Effectiveness and cost-effectiveness of repeated implementation intention formation on adolescent smoking initiation: A cluster randomized controlled trial.

Randomization and Masking

School was the unit of randomization. Schools were randomized by random number generator to intervention or control conditions on a 1:1 ratio by the trial statistician (RW). Randomization took place before recruitment of participants within each school. Due to the nature of the intervention, adolescents, teachers administering the intervention, heads of school, and data collection assistants were aware of group allocation. The trial statistician who conducted the analyses was initially blinded to condition.

Procedures

Self-reported data were collected at baseline plus 12, 24, 36, and 48 months postbaseline by research staff (present to answer questions) via questionnaire in groups (classes or year group assemblies) with adolescents requested not to confer. At each time point a smokerlyzer measure of breath carbon monoxide levels was taken individually with readings not available to adolescents.

The eight intervention sessions took place separately to data collection in classroom time (with each session containing approximately 26 adolescents) approximately every 6 months starting within 2 months of baseline data collection and were each led by a teacher. The content of sessions was designed to be matched (in relation to duration and frequency plus the use of written motivational materials and an implementation intention formation task) across the two conditions but focusing on smoking (intervention condition) or completing homework (control condition) as an unrelated behavior. Adolescents engaged with motivational materials (read antismoking messages or prohomework messages plus engaged in related tasks designed to increase engagement with the messages) and then completed implementation intentions sheets in relation to the target behavior (not smoking in intervention condition; completing homework in control condition). The target behavior in the control condition (completing homework) was selected to be a nonhealth related behavior appropriate for adolescents. The interventions were designed to run within a standard classroom session (50 min) with the majority (60%) of the time devoted to the messages.

Implementation intention formation was consistent across intervention sessions. Adolescents were first required to tick an option to indicate how they could refuse smoking this school term (“Tick ONE of the following things you could say if you were offered a cigarette or if you were tempted to smoke . . .; No thanks, smoking makes you smell awful; No, I do not want yellow teeth; No, I do not want to get addicted; No thanks, if you’re buying cigarettes you’re buying cancer; No it’s really bad for my asthma”). They were then requested to write in the selected response or generate a new response of their own to complete a statement (“If someone offers me a cigarette, then I will say . . .; e.g., No cancer sticks for me”). Adolescents were then required to indicate where they would not smoke (“Tick ALL the places where you will not smoke: I will not smoke at school; I will not smoke at home; I will not smoke at a party; I will not smoke with my friends; I will not smoke if I’m offered a cigarette”) and to respond to a question about smoking this school term (“I think I can make sure I do not smoke this term: yes, no”). The task was similar in the control condition but completed in relation to completing homework. Participants completed the implementation intention task individually by ticking boxes and writing down responses. The implementation intention sheets were collected in by the teacher and returned to the research team.

The motivational materials provided antismoking or prohomework messages and were paper based. The motivational materials were different in each session (i.e., eight sets of materials), were all judged to be age-appropriate by an experienced school teacher, and were similar in content to that used in our previous work ( Conner & Higgins, 2010 ; Higgins & Conner, 2003 ). For example, the first set of antismoking materials (“Smoking: It’s not worth it”) focused on 10 reasons not to smoke and included text and pictures along with a quiz designed to promote engagement with the materials. Full copies of the implementation intention sheets and motivational materials can be obtained from the first author.

Training sessions were run with teachers in each year of the study. These were 45-min sessions run in each school that focused on the broad purpose of the intervention and details of the intervention content (motivational messages and implementation intention sheets plus a plan of how to run the session). An opportunity to discuss the content and any potential problems with delivery was provided. The need to stick to the planned content and ensure all implementation intention sheets were fully completed was emphasized. A teacher in each school acted as a coordinator and monitored the delivery of all sessions and was available to answer teachers’ questions.

Outcome Measures

Four measures of smoking were used as outcomes at the 48-month follow-up. Self-reported cigarette use was assessed at each time point using a standardized measure ( Office for National Statistics, 1997 ); adolescents ticked one of: (a) I have never smoked; (b) I have only tried smoking once; (c) I used to smoke sometimes, but I never smoke cigarettes now; (d) I sometimes smoke cigarettes now, but I do not smoke as many as one a week; (e) I usually smoke between one and six cigarettes a week; (f) I usually smoke more than six cigarettes a week. This was used to create our first two measures of smoking: ever smoking (ticking response a coded 0; ticking responses b–f coded 1); regular smoking (ticking responses a–d coded 0; ticking responses e–f coded 1).

Any smoking (last 30 days) was assessed at 48-month postbaseline only (self-reported number of days in last 30 days using each of cigarettes, cigars, pipes, or sheesha/hookah was recorded and summed). Any smoking (last 30 days; 0 days coded 0; ≥1 day coded 1) was our third smoking measure.

Breath carbon monoxide (CO) levels (in parts per million; COppm) were assessed using the Micro + Smokerlyzer® CO Monitor (Bedfont Scientific Limited, Kent, United Kingdom) at each time point. However, the short half-life (four-six hours) of breath CO means that such measures are only reliable and valid for assessing recent cigarette smoking ( Bedfont, 2017 ; Jarvis, Tunstall-Pedoe, Feyerabend, Vesey, & Saloojee, 1987 ; Stookey, Katz, Olson, Drook, & Cohen, 1987 ). A variety of cut-offs have been used in the literature to indicate smoking in adults. We used the cut-off recommended by the device manufacturer as a clear indication of recent smoking in adolescents (≤6 ppm CO coded as 0; >6 ppm CO coded as 1). Breath CO >6 ppm was our fourth smoking outcome measure.

For the three smoking measures taken at each time point (ever smoking, regular smoking, breath CO >6 ppm) we also created measures of smoking across Time Points 2 to 5 (based on being categorized as smoking on a measure on at least one of the time points).

Other Measures

Other measures were assessed as covariates and/or moderators and measured at 48 months follow-up. At the school level we recorded geographical area (Leeds; Staffordshire) and size (number of pupils), and area level socioeconomic status (percentage of pupils in a school receiving free school meals; Croxford, 2000 ). At the individual level we assessed gender, ethnicity (self-reported classification dichotomized into non-White vs. White) and individual-level socioeconomic status (four-item Family Affluence Scale [FAS] scored 0–9 with higher scores indicating greater affluence; Boyce, Torsheim, Currie, & Zambon, 2006 ).

Fidelity checks assessed adherence, quality of delivery, and exposure to the intervention. The study coordinator in each school was requested to monitor adherence and provide feedback on the number of intervention sessions in their school not run as planned. Teachers were requested to return to the study coordinator completed implementation intention sheets after each session. These were subsequently collected from each school. For approximately half of these sessions, teachers were also requested to complete feedback sheets on session delivery. The feedback sheets included a rating of how well the session went (“The lesson went incredibly well;” strongly disagree , disagree , neither agree nor disagree , agree , strongly agree ). Quality of delivery was also assessed in observation of sessions by researchers. Approximately 7% of sessions were observed by researchers, including at least one session in each school. Observation sheets included a rating of overall quality of delivery (“Overall session quality was . . .;” low , moderate , satisfactory , good , high ). Exposure to the intervention was assessed by self-reported questions from participants at the final follow-up. Those in the intervention (antismoking) condition were asked to indicate which sessions they attended by checking a box next to each session (identified by number, short title, and image of the antismoking information) to give a score between 0 and 8. All participants were requested to indicate if they had moved school since the beginning of the study and to specify the old school and year of change (coded into total numbers changing school, numbers moving between schools in different conditions, numbers moving from nonstudy schools or nonspecified schools).

Data relevant to costing the intervention fully were also collected. A number of other measures were taken but are not reported here (full details available from first author along with intervention materials, analysis scripts, and raw data).

Statistical Analyses

Based on a power of 90% to detect a 5% difference in smoking rates, an intraclass correlation (ICC) of .01, and alpha of .05, prior sample size calculations indicated the need for at least 3,672 adolescents from 36 schools in the analyses ( Conner et al., 2013 ). We first summarized the measures taken for the full sample and the intervention and control conditions. The main analyses tested for differences between the intervention and control conditions at 48-month postbaseline in each of the four smoking measures among those who were self-reported never smokers at baseline. Those who self-reported ever smoking at baseline ( N = 301) were removed from all analyses. The largest amount of missing data was for baseline ever smoking, principally due to a failure to match individually generated codes. Missing self-reported ever smoking at baseline was imputed to be zero (i.e., never smoking). Missing data from other variables ranged from 0.2% for gender to 5.8% for any smoking (in last 30 days; see Table 1 for details of numbers of missing data points for each variable) and only 88% of the 6,115 never smokers in the sample would have been available for analysis under the traditional listwise deletion method across these variables. Data were primarily missing due to item nonresponse. We addressed the problem of missing data through multiple imputation using chained equations (MICE; van Buuren & Groothuis-Oudshoorn, 2011 ) after confirming that the missing values were missing at random. The mice command in R was used to generate 20 imputed data sets that were analyzed using the pooled command. Imputed values compared reasonably with observed values and the results using listwise deletion were similar to multiple imputation, so imputed results are presented.

Based on the distribution and frequency of outcomes, log binomial regressions, implemented in R were used to predict each smoking outcome (ever smoking; any smoking in the last 30 days, regular smoking, breath CO >6 ppm) controlling for the clustering among schools (multilevel modeling). Condition and percentage free school meals were Level 2 variables in these models, while gender, ethnicity, and the FAS scores were Level 1 variables. We report the risk ratio (RR), the 95% confidence interval around the risk ratio (95% CI), and the p value for each predictor variable in these regressions. The RR is the ratio of likelihood of the outcome (in this case smoking) across the compared conditions (intervention vs. control). For each step we also report the ICC. At Step 1 condition was entered, while at Step 2 we examined the effects of controlling for demographic variables (school SES; boys vs. girls; non-White vs. White ethnicity; individual level of socioeconomic status based on FAS). At Step 3 we tested whether each of these demographic variables significantly moderated the effects of the intervention. For outcome measures taken at each of the postbaseline time points (ever smoking, regular smoking, and breath CO >6 ppm), sensitivity analyses assessed intervention effects on smoking on at least one time point (i.e., for each smoking measure an outcome was created: 0 = not smoking at any time point; 1 = smoking at one or more time points). Fidelity analyses also examined whether attending no smoking intervention sessions versus a few or most smoking intervention sessions influenced the key findings. Fidelity analyses also examined whether the key findings were influenced by excluding participants who self-reported changing school.

The economic evaluation was based on the incremental cost of the intervention per averted smoker at age 15–16 years. The costs of implementing the intervention were gathered by researchers during the study and expressed in United Kingdom sterling in 2017 prices (converted to U.S. dollars) based on wages and transport costs as at August 2017 provided by the Office for National Statistics. Costs included intervention development (printing material), delivery (travel and time incurred in providing training and support), and receipt (teacher time in undertaking training). Costs over the 4-year period were discounted at 3.5% per annum consistent with NICE guidelines ( National Institute for Health & Care Excellence, 2018 ). An incremental cost-effectiveness ratio (ICER) was calculated based on the incremental cost per adolescent of implementing the intervention divided by the difference in the proportion not smoking across conditions.

Predicting Smoking Outcomes

Log binomial regressions in a multilevel model that controlled for the effects of school (and excluded baseline ever smokers) showed that self-reported ever smoking was significantly lower in the intervention compared to the control condition (Model 1, Table 2 ). This effect for condition on ever smoking remained significant when also controlling for demographic variables (Model 2, Table 2 ). Free school meals, gender, White ethnicity, and FAS were also significant predictors of ever smoking at this step, with higher levels of ever smoking being associated with more free school meals, being female, being White and with lower family affluence. None of the demographic covariates significantly moderated the effects of the intervention on ever smoking (free school meals: RR = 1.00, p = .790; gender: RR = 1.07, p = .387; ethnicity: RR = 1.14, p = .246; family affluence scale: RR = 1.02, p = .402). Sensitivity analyses also showed that the effect of the intervention on reducing ever smoking at any time point was significant when controlling for covariates ( p = .003) or not ( p = .016; see supplementary Table 1 ).

A similar pattern was apparent for any smoking (last 30 days ) with the condition effect being significant (i.e., less smoking in the intervention condition) when not controlling (Model 1, Table 2 ) or controlling (Model 2, Table 2 ) for demographic variables. Lower family affluence (FAS) was significantly associated with higher levels of any smoking (last 30 days ). None of the demographic covariates significantly moderated the effects of the intervention on any smoking (last 30 days ; free school meals: RR = 1.00, p = .851; gender: RR = 1.04, p = .732; ethnicity: RR = 1.25, p = .177; family affluence: RR = 0.98, p = .513).

The intervention had a weaker effect in relation to regular smoking and breath CO >6 ppm. Although regular smoking was lower in the intervention compared with the control condition (see Table 1 ), this difference was not statistically significant (Model 1, Table 2 ). The condition effect remained nonsignificant ( p = .092) when also controlling for demographic variables (Model 2, Table 2 ). More regular smoking was significantly associated with being in the control compared to the intervention condition, more free school meals, being White, and being from a less affluent family (FAS). None of the demographic covariates significantly moderated the effects of the intervention on regular smoking (free school meals: RR = 1.00, p = .945; gender: RR = 1.23, p = .283; ethnicity: RR = 1.47, p = .197; family affluence: RR = 0.91, p = .102). Sensitivity analyses showed that the effect of the intervention on reducing regular smoking at any time point was significant when controlling for demographic covariates ( p = .036) but was not significant when not controlling for covariates ( p = .156; see supplementary Table 1 ). Similarly, breath CO >6 ppm was lower in the intervention compared to the control condition (Model 1, Table 2 ), although this difference was not statistically significant ( p = .066). The effect for condition on breath CO >6 ppm became significant ( p = .024) when also controlling for demographic variables (Model 2, Table 2 ). Lower family affluence scores (FAS) were also significantly related to higher levels of breath CO >6 ppm. None of the demographic covariates significantly moderated the effects of the intervention on breath CO >6 ppm (free school meals: RR = 0.92, p = .061; gender: RR = 0.86, p = .689; ethnicity: RR = 1.71, p = .358; family affluence: RR = 0.86, p = .149). Sensitivity analyses showed that the effect of the intervention on breath CO >6 ppm at any time point was not significant when controlling for covariates ( p = .079) but was significant when not controlling for demographic covariates ( p = .027; see supplementary Table 1 ).

Fidelity Analyses

School coordinators informed us about a total of 37 sessions (approximately 1.6%) across control and intervention conditions that did not run as planned. We received teacher feedback on 797 (88%) out of 905 sessions from teachers, with 11 (1%) responding strongly disagree , 69 (7%) disagree , 348 (44%) neutral , 337 (42%) agree , and 30 (4%) strongly agree to the statement that the “lesson went incredibly well.” A total of 73 individual sessions (30 control, 43 intervention) were observed and scored for quality. In relation to overall session quality, none were rated as unsatisfactory , four (5%) as moderate , 16 (22%) as satisfactory , 46 (63%) as good , and seven (10%) as high quality . Lower ratings were mainly attributable to disruptive student behavior impacting on learning, too large a group to process the activities interactively, insufficient time to complete the activities, or insufficient staff input (e.g., no exploration of the activities as a group/students completed independently).

Completed implementation intention sheets were returned for approximately 91% of adolescents (89% in intervention; 95% in control) and the vast majority of sheets were scored as complete (88%) with no difference between conditions (87% in intervention; 90% in control). It was not possible to match individually generated codes to data on smoking thus precluding an analysis of the impact of completion on intervention effectiveness.

In the intervention condition at the final time point, a total of 496 (14%) participants reported attending no (0) intervention sessions, 542 (15%) reported attending a few (one to four) intervention sessions, and 2,590 (71%) participants reported attending most (five to eight) intervention sessions. Analyses indicated that controlling for clustering by schools and covariates there were generally few differences between those attending no intervention sessions and those who attended a few intervention sessions for each of the smoking outcomes ( ever smoking : RR = 1.11, 95% CI [0.96, 1.29], p = .143; any smoking (last 30 days ): RR = 0.78, 95% CI [0.62, 0.98], p = .032; regular smoking: RR = 0.96, 95% CI [0.69, 1.34], p = .811; breath CO >6 ppm: RR = 0.92, 95% CI [0.48, 1.76], p = .794). In contrast, there were significantly lower rates of smoking in those who attended most intervention sessions compared with those attending no intervention sessions for each of the smoking outcomes ( ever smoking : RR = 0.76, 95% CI [0.68, 0.86], p < .001; any smoking (last 30 days ): RR = 0.61, 95% CI [0.52, 0.72], p < .001; regular smoking : RR = 0.57, 95% CI [0.45, 0.73], p < .001; breath CO >6 ppm: RR = 0.41, 95% CI [0.24, 0.70], p = .001).

A total of 540 (9%) participants reported a change of school within the study period. A total of 82 moved between schools in the same condition, 86 moved between schools in different conditions, and 372 moved in to the study from nonstudy schools or did not specify the school they had moved from. These numbers were similar for the control and intervention conditions. Sensitivity analyses indicated that excluding these 540 participants did not substantively alter the main findings (i.e., no change in significance of condition for any smoking outcomes).

Economic Analyses

The intervention was costed at $1,391 (£1,031) per school over the 4-year program duration. Twenty-eight percent of the intervention cost was associated with covering teacher time (seven members per school) to attend an intervention training session (45-min duration) pertaining to how to deliver the intervention, conservatively assuming that: (a) teachers attend the training every year, and (b) their workload is already fully allocated. The rest of the intervention cost stemmed from printing and delivering material to schools, researcher travel, and time incurred in training teachers and various administrative support tasks.

Based on an average school size of 160 pupils (initially aged 12–13 years), the intervention cost is $8.69 (£6.44) per adolescent (see supplementary Table 2 for further cost details). When comparing this intervention cost to the trial arm difference in the proportion who were never smoking at follow-up (6.5% more on ever smoking measure in intervention arm; Table 1 ), the intervention yields an incremental cost-effectiveness ratio (ICER) of $134 (£99) per ever smoker avoided at age 15–16 years. A sensitivity analysis including the sunk cost of designing the intervention ($21,159 [£15,680]) and assuming teachers’ delivery of sessions requires extra time (i.e., they are done outside existing sessions), the intervention cost increases to $18.99 (£13.33) per adolescent. This yields an ICER of $292 (£205) per smoker avoided at age 15–16 years based on ever smoking assessment.