|Home | About | Journals | Submit | Contact Us | Français|
The impact of lung cancer screening on smoking behavior is unclear. The aims of this ancillary study of the Prostate Lung Colorectal and Ovarian Cancer Screening Trial were to produce risk prediction models to identify individuals at risk of relapse or continued smoking and to evaluate whether cancer-screening variables affect long-term smoking outcomes.
Participants completed a baseline questionnaire at trial enrollment and a supplemental questionnaire 4–14 years after enrollment, which assessed several cancer-related variables, including family history of cancer, comorbidities, and tobacco use. Multivariable logistic regression models were used to predict smoking status at completion of the supplemental questionnaire. The models’ predictive performances were evaluated by assessing discrimination via the receiver operator characteristic area under the curve (ROC AUC) and calibration. Models were internally validated using bootstrap methods.
Of the 31 694 former smokers on the baseline questionnaire, 1042 (3.3%) had relapsed (ie, reported being a current smoker on the supplemental questionnaire). Of the 6807 current smokers on the baseline questionnaire, 4439 (65.2%) reported continued smoking on the supplemental questionnaire. Relapse was associated with multiple demographic, medical, and tobacco-related characteristics. This model had a bootstrap median ROC AUC of 0.862 (95% confidence interval [CI] = 0.858 to 0.866) and a calibration slope of 1.004 (95% CI = 0.978 to 1.029), indicating excellent discrimination and calibration. Predictors of continued smoking also included multiple demographic, medical, and tobacco-related characteristics. This model had an ROC AUC of 0.611 (95% CI = 0.605 to 0.614) and a slope of 1.006 (95% CI = 0.962 to 1.041), indicating modest discrimination. Neither the trial arm nor the lung-screening result was statistically significantly associated with smoking outcomes.
These models, if validated externally, may have public health utility in identifying individuals at risk for adverse smoking outcomes, who may benefit from relapse prevention and smoking cessation interventions.
A total of 226 160 new cases of lung cancer and 160 340 lung cancer deaths are expected in the United States in 2012 (1). Although smoking abstinence is the best way to reduce lung cancer mortality, early detection and treatment of lung cancer may also reduce mortality from this disease (2). Two randomized US cancer-screening trials—the Prostate Lung Colorectal and Ovarian Cancer Screening Trial [PLCO (3)] and the National Lung Screening Trial [NLST (4)]—have assessed whether lung screening reduces lung cancer mortality. Results of the PLCO indicated no difference in lung cancer mortality between subjects who received annual chest x-ray and those who received usual care (5). By contrast, the NLST results indicated a 20% mortality reduction in the spiral computed tomography arm compared with the chest x-ray arm (4).
Participants in these two screening trials provide an important source of data for understanding smoking-related behaviors among current and former smokers who are not seeking a tobacco-related intervention. In addition, lung cancer screening may provide a critical “teachable moment (6)” for altering smoking behaviors. Teachable moments are defined as naturally occurring events, such as screening tests, hospitalizations, pregnancies, or new illnesses, that “motivate individuals to spontaneously adopt risk-reducing health behaviors” (6). In short term (ie, less than 3 years) smoking outcome studies (7–9), current and former smokers who received a false-positive lung-screening result had improved smoking outcomes compared with those who received a negative screening result. Additional factors associated with smoking outcomes among current smokers in these screening studies included healthier lung function, lower cancer-related anxiety, and a lower perceived benefit of quitting (7–10). Factors associated with relapse among the former smokers included shorter duration of abstinence and a longer smoking history (9,10).
In addition to these studies, three randomized lung cancer–screening trials (11–13) and one single-arm trial (14) have also reported the effect of lung cancer screening on the smoking behavior of participants. In the 1-year follow-up of the Danish Lung Cancer Screening trial comparing annual low-dose computerized tomography screening vs no screening (11), a positive screening result was associated with an increase in smoking cessation and a decrease in smoking relapse. However, there were no differences between the screening and control arms in terms of quit rates (11.9% and 11.8%, respectively) or relapse rates (10.0% and 10.5%, respectively). At the 2-year follow-up in the Dutch–Belgian trial (13), screened participants were statistically significantly less likely than the control (no screening) participants to quit (14.5% vs 19.1%, respectively; P = .04). However, the overall quit rate of 16.6% suggested that the trial participants were highly motivated to quit smoking. There were no differences in relapse rates between the two arms of the Dutch–Belgian trial, nor was there a difference in the prolonged abstinence rate between those with a false-positive screening result and those with a negative result (15). The Mayo Lung Project (12) reported no difference in the quit rates between the screening (chest x-ray) and usual care arms at the 1-year follow-up (overall quit rate was approximately 15%). However, the impact of the screening result on smoking outcomes was not reported (12). Each of these trials reported that several tobacco-related and demographic factors (eg, pack-years of smoking, nicotine dependence, motivation to quit, education, race or ethnicity, and age) predicted continued smoking among current smokers (11–13) and that pack-years, abstinence duration, and age predicted relapse among former smokers (11). Finally, in a single-arm study, the Early Lung Cancer Action Program (ELCAP) investigated the short- and long-term impacts of computerized tomography screening on smoking behavior (14). Among subjects who were current smokers at baseline, those with a negative screening result had a statistically significant 28% higher likelihood of continued smoking at follow-up (1–72 months after baseline) compared with those with a positive screening result [P < .004 (14)]. This difference was no longer statistically significant at the 6-year follow-up. Screening results were not associated with relapse rates at the short- or long-term follow-up.
To further examine factors associated with tobacco use in the context of cancer screening, we focused on two study aims. First, we developed two prediction models to identify individuals at high risk for two adverse smoking outcomes: relapse among former smokers and continued smoking among current smokers. To our knowledge, no prediction models of tobacco use have been applied in the context of cancer screening. In fact, to our knowledge, only the prior prediction model has been used to identify those at high risk for continued smoking (16). This model indicated that among a population-based sample of California smokers, demographic and tobacco-related predictors resulted in a receiver operator characteristic area under the curve (ROC AUC) of 0.693, indicating moderate predictive ability at 1–2 years of follow-up (16). We have extended this work by using a substantially longer follow-up and a broader set of predictors than those included in previous cancer-screening studies.
Second, we evaluated associations between screening- and trial-related variables and adverse smoking outcomes. Studies of whether lung cancer screening alters smoking behaviors have produced mixed results; we sought to assess the long-term impact of screening and screening results on smoking outcomes. We expected that receipt of a false-positive lung-screening result would be associated with an increased likelihood of quitting smoking and a lower likelihood of relapse compared with receipt of negative screening results. This is an important question to address, given that many individuals are likely to undergo lung cancer screening following the recent publication of the NLST results (4). As a result, even a small improvement in cessation and relapse prevention following screening could have an important public health impact.
The PLCO is a multisite, randomized controlled trial designed to assess whether cancer screening reduces disease-specific mortality for prostate, lung, colorectal, and ovarian cancers (3). The 10 screening centers of the PLCO are located at Aurora, CO; Washington, DC; Honolulu, HI; Detroit, MI; Minneapolis, MN; St Louis, MO; Pittsburgh, PA; Salt Lake City, UT; Marshfield, WI; and Birmingham, AL. At enrollment, trial participants (N = 154 907) were men and women between the ages of 55 and 74 years with no history of prostate, lung, colorectal, or ovarian cancer. Subjects were accrued from November 1993 through June 2001, randomly assigned to the screening or control arm, and were assessed for up to 14 years. Participants assigned to the screening arm received a screening examination at enrollment (baseline) and subsequent screenings for all cancers for which they were eligible. Screening arm participants received a maximum of four annual chest x-rays. Participants assigned to the control arm did not receive any screening examinations as part of the study and were instructed to follow their usual health-care routine.
A baseline questionnaire was administered to all participants at trial enrollment. During 2006–2007, a supplemental questionnaire was mailed to all PLCO participants who were thought to be alive at the time. The baseline and supplemental questionnaires assessed several cancer-related topics, including family history of cancer, comorbidities, and tobacco use. The median follow-up (ie, interval between the baseline and supplemental questionnaires) was 8.5 years (range = 4–14 years).
Figure 1 presents the flow of participants in the trial. Participants excluded from this analysis were those who were missing the baseline or supplemental questionnaire, had never smoked regularly, died before completing the supplemental questionnaire, withdrew from the trial, or had inconsistent information on the tobacco items (eg, self-identified as a current smoker on the baseline questionnaire and as a never smoker on the supplemental questionnaire). To prevent the premature release of PLCO outcomes, we excluded participants who were diagnosed with any type of cancer between completion of the baseline and supplemental questionnaires from these analyses (57% of the cancers diagnosed were prostate, lung, colorectal, or ovarian cancers).
Of those who completed the baseline questionnaire and identified themselves as ever smokers (40 117 control arm participants and 40 635 screening arm participants), approximately two-thirds returned the mailed supplemental questionnaire (25 690 control arm participants [64.0%] and 26 898 screening arm participants [66.2%]). The final sample of current and former smokers included 18 791 in the control arm and 19 710 in the screening arm. Tobacco use groups were defined based on the combined smoking status as reported on the baseline and supplemental questionnaires: current smoker on both questionnaires, current smoker on the baseline questionnaire and former smoker on the supplemental questionnaire, former smoker on both questionnaires, or former smoker on the baseline questionnaire and current on the supplemental questionnaire (Figure 1). All participants provided written informed consent for participation in the PLCO trial. Institutional review board approval was obtained from all 10 screening centers.
Demographics and Medical History. On the baseline questionnaire, participants provided information on their age, sex, race or ethnicity, education level, marital status, employment status, weight, and height. On the supplemental questionnaire, participants indicated their family history of lung cancer and current family income (categorized into five levels: level 1, <$20 000; level 2, $20 000–$49 000; level 3, $50 000–$99 000; level 4, $100 000–$200 000; level 5, >$200 000).
Comorbidities. On the baseline questionnaire, participants indicated their previously diagnosed illnesses from a list of 21 conditions. On the supplemental questionnaire, participants again indicated all previously diagnosed illnesses to date from the same list of 21 conditions. Newly diagnosed tobacco-related diseases, including emphysema, coronary heart disease, heart attack, and stroke, were defined as those indicated on the supplemental questionnaire but not on the baseline questionnaire.
Tobacco Use Variables. On the baseline questionnaire, participants indicated the age they started smoking regularly (no additional information was available regarding the definition of “regularly”); the number of years they had smoked and the number of cigarettes smoked per day, and the number of pack-years [packs per day multiplied by years smoked]); their use of other tobacco products (ie, pipes or cigars); and the type of cigarettes smoked (filtered, unfiltered, or no usual type). Former smokers indicated their age when they had last quit smoking. From the supplemental questionnaire, we obtained the type of cigarette smoked (ultralight, light, regular, or no usual type, and menthol, non-menthol, or no usual type) and information on secondhand smoke exposure (none, moderate, or heavy) during childhood, adulthood at home, and adulthood at work. We calculated a secondhand smoke exposure score by summing the exposure level (none = 0, moderate = 1, and heavy = 2) during childhood, adulthood at home, and adulthood at work. Participants who indicated they were current smokers on the supplemental questionnaire reported the number of cigarettes smoked per day over the previous 30 days, whether they smoked some days or every day during the previous 30 days, intention to quit [within next 30 days, within next 6 months, or not thinking of quitting in the next 6 months (17)], and their nicotine dependence level based on the 6-item Fagerström Test for Nicotine Dependence [FTND (18)]. These dependence variables were not assessed among former smokers.
Smoking Status. The baseline questionnaire defined former smokers as those who had ever smoked regularly for at least 6 months but did not currently smoke cigarettes regularly. The supplemental questionnaire defined former smokers as those who had smoked 100 cigarettes or more in their lifetime but had not smoked during the previous 30 days. The baseline questionnaire defined current smokers as those who had smoked cigarettes regularly for at least 6 months and who currently smoked regularly. The supplemental questionnaire defined current smokers as those who had smoked 100 cigarettes or more in their lifetime and had smoked during the previous 30 days. The definitions of former and current smokers differed on the baseline and supplemental questionnaires because a decision was made by PLCO investigators when constructing the supplemental questionnaire to improve the measurement of smoking status, which meant that the two assessments could not be identical.
For baseline-questionnaire former smokers who relapsed (ie, those who self-reported as a current smoker on the supplemental questionnaire) and for baseline-questionnaire current smokers who quit (ie, those who self-reported as a former smoker on the supplemental questionnaire), the supplemental questionnaire assessed the participant’s age, but not the precise date, on which the change in smoking status occurred. Thus, the time-to-event data are the number of years smoked and the number of years since quitting, based on participant’s age when they completed the supplemental questionnaire (ie, time-to-event data were not available in months).
Trial-Related Variables. Trial-related variables included trial randomization (screening vs control), number of years in the trial (calculated as the date of supplemental questionnaire completion minus date of baseline questionnaire completion), and PLCO screening site (Colorado, Washington DC, Hawaii, Michigan, Minnesota, Missouri, Pennsylvania, Utah, Wisconsin, and Alabama). We defined a false-positive screening result as a positive screening result at one or more of the four annual chest x-rays that was not followed by a lung cancer diagnosis within 3 years (19). However, a positive screening result was necessarily a false-positive result because all trial participants who were diagnosed with any form of cancer had been removed from this dataset. Although it is possible that in some instances, the false-positive screening result may have occurred after an individual had changed his or her smoking behavior, the absence of data prevents us from estimating the frequency of such occurrences.
Because the univariate associations between adverse smoking outcomes (relapse and continued smoking) and demographic, medical, and trial- and tobacco-related variables were similar across study arms (data not shown), we used data pooled from both trial arms for subsequent analyses. Differences in distributions were based on t tests for continuous variables (age, body mass index, age started smoking, number of years smoked, number of cigarettes smoked per day, number of pack-years, secondhand smoke exposure score, and number of years in trial) and on χ2 tests of trend for categorical variables (sex, race, education level, marital status, family income, number of comorbidities, number of new tobacco-related diseases, family history of lung cancer, cigarette type (filtered or unfiltered cigarettes, pipes or cigars), trial arm, number of false-positive x-rays). Cut points for income, comorbidities, new tobacco-related diseases, and false-positive screening tests were based on the distributions of the variables. All statistical tests were two sided.
Data were missing for fewer than 5% of participants for all study variables except income, which was missing for 17.5% of participants. Missing income data were imputed using 10 imputation sets and predictive mean matching (20) with the use of the Stata Multiple Imputation package (21). All predictor and outcome variables in the final models were used in the imputation models. Predictor variable coefficients, SEs, and summary statistics from the multiple imputation sets were pooled according to the method of Little and Rubin (22).
Prediction Models. Logistic regression models were prepared to predict 1) relapse among participants who were former smokers on the baseline questionnaire and 2) continued smoking among participants who were current smokers on the baseline questionnaire. A priori selection of demographic, medical, or tobacco-related predictor variables thought to be related to smoking behavior was based on the literature or expert opinion (7–10). To reduce model complexity, we excluded variables with a P value greater than .20 from the final models.
We evaluated nonlinear effects of continuous variables with restricted cubic splines using four knots and three splines (23). Knot placement was based on the percentile distribution of data to ensure adequate coverage of the data (23,24). For smoking duration, knots were placed at 6, 20, 33, and 47 years, and for smoking quit time, knots were placed at 2.0, 14.1, 25.5, and 38.1 years. Interactions of selected predictors in the final models were evaluated by inclusion of interaction and main effects terms. To ensure model simplicity, only a priori interactions with P values less than .05 were considered for model inclusion. No such interactions were observed in this study. In addition, we assessed the model’s ability to discriminate (classify correctly) current smokers from former smokers using the ROC AUC (24). Model calibration, which determines whether model-predicted probabilities match observed probabilities, was assessed by evaluating how much the intercept and slope of the calibration line deviated from the ideal values of 0 and 1, respectively, when the predicted probabilities were plotted vs the observed probabilities. Model building attempted to optimize the aforementioned performance statistics. Internal validation of models was carried out by applying the final prediction models to 200 bootstrap samples and evaluating the bootstrap medians and 95% confidence intervals (CIs) for the ROC AUC and calibration line intercepts and slopes over the 200 samples. Models, statistics, and figures were prepared using Stata/MP 12.1 (25).
Association Models Testing Trial-Related Variables. We prepared two multivariable association models using logistic regression analyses for relapse and continued smoking. We included the demographic, medical, and tobacco-related variables and further added the trial-related variables, which included screening center, randomization group, time elapsed between the baseline and supplemental questionnaires, and the number of false-positive chest x-rays. The associations between screening variables and smoking behaviors were tested, with two-sided P values less than .05 considered statistically significant.
The distributions of demographic, medical, tobacco-related, and trial-related variables by the two adverse smoking outcomes (relapse and continued smoking) are presented in Table 1. Of the 31 694 self-reported former smokers on the baseline questionnaire, 1042 (3.3%) had relapsed (ie, self-reported as a current smoker on the supplemental questionnaire). Of the 6807 self-reported current smokers on the baseline questionnaire, 4439 (65.2%) reported continued smoking on the supplemental questionnaire. Overall, participants were approximately 60 years old, one-half were men, 90% were white, one-half had two or more comorbidities, the average age at the start of smoking was 18–19 years, and 10% had received one or more false-positive chest x-rays (Table 1). Furthermore, on most of the demographic, medical, and tobacco- and trial-related variables, former smokers who relapsed differed statistically significantly from the former smokers who remained abstinent, and current smokers who continued smoking differed statistically significantly from those who quit (Table 1).
Several nicotine dependence items were only assessed on the supplemental questionnaire among current smokers. Therefore, these variables could not be included in the prediction models and are described here. Substantial nicotine dependence was evident based on standard measures of dependence: 66.3% smoked within 30 minutes of waking, the average FTND score was 4.95 (SD = 2.5; range = 0–10), and 47.3% scored greater than 6 on the FTND (26). Regarding packs per day, 51.0% smoked less than 1 pack per day, 32.6% smoked 1 pack per day, and 16.4% smoked more than 1 pack per day. Concerning intention to quit, 18.7% of current smokers planned to stop smoking within 30 days, 36.5% planned to stop within 6 months, and 44.9% were not considering quitting within 6 months. This distribution of intention to quit is similar to that in US (27), European (28–31), and Canadian (32) population-based studies.
Table 2 presents the multivariable risk prediction model for relapse. The predictive performance of this model was evaluated in 200 bootstrap samples. The bootstrap median ROC AUC was 0.862 (95% CI = 0.858 to 0.866), and the bootstrap median calibration line intercept and slope were −0.0001 (95% CI = −0.0007 to 0.0004) and 1.004 (95% CI = 0.978 to 1.029), respectively. These predictive performance statistics support the conclusion that the relapse model has high discrimination and calibration. Relapse was more likely among those who were younger at completion of the baseline questionnaire, black or Hispanic (vs white), less educated, or unmarried, and those with a lower income, a lower BMI, or no family history of lung cancer. The following tobacco-related variables were associated with relapse: more secondhand smoke exposure, fewer cigarettes smoked per day, more pack-years, and smoking light or ultralight cigarettes (vs regular cigarettes or no usual type) or pipes or cigars (vs neither); relapse was also more likely among longer-term smokers and recent quitters. The latter two variables had statistically significant nonlinear associations with relapse, as depicted in Figures 2 and and33 (both P < .001). To facilitate estimation of individual risk using the relapse prediction model, we present the beta coefficients for predictor variables and the model intercept in Supplementary Table 1 (available online). These model parameters allowed us to calculate the probability of relapse during follow-up (median duration = 8.5 years), given an individual’s specific predictor values.
To assess whether this smoking relapse model was relevant beyond randomly assigned trial participants, who are typically more highly educated than the general population, we limited the analysis to trial participants with high school education or less (N = 11 306) and found that the ROC AUC was 0.857 and the calibration intercept and slope were −0.0002 and 0.982, respectively (Supplementary Table 2, available online). These findings suggest that the prediction model may be generalizable beyond the PLCO sample.
Table 2 also presents the multivariable associations between 1) the demographic, medical, and tobacco- and trial-related variables and 2) relapse. All of the demographic, medical, and tobacco-related predictors had odds ratios and confidence intervals that were virtually identical to those described in the prediction model. With regard to the trial-related variables, relapse was more likely among those who were enrolled in the trial for a shorter time (OR = 0.89, 95% CI = 0.85 to 0.93), but it was not associated with trial arm, screening center, or receipt of a false-positive chest x-ray.
Table 3 presents the multivariable risk prediction model for continued smoking. The predictive performance of the model was evaluated in 200 bootstrap samples. The bootstrap median ROC AUC was 0.611 (95% CI = 0.605 to 0.614) and the bootstrap median calibration line intercept and slope were −0.004 (95% CI = −0.027 to 0.025) and 1.006 (95% CI = 0.962 to 1.041), respectively. These predictive performance statistics indicate that this model has moderate ability to discriminate. Continued smoking was more likely among those who were younger at completion of the baseline questionnaire or black or Hispanic (vs white) and those with a lower income, lower BMI, or had no new tobacco-related diseases. The following tobacco-related variables were associated with continued smoking: smoking more cigarettes per day; smoking light or ultralight cigarettes (vs regular cigarettes or no usual type); smoking filtered cigarettes (vs unfiltered); and more secondhand smoke exposure. To facilitate estimation of individual risk using the continued smoking prediction model, we present the beta coefficients for predictor variables and the model intercept in Supplementary Table 3 (available online). We did not conduct an analysis limited to those with less than a high school education for the continued smoking model given the moderate prediction capabilities of the overall model.
Table 3 also presents the multivariable associations between the demographic, medical, and tobacco- and trial-related variables and continued smoking. Although the statistical significance level became non–statistically significant for two variables (age and pack-years), the remaining demographic, medical, and tobacco-related predictors had virtually identical odds ratios and confidence intervals as described in the prediction model. Continued smoking was statistically significantly more likely among those who were enrolled in the trial for a shorter period (OR =.85, 95% CI = 0.82 to 0.88) but was not statistically significantly associated with trial arm, screening center, or screening result.
The goals of this PLCO ancillary study were to develop and internally validate prediction models that identify former smokers at risk of relapse and current smokers at risk for continued smoking. Among the 3.3% of former smokers who relapsed, relapse was statistically significantly associated with multiple demographic, medical, and tobacco-related characteristics and indicated excellent discrimination and calibration. Among the 65.2% of smokers who continued smoking, predictors of continued smoking also included multiple demographic, medical, and tobacco-related characteristics and indicated modest discrimination and calibration. In addition, we tested two hypotheses regarding the impact of trial-related factors on smoking outcomes and found that receipt of a false-positive lung-screening result was not associated with an increased likelihood of quitting smoking or with a lower likelihood of relapse compared with those with negative screening results.
Relapse rates and abstinence rates in this study were similar to those reported in previous screening studies. Among former smokers who had quit before PLCO enrollment, 3.3% had relapsed by the follow-up assessment, which is comparable to the 2.0%–4.7% relapse rates reported by similarly defined former smokers (10,11,14). Among current smokers, 34.8% reported not smoking during the 30 days preceding the follow-up assessment, a rate that is comparable to similarly defined quit rates in other long-term lung cancer–screening studies: 35% at 6 years (14) and 24% at 3 years (9). Studies with shorter follow-ups reported lower quit rates: 16%–19% at 2 years (13,33) and 6%–16% at 1 year (11,34,35).
The relapse prediction model had excellent discrimination and calibration and suggested that relapse was more likely among those who were younger, less educated, unmarried, black or Hispanic, and those with a lower income. Relapse was also more likely among longer-term smokers, recent quitters, smokers of light or ultralight cigarettes, and pipe or cigar smokers. Almost identical results were obtained when the sample was limited to those with less education, suggesting that the prediction model may be relevant to those who are less educated than randomly assigned trial participants (who are typically more highly educated than the general population). Our model is similar to other association models of relapse (10,11,14); however, it also included several additional demographic, medical, and tobacco-related variables that other studies have not reported. This relapse prediction model may be useful for identifying former smokers who may benefit most from relapse prevention interventions.
The continued smoking prediction model had moderate discrimination and calibration and suggested that continued smoking was more likely among those who were younger, had lower income, and were black or Hispanic. Continued smoking was also more likely among heavier smokers, smokers of light or ultralight cigarettes, and those with greater secondhand smoke exposure. However, current smokers with higher body mass index or new tobacco-related diseases and smokers of unfiltered cigarettes were less likely to continue smoking. These characteristics may be useful in identifying smokers who are most in need of a smoking cessation intervention. This model is similar to the one other prediction model of smoking cessation tested in a population-based study of 2000 smokers in California [ROC AUC = 0.693 (16)]. The improved predictive performance in that model (vs the ROC AUC of 0.611 in our model) may be due to the shorter follow-up period and the inclusion of a detailed assessment of nicotine dependence in the California study (16). However, these two prediction models are not directly comparable because they come from different study populations with different distributions of predictor variables (36).
Contrary to our expectation, neither the PLCO trial arm nor the arm receiving a false-positive chest x-ray was associated with smoking outcomes, as has been observed in shorter-term studies of smoking outcomes (7–9,15). This negative finding concerning the false-positive x-ray has also been reported after 6 years of follow-up in the ELCAP (14), suggesting that screening results may have only a short-term effect on smoking behavior. However, we found that being diagnosed with a new noncancer, tobacco-related disease was inversely associated with continued smoking, suggesting another possible teachable moment for altering smoking behaviors: Upon receiving a new diagnosis, smokers may be motivated to quit smoking and become more amenable to formal cessation programs.
This study has several notable strengths. To our knowledge, this is the first study to develop and internally validate a prediction model for smoking relapse that has high predictive accuracy. The predictive ability of our continued smoking model was only moderate, but it too may have public health utility. These models can serve as a starting point for external validation in other populations. In addition, these models suggest important variables that should be considered in the development of effective intervention methods for long-term, heavily dependent smokers who are likely to be well represented in lung cancer–screening programs. Furthermore, this study reports the longest follow-up completed in a lung cancer–screening trial or in an observational study and calls into question whether the immediate impact of screening results (7,8,15) on smoking outcomes persist in the long term. Our findings emphasize the need for relapse prevention and cessation interventions to be available immediately following screening, when a subject’s motivation to remain abstinent or to stop smoking may be increased compared with the situation before screening. Given the eligibility criteria for undergoing lung cancer screening, interventions will need to be specifically designed for older, long-term smokers who are highly nicotine dependent. Finally, we replicated the findings of three other studies of smoking status that were conducted within a randomized screening trial, by showing that regular screening did not have an impact on adverse smoking outcomes compared with usual care (11–13,15).
There are several study limitations that should be considered. First, we were missing some potentially important data, including smoking status event times between completion of the baseline and supplemental questionnaires, a detailed baseline assessment of nicotine dependence, and repeated assessments of smoking behavior. The availability of these data may have improved the prediction models and allowed for a more detailed understanding of relapse and cessation in this study. Second, we did not have biochemical verification of smoking status. However, Studts et al. (37) reported 100% agreement between biochemically verified and self-reported smoking status in the context of a randomized controlled trial, and several studies (10,38,39) have reported no statistically significant difference in the percentage of smokers between the two measurement strategies. Third, because we excluded all participants who were diagnosed with any cancer between completion of the baseline and supplemental questionnaires (as required by the PLCO investigators), we were unable to evaluate associations between new cancers and smoking outcomes. Thus, these analyses are relevant for the majority of people who will undergo screening in clinical programs, because most will not be diagnosed with cancer as a result of the screening. Fourth, we excluded individuals who died after completing the baseline questionnaire but before completing a supplemental questionnaire because they lacked a follow-up assessment of smoking status. This exclusion, while unavoidable, may have resulted in a healthier sample. Finally, PLCO participants had, on average, a higher socioeconomic status than the general population, which may limit the generalizability of the findings (40,41). However, the predictive ability of the smoking relapse prediction model remained very high when we limited the model to PLCO participants who had only a high school education or less. In addition, the long-term, chronic smoking histories of PLCO participants are likely to be similar to the smoking histories of those who will be eligible to undergo clinically based lung cancer screening. Thus, our model may be of value in identifying individuals in the general population who are at high risk for relapse, which would allow for the application of targeted relapse prevention programs.
Directions for future research include multiple longitudinal assessments of the short- and long-term impacts of cancer screening on individuals’ smoking behaviors. In addition, an assessment of formal smoking cessation and relapse prevention interventions within lung cancer–screening programs is needed to examine whether it is possible to improve upon and sustain the short-term quit rates reported in previous studies (7,8,10,13,39). Whether an intervention can extend the immediate increase in cessation and interest in quitting observed in these prior studies is unknown. Given the 20% decline in lung cancer mortality reported by the NLST (4), it is likely that many individuals will undergo lung cancer screening in the near future. Thus, even a small percentage change in cessation and relapse prevention could have an important public health impact. Finally, additional research is needed to evaluate the impact of other events that occur during screening, such as the diagnosis of a new illness, on smoking behavior and how such factors may be used to promote cessation or to prevent relapse.
Our models, if validated externally, may have public health utility for identifying individuals who are at increased risk for adverse smoking outcomes in lung cancer–screening programs. Furthermore, these results may also be generalizable to the general population given that the relapse model limited to those who had completed high school or less demonstrated the same excellent predictive performance as observed in the entire PLCO population. Such information may allow the application of targeted smoking cessation and relapse prevention programs in an effort to reduce the negative health impact of long-term tobacco use among individuals who present for lung cancer screening and in the general population.
National Cancer Institute, the National Institutes of Health (PLCO Contract N01-CN-25522).
The funder did not have any involvement in the design of this ancillary study; the collection, analysis, and interpretation of the data; the writing of the manuscript; or the decision to submit the manuscript for publication.