Kellam 1998 (GBG).

Methods	Country: USA. Site: 19 elementary schools in Baltimore. 'Good Behaviour Game'. Focus: Smoking prevention by changing behaviour predicting later smoking uptake. Design: Cluster RCT (Group 1: never‐smoking prevention cohort).
Participants	Baseline: 2311 (analysis limited to 1604 nonsmokers at baseline (Cohort 1 = 818 entered 1st grade in 1985, Cohort 2 = 786 entered 1st grade in 1986) . Age: 5 ‐ 6 yrs. Gender: 49.6% F Ethnicity: Greek and Italian 14%, White 16%, A‐A 70%. Baseline smoking data: Tobacco users: N = 502 (for group which entered 1st grade in 1985 = 275, for group which entered 1st grade in 1986 = 227); Tobacco nonusers = 1102 (for group which entered 1st grade in 1985 = 543, for group which entered 1st grade in 1986 = 559).
Interventions	Category: Other interventions vs control. This intervention did not align with the main 5 categories; the programme intervenes by rewarding good behaviour in the classroom. Programme deliverer: Teachers. Intervention: 2 yrs (1st and 2nd grades). Compared 2 programmes designed to reduce future tobacco usage by addressing risk factors for uptake: Good Behaviour Game, led by classroom teachers during regular classes. They defined and posted undesirable behaviours (fighting, shouting out of turn, and teasing), and the class with the most points for good behaviour won prizes. The game was played initially for 10 mins 3 times weekly, increasing in frequency and duration. Mastery Learning for reading. Students proceeded to the next unit only when they mastered 85% of the learning objectives, small groups, formative testing, and individual instruction. Control: "customary school programs" / reading skills intervention.
Outcomes	Definition of smoking: "tobacco user", 'tobacco nonuser'. Assessed at individual interview. Follow‐up: From age 8 to 14.
Notes	Quality of intervention delivery: No process analysis. Statistical quality: Was a power computation performed? No. Was an intention‐to‐treat analysis performed? Yes, three different statistical methods were performed. Was a correction for clustering made? "To accommodate clustering of students within initial elementary schools, this analysis involved presorting of students into strata defined by school attended in first grade". Were appropriate statistical methods used? Yes. "We used standard life table and survival analysis methods to compare risk of initiating tobacco use for 2 interventions and all internal and external control classrooms. Kaplan‐Meier survival curves for each group were compared via log‐rank statistics as an aid to interpretation. Adjusted estimates for the relative risk of tobacco smoking were also obtained via conditional forms of Cox proportional hazards modelling; EGRET was used in calculating estimates".
Risk of bias
Bias	Authors' judgement	Support for judgement
Random sequence generation (selection bias)	Unclear risk	"Five urban area were defined with socioeconomic levels ranging from very poor to middle class. In each area, 3 to 4 public elementary schools with similar racial/ethnic profiles were selected. Within each area, the Good Behavior game was randomly assigned to 1 school and Mastery Learning to another; 1 or 2 schools were controls". Method of randomisation not stated. Clusters: Schools Cluster constraint: Schools with geographical area stratified according to socioeconomic levels. Baseline comparability: "There were no significant differences between the intervention groups on baseline characteristics of teacher ratings of aggressive, disruptive behavior, fall‐of‐first ‐grade achievement, or free or reduced‐price school lunch when we took into account school as a random factor".
Allocation concealment (selection bias)	Unclear risk	No statement
Blinding of outcome assessment (detection bias) All outcomes	Unclear risk	No statement
Incomplete outcome data (attrition bias) All outcomes	Unclear risk	"When attrition occurred it was unrelated to intervention status". Attrition 31%; Kellam 2008 reported on students who were re‐interviewed at age 19‐21; For Cohort 1, Table 6 lists 278 males and Table 7 348 females (total N = 626, (i.e. 77% of Cohort 1 in Grade 1 in 1985). The analysis offers 3 different intention‐to‐treat analyses, but these data are for a smaller sample size than baseline.
Selective reporting (reporting bias)	Low risk	No selective reporting