in the second week. At day 0 post-TQD users of varenicline report approximately a 0.41-point decrease in craving at the morning prompt compared with nicotine patch. The plot for exponentiated β(t_i,j) denoted by “β-path” shows that holding constant the effect of the treatment group, the odds of smoking increases with a unit increase in craving throughout the first 3 weeks and stabilizes in week 4. At day 0 post-TQD, the odds of reported smoking increases by ~11% for every unit increase in craving. The odds of smoking due to increasing craving has an increasing trend across the first 4 weeks of a quit attempt. Both “α-path” and “β-path,” ie, the time-varying treatment effect on craving and the time-varying effect of craving on smoking status adjusted for the treatment group, are statistically significant throughout the study period.

Figure 1.

Mediation effect of craving and α- and β-path effects with 95% CI estimated for the binary outcome (smoking status—Yes/No). The left column of the figure shows results of mediation analysis for varenicline vs. nicotine patch and the right column shows results of mediation analysis for cNRT vs. nicotine patch. The top panels show the time-varying mediation effect estimates at each timepoint of interest. The y-axis is on the odds scale. The middle panel shows the time-varying effect of varenicline/cNRT on craving compared with the patch. The y-axis is on the linear scale. The bottom panel shows the effect of craving on smoking status when the treatment is held constant. The y-axis is on the odds scale. CI = confidence interval, cNRT = combination nicotine replacement therapy.

The second column in Figure 1 presents the comparison between the cNRT and nicotine patch groups. The time-varying pattern for the mediated effect, “α-path” and “β-path” curves observed in the cNRT group is similar to the varenicline group. However, early in week 2, the mediated effect and the treatment effect on craving are not statistically significant (ie, the 95% confidence intervals include 1). The time-varying effect of craving on smoking status adjusted for the treatment group is statistically significant throughout the study period.

Mediation via Negative Mood and Cessation Fatigue

Supplementary Figures 4 and 5 show the effects of varenicline and cNRT on the daily smoking status, mediated via negative mood and cessation fatigue, respectively. The treatment effects of cNRT and varenicline vs. nicotine patch on smoking status mediated via negative mood and cessation fatigue are not statistically significant throughout the 4 weeks. Descriptively, compared with the reference group, the odds of smoking in the varenicline and cNRT groups mediated via negative mood are lower until mid-week 2 and stabilize after an intermediate increase. This effect, though not statistically significant, appears to be driven by treatment suppression of negative affect in the first few weeks (α path). Descriptively, the odds of smoking in the varenicline group mediated via cessation fatigue are higher until mid-week 2 and stabilize after an intermediate decrease, suggesting that fatigue is more likely to drive smoking in the varenicline group compared with the reference group. Compared with the reference group, the odds of smoking in the cNRT group mediated via cessation fatigue are lower and remain stable throughout the 4 weeks, although the mediation effect is not statistically significant.

Discussion

The main goal of this study was to identify “how,” “when,” and “why” the associations among the variables of interest change during the quit attempt using EMA data to better inform therapeutic approaches. The time-varying mediation analysis in this paper addresses the first two components, ie, the “how” and “when,” using a localized estimation strategy. It has the advantage of capturing the nonlinearity and time-varying characteristic of the mediation effect. In addition, because the distribution of the time-varying mediation effect is unknown and not necessarily normal, a nonparametric approach (percentile bootstrap) was used to estimate 95% point-wise confidence bands and test the statistical significance of the effect estimates.

The findings from this study support prior findings that the effect of varenicline on smoking status is significantly mediated via craving reduction.^38,39 A closer look at the individual components (α- and β-paths) of the mediated effect explain this finding further—compared with nicotine patch, individuals in the varenicline group reported declining levels of craving (α-path) and, craving remains a key predictor of smoking status, since, with increasing craving, the odds of smoking increases with the treatment effect held constant (β-path). Thus, the findings suggest that in the first week post-TQD, individuals in the varenicline group were less likely to report smoking because they had reduced cravings compared with those in the nicotine patch group. However, the effect of varenicline on craving alleviation plateaus beyond the initial post-TQD period. This reduction in craving alleviation is commensurate with earlier findings suggesting that beyond the initial post-quit period, as self-efficacy^21,40 may dwindle, additional support is required to address cravings, which continue to present a relapse risk.

The effect of varenicline, a nonnicotine partial agonist, on craving was not replicated with cNRT. cNRT suppresses craving significantly in the first week post-TQD and thereby reduces odds of smoking; however, this mediated effect of cNRT is not significantly different than nicotine patch beyond day 8 post-TQD. Although the effects of cNRT on craving, and smoking status mediated via craving, are not statistically different from nicotine patch week 1 post-TQD, findings suggest that craving is still a key predictor of smoking status throughout the 4 weeks. Essentially, compared with the nicotine patch, cNRT offers no added benefit of craving reduction beyond week 1 post-TQD (α-path), but with increasing craving, individuals’ risk of reverting to smoking also increases (β-path) regardless of treatment condition. Therefore, irrespective of whether it is a predictor or a mediator, craving is a crucial treatment target, and monitoring sudden peaks and declines in craving, especially week 1 post-TQD onwards, may identify people at increased risk for relapse and an ideal time for a just-in-time intervention, especially for those on nicotine patch and cNRT.

We did not observe significantly different mediated effects through negative mood and cessation fatigue for both varenicline and cNRT, compared with the nicotine patch. This illustrates the key treatment challenge clinicians, researchers, and people who smoke face: we can identify risk factors for relapse (ie, craving, negative affect, cessation fatigue), however even the most efficacious pharmacotherapies do not differentially reduce these risks.

Extant literature documents associations among negative mood, cessation fatigue, and craving extensively.^15–19 In this study, participants who report smoking report less negative mood (see Supplementary Figures 1b, 2b, and 3b) and cessation fatigue (Supplementary Figures 1c, 2c, and 3c). These consistent patterns of results and their interrelations suggest that a multiple mediator model could be explored in future studies. In addition, the temporal order of these constructs is ambiguous. Negative mood can trigger craving and vice versa, illustrating another area for future research. Finally, these results could be due to measurement error due to the single-item assessments of negative mood and cessation fatigue. Further evaluations of these constructs must be conducted with better measurement.

We address the third component, “why” (ie, causality), partially. By modeling the effect of symptoms observed at the immediate past prompt on the smoking status recorded at the current prompt, we retained the temporal sequence of the variables of interest. Under the potential outcomes framework, defining the indirect effects as “causal” requires the strong assumptions of no unmeasured confounding of the different treatment → mediator → outcome pathways. Because the pretreatment covariate distributions were similar across the exposure groups, the parent study was a randomized controlled trial, and the participants were motivated to quit smoking, we can assume that there is no treatment → mediator or treatment → outcome confounding. In addition, there must not be post-treatment confounders of the mediator → outcome pathway (ie, confounders of the mediator and outcome that have themselves been influenced by the treatment). Obviously, prior smoking status is a potential post-treatment confounder of symptoms and current smoking status. To examine sensitivity to post-treatment confounding, we modeled the effect of treatment on symptoms mediated via smoking status at the immediate past prompt and found that the time-varying effect was not statistically significant. We also found that prior smoking status (smoking status reported at time “t − 1”) does not mediate the treatment effect on current smoking status (smoking status reported at time “t”), however, it is a strong predictor of current smoking status. Thus, we cannot negate the possibility of post-treatment time-varying confounding of the mediator → outcome path by the outcome measured at previous timepoint. The coefficient scales of the logistic regression model change with the inclusion of additional covariates in the model, irrespective of the strength of association between the covariates and the outcome. Including past smoking status in the logistic regression model of the current smoking status will lead to a conflation of the total variance and result in a biased mediation effect estimate, especially when the two smoking statuses are highly related to one another. Thus, we refrained from adjusting for past smoking status in the primary analysis.

Using time-varying mediation models, we found additional support for the notion that craving suppression is related to cessation success, especially early in the quit attempt and that this is one pathway via which varenicline supports smoking cessation. Defining the effect estimates from this study as causal estimands under the potential outcomes framework requires further research. Future research is also needed to assess the causal time-varying effect of single and multiple correlated mediators. Despite this limitation, these findings provide insight into the smoking cessation process and illustrate the importance of targeting craving reduction early in the quit attempt and the need for developing treatments that can address other risks for relapse (eg, negative mood, cessation fatigue) through the estimation of time-varying mediation effects.

Supplementary Material

A Contributorship Form detailing each author’s specific involvement with this content, as well as any supplementary data, are available online at https://academic.oup.com/ntr.

Funding

This work was supported by the National Institutes of Health grant R01CA229542-01 from the National Cancer Institute and the National Institutes of Health’s Office of Behavioral and Social Science Research.

Declaration of Interests

None declared.

Acknowledgments

We gratefully acknowledge all funding listed above. The views expressed are those of the authors and not necessarily those of the funders.

Data Availability

The data used in this article were provided by the University of Wisconsin Center for Tobacco Research and Intervention (UW-CTRI) by permission. Data will be shared on reasonable request to the corresponding author with permission of UW-CTRI. The data analysis code is available in Supplementary Materials.