This systematic review and meta-analysis was conducted in accordance with the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) [35]. The PRISMA checklist can be found in Multimedia Appendix 1.

Studies were included if they (1) examined an individual-level text messaging intervention to promote smoking cessation, (2) used a RCT design, (3) assessed smoking outcomes (eg, abstinence, quit attempt), (4) provided sufficient statistical information to calculate ESs, and (5) were available (including electronic publications and dissertations) by December 31, 2014. Studies that examined text messaging interventions with the goal of maintaining (vs initiating) smoking abstinence among recently quit smokers were excluded (eg, Snuggs, McRobbie [36]).

Multiple electronic reference databases (PubMed (1946+), PsycINFO (1872+), ProQuest Dissertations and Theses Full Text (1973+), CINAHL (1981+), Global Health (1973+), The Cochrane Library (1992+), Communication & Mass Media Complete (1915+), and EMBASE (1947+)) were searched using a Boolean search strategy: (tobacco OR smok*) AND (text messag*) OR (cellular phones) OR (cell AND phone) OR (mobile) OR (mobile devices) OR ("short message service") or ("multimedia messaging service") AND (interven* or prevent*). Because many electronic databases have specific search methods (eg, Medical Subject Heading (MeSH) terms used in PubMED are not used in other databases such as PsycINFO), our basic search strategy was modified to accommodate the specific search parameters for each electronic database. No language restrictions were applied. The electronic reference database searches were initially conducted in April 2014 and updated in January 2015 to ensure that we retrieved all available studies through December 31, 2014. Reference sections of relevant manuscripts (including published reviews obtained through the electronic reference database searches) were also reviewed. Finally, we searched the tables of contents of relevant journals (Journal of Medical Internet Research; Telemedicine & eHealth) for relevant papers.

Study titles and abstracts were initially screened for possible inclusion. Full-text manuscripts of potentially relevant records and references from relevant manuscripts were retrieved and reviewed for final inclusion. Studies that fulfilled the inclusion criteria (see Eligibility Criteria) were included. When authors reported details, ancillary information (eg, results from a pilot study), and/or study information across multiple manuscripts, those manuscripts were linked in the database and represented as a single unit and the manuscript reporting the most complete study sample was selected as the primary manuscript.

Two of three independent coders (LAJSS, RCL, or EGJ) extracted study information (eg, publication year), sample characteristics (eg, gender, ethnicity), design specifics (eg, recruitment method), intervention procedures (eg, number and frequency of text messages), and components (eg, personalized feedback, goal setting) from each study. Methodological quality was assessed using 14 items adapted from validated measures [37,38]. Interrater reliability was assessed for all study, sample, and methodological variables. For the categorical variables, raters agreed on a mean of 79% of the judgments (mean Cohen’s κ =.60). Reliability for the continuous variables yielded an average intraclass correlation coefficient (ρ) of 0.89 across categories (median=0.99). Disagreements between pairs of coders were resolved through discussion.

Study outcomes included dichotomous (eg, frequencies) and continuous (eg, means) assessments of smoking cessation including abstinence, quit attempts, and cigarette use. Smoking abstinence was assessed using a number of methods including point prevalence abstinence (ie, abstinence from a specific time-point to follow-up assessment), continuous abstinence (ie, abstinence from quit date to follow-up assessment), prolonged or sustained abstinence (ie, sustained abstinence between two assessments), and repeated point prevalence abstinence (ie, abstinence from at least two specific time-points to follow-up assessments; see Hughes et al [39] for details regarding smoking abstinence definitions). Other measures included making a quit attempt and the quantity of cigarettes smoked per day or week. Finally, we assessed nicotine dependence, which was measured using validated measures (eg, Heaviness of Smoking Index [40]).

ESs were calculated for each study by two of the three independent coders (LAJSS and HT or EGJ). Because the studies assessed smoking outcomes using dichotomous (eg, abstinence) and continuous (eg, quantity of cigarettes smoked per week or month) variables, we used two ES indices to represent the outcomes. For dichotomous outcomes, we estimated a summary odds ratio from 2 × 2 tables by calculating the odds ratio and transforming it to a log OR (with the corresponding standard error) [41]. For continuous outcomes, ESs were calculated using the standardized mean difference [42]. In the absence of means and standard deviations (SD), other statistical information (eg, F test) was used [33,43]. ESs (Cohen’s d) were corrected for sample size bias [44].

Multiple ESs were calculated from individual studies when the study included multiple intervention conditions (2 studies), measured more than one outcome (18 studies), used (or provided sufficient data to calculate) more than one statistical method to analyze the outcomes (ie, intent-to-treat and complete-case analysis; 19 studies), or assessed outcomes at multiple follow-ups (5 studies). ESs calculated for each intervention were treated as separate studies to avoid violating the assumption of independence [33,43]. Positive ESs indicated that participants who received the text messaging intervention reported higher smoking abstinence rates, more quit attempts, smoked fewer cigarettes per day or week, and scored lower on measures of nicotine dependence compared with a control or comparison condition. All ESs were reviewed for accuracy; discrepancies were resolved through discussion and final calculations.

ESs were analyzed separately by smoking outcome and stratified by type of analysis (ie, intent-to-treat or complete case analysis). The overall ES for smoking abstinence using standard and cumulative meta-analytic approaches was also assessed (a cumulative meta-analysis provides an updated pooled ES estimate each time a new trial is added). Because smoking abstinence was measured using a number of methods (point prevalence, continuous abstinence, prolonged or sustained abstinence, and repeated point prevalence), only a single measure from each study was included to avoid violating the assumption of independence [43]. Decisions regarding which ES to include for each intervention were based on the most commonly reported measure (ie, 7-day point prevalence), followed by the next common measure, and so forth until all studies reporting an abstinence measure were represented. Two studies (k=3) did not include measures of smoking abstinence [45,46].

Summary ESs were calculated using random-effects procedures such that ES were weighted by the inverse of their random-effects variance [33,47]. The homogeneity statistic, Q, was calculated; a significant Q indicates a lack of homogeneity and an inference of heterogeneity. The I ² index and the corresponding 95% CIs were also calculated to assess the observed dispersion [48-50] The I ² index ranges from 0% to 100% with 25%, 50%, and 75%, considered low, moderate, and high levels, respectively, of observed variance reflecting true differences in ESs [51]. Analyses were conducted in Comprehensive Meta-analysis [52] and Stata [53] using publicly available macros [33].

Moderator analyses examined the overall effects of smoking abstinence using a modified weighted least squares regression analyses or the meta-analytic analogue of an analysis of the variance (ANOVA) with weights equivalent to the inverse of the variance plus the random variance component for each ES [33,54,55]. Random-effects models (methods of moment) were estimated. Analyses examined a priori determined moderators of smoking abstinence. Sample and methodological characteristics (eg, gender, age, region, recruitment method), intervention type (text only vs text plus other intervention components), intervention dose and delivery (eg, number of texts sent, frequency of texts), message type (targeted), and intervention content (eg, personalized feedback, goal-setting) were examined. All moderator analyses were conducted in Comprehensive Meta-analysis [52].

Asymmetries in the distributions of ESs, indicating a possible reporting bias [56], were examined by inspecting funnel plots and assessing the degree of asymmetry [57-59]. As recommended by Stern and colleagues [60], we conducted tests for publication bias only if the dependent variable included 10 or more studies. Trim and fill procedures [61,62] are used to estimate and correct for the possibility of missing studies (based on a rank-based data augmentation procedure) if publication bias is detected using funnel plot asymmetry tests [58,59].

Our search strategy yielded a total of 3678 unique records. Following title and abstract screening, 90 full-text manuscripts were assessed for eligibility. Of these 90 manuscript, 51 were excluded because they did not include a text messaging intervention and/or a control or comparison condition, used a quasiexperimental design, assessed only the maintenance of smoking abstinence (rather than initial smoking cessation), did not measure smoking behavior, or did not provide data suitable for meta-analysis (eg, protocol, qualitative). Our final sample included 20 studies reporting 22 interventions (k) (Figure 1). An additional 19 manuscripts were retained as supplemental information for the included studies (details of the 20 studies included in the meta-analysis are provided in Table 1).

The characteristics of the studies, samples, and interventions included in the meta-analysis are described below. Full details are provided in Table 2.

The overall summary ES for smoking abstinence was significant, OR=1.37 (95% CI=1.25, 1.51; k=19). That is, participants who received a text messaging intervention were 1.37 times more likely to abstain from smoking relative to controls. The hypothesis of homogeneity was supported: Q₁₈=19.36, P=.370. I ² was 7% (95% CI 0, 42). The confidence intervals surrounding I ² did not exceed the 50% threshold indicating that the proportion of observed heterogeneity is low (a forest plot of the overall smoking abstinence is provided in Figure 2).

The cumulative meta-analysis was performed using the final completion date for data collection for each study (Figure 3). Results showed that the benefits of using a text messaging approach for smoking cessation was established by 2009 (end of data collection for Free et al [30]), after only five RCTs involving 8383 smokers (OR=1.39, 95% CI=1.15, 1.67, P<.001). Results from the additional 13 studies (k=14) with 6870 participants did not change the established efficacy of text messaging interventions for smoking cessation. The CIs surrounding the ES estimates narrowed as the data accumulated. Exploratory analyses restricted to the 10 studies that were of moderate to high methodological quality (ie, studies satisfying at least 65% of the methodological quality criteria; median=70%, range, 65%-80%) also indicated that the efficacy of text messaging for smoking cessation was established by 2009 (end of data collection for Free et al [30]), after only three moderate to high quality studies involving 6388 smokers, OR=1.46, 95% CI=1.29, 1.64. A cumulative plot for smoking abstinence restricted to studies meeting criteria for moderate to high methodology quality is provided (Figure 4).

Moderator tests were conducted for overall smoking abstinence (Table 4). Compared with controls, text messaging interventions were more successful in increasing smoking abstinence when the trials included fewer women (B=0.58, SE=0.26, P=.025) and were conducted in Asia (OR=2.16, 95% CI=1.02, 4.61), North America (OR=1.94, 95% CI=1.41, 2.67), or Europe (OR=1.46, 95% CI=1.31, 1.62) versus Oceania (OR=1.12, 95% CI=0.97, 1.30), Q₃=14.38, P=.002. Text messaging interventions were more successful at increasing smoking abstinence when the participants were recruited via the Internet (eg, Web-based advertisements, quitline; OR=1.72, 95% CI=1.41, 2.11) versus offline (eg, schools or clinics; OR=1.45, 95% CI=1.05, 2.00) or via a combination of Web-based and offline approaches (eg, Internet and community ads; OR=1.30, 95% CI=1.18, 1.43), Q₂=6.39, P=.041.

Both graphical and statistical tools were used to test for the possibility of publication bias. Results from Begg’s test [58] and Egger’s regression asymmetry test [59] revealed no evidence of publication bias for the dependent variables with sufficient cases for assessment (ie, 7-day point prevalence abstinence and overall smoking abstinence). The funnel plots and results of the statistical tests are available in Multimedia Appendix 2.

This meta-analysis evaluated the impact of text messaging interventions on smoking outcomes among 20 RCTs reporting on 22 interventions among 15,593 smokers. The results of this meta-analysis provide evidence for the efficacy of text messaging interventions on smoking outcomes. The overall odds of smoking abstinence were 1.37 times higher in the text messaging versus control or comparison groups. This finding is comparable to other meta-analyses evaluating text messaging interventions for smoking cessation [23-25]. Furthermore, our cumulative meta-analysis showed that the benefits of text messaging interventions for smoking cessation were established by 2009, after only five RCTs involving 8383 smokers and culminating with Free et al [30], although the results of the trial were not published until 2011. Most of the subsequent RCTs were already underway or completed by 2011. The robustness of this finding (coupled with limited evidence of heterogeneity) clearly indicates that conducting any future RCTs with the primary goal of assessing the efficacy of text messaging for smoking cessation is unnecessary.

It is noteworthy that the summary ESs favored the treatment groups even when 90% (18/20) of the controlled trials used an active control and 67% (12/18) of these active controls included some smoking-related content. These active controls with smoking-related content used a variety of means to disseminate smoking cessation information including smoke-free websites, self-help guidebooks, and smartphone apps. Some of the active controls included in the overall smoking abstinence analyses also provided text messaging, but the text messaging was a weaker form of that offered to the intervention groups or contained unrelated content (eg, diet and physical activity, importance of study participation). For example, Pollak et al [82] compared text messaging support messages that used a scheduled, gradual smoking reduction approach with standard smoking-related text messaging support messages and Free et al [30] compared smoking-related text messages with study-related participation texts (eg, thanking participants for taking part in the study, requests for updating contact details). Further probing revealed that the use of an active comparison condition with or without text messaging did not moderate smoking abstinence (active comparisons with any text messages: OR=1.37, 95% CI=1.17, 1.61, k=8; active comparisons without text messages: OR=1.39, 95% CI=1.18, 1.64, k=10; Q₁=0.02, P=.897). Furthermore, there was no difference between active comparisons that included smoking-related (OR=1.55, 95% CI=1.02, 2.38, k=3) or nonsmoking (OR=1.35, 95% CI=1.10, 1.66, k=5) text messaging, Q₁=0.38, P=.561. Thus, our meta-analysis provides evidence that text messaging interventions for smoking cessation improves smoking abstinence above and beyond other (weaker) smoking cessation delivery modalities with or without text messaging.

Our meta-analysis also demonstrated that the magnitudes of the summary ESs across measures of smoking abstinence were weaker in text messaging studies using an intent-to-treat analytic approach versus a complete case analysis (ORs ranged from 1.38-2.60 vs 1.43-3.62). Excluding participants after randomization from analyses introduces bias that may alter the conclusions made about individual studies’ treatment effects [34]. Failure to account for these biases can also affect meta-analytic results. Future studies should include intent-to-treat analyses when presenting their results to minimize potential biases, and perhaps make efforts to examine patterns of attrition.

The results from our moderator analyses revealed three important moderators of text messaging interventions for smoking cessation. First, text messaging interventions conducted in North America, Europe, and Asia produced better results than those conducted in Oceania. All of the RCTs located in North America were conducted in the United States Over the past 20 years, smoking has become far less socially acceptable in the Unites States than in previous decades [97]. Smoking is no longer permitted in government offices and other government facilities, and in most of the Unites States smoking is not permitted in restaurants and many other public venues [98]. Smoke-free laws are also in place in many European Union countries, and many countries represented in this study have completely banned smoking in public places, workplaces, or on public transportation [99]. Smoke-free environments have also been introduced in many Asian countries such as China and Turkey [100,101]. It may be that light-touch/low-intensity interventions such as text messaging are most effective when the surrounding environment supports cessation, or at least is actively unsupportive of continued smoking.

Second, the efficacy of the text messaging interventions for smoking cessation differed by men and women. That is, studies with larger proportions of women participants were less successful at improving smoking abstinence. Prior research has suggested that interventions to increase smoking cessation may be less effective for women than men [102]. Women may be less likely than men to quit smoking for a number of reasons including weight concerns, less social support for quitting, genetic variants that affect the efficacy of pharmacotherapies, and mood regulation [103]. Interventions that specifically address women’s concerns can help women stop smoking [104]. Only two text messaging interventions included in this meta-analysis were targeted to women, specifically pregnant smokers, and found mixed results [81,82]. One study found no differences in smoking abstinence between the text messaging intervention and no message control group, while the other study showed a significant difference in smoking abstinence among women who received text messages that used a scheduled gradual smoking reduction approach versus support text messages alone. Future text messaging interventions for smoking cessation should address the specific treatment needs of women with further attention to the type and intensity of the text messages desired by women smokers.

Finally, recruitment method was associated with increases in smoking abstinence. Studies in which participants were recruited exclusively via the Internet achieved higher rates of smoking abstinence relative to studies that recruited participants using an offline or combination of Web-based and offline recruitment methods. It may be that Web-based recruitment is more culturally consistent with use of a technology-delivered intervention, and individuals who respond to Web-based recruitment may be more comfortable with and better able to respond to a text-message delivered program compared with other individuals. More work is needed to identify the characteristics of individuals who do well with technology-delivered interventions versus those who would respond better to more traditional (eg, in-person) therapies.

Prior research comparing four Web-based (3: health risk assessment; advertisements; quit line screening) and offline (1: offline materials such as television advertisements) recruitment methods also showed that Web-based advertisements had higher yield rates and were more cost-effective than other approaches [105]. Despite the potential benefits of Web-based recruitment, other research shows that participants recruited via the Internet (vs offline) communication are less likely to participate through follow-up [106]. Retention rates for studies using Web-based, offline, and combined recruitment methods in this meta-analysis were 70%, 83%, and 78%, respectively. Retaining smokers in smoking cessation interventions is an ongoing concern but these concerns may be mitigated by the added convenience, potential participation rates, and cost-effectiveness of Web-based recruitment. Nonetheless, text messaging interventions for smoking cessation should employ strategies known to be effective for increasing retention (eg, emphasizing the benefits of participation, reminders [107]) when recruiting smokers via the Internet.

Several limitations should be considered when interpreting our findings. One limitation of evaluating the effect of smoking cessation interventions is that studies used many different measures (eg, variety of measures of abstinence such as point prevalence and continuous abstinence; for a discussion, see [39]). Because some of these measures (eg, repeated point prevalence) were used in a limited number of studies, we could not assess moderators of these individual types of measures separately and necessitated pooling measures for our overall analyses. Second, evaluation of this literature is also limited as few of these studies systematically assessed the additive effect of text messaging. That is, most of the text messaging interventions included text messaging plus some other smoking-related content (eg, counseling, supporting website) rather than testing the effects of the same smoking cessation intervention with or without text messaging. Future studies should examine the additive effects of text messaging plus other smoking-related content. Third, identification and retrieval of relevant studies may have been hindered by the author’s use of keywords (eg, failure to include ‘intervention’ as a keyword because the intervention modality was the focal point of the study). To reduce the possibility of inadvertently missing studies, we searched multiple electronic databases, tables of contents of relevant journals, and reference sections of relevant papers and reviews [108]. Fourth, we focused our analyses on the final post-intervention assessment because most of the studies (15/20, 75%) did not provide data from multiple post-intervention assessments. Using the last post-intervention assessment as the point of analysis provides a stronger test of the potential effects of the intervention on long-term cessation because initial intervention effects tend to decay over time (cf. Johnson et al [109]). In this meta-analysis, however, the final post-intervention assessment occurred most often immediately following the completion of the intervention (median=1 week; range, 0-44 weeks). Future research with follow-up periods extended to 1 year or longer is needed to determine whether reductions in smoking behavior are sustained over time. Finally, our moderator tests were limited to the available data. Potentially important moderators could not be tested because there were too few or too many cases to be evaluated (eg, 22/22, 100% of the interventions provided self-management training), and thus were omitted from our analyses.

The prevalence of cigarette smoking has declined over the past decade; however, more than 1 billion adults continue to smoke worldwide [1]. Text message interventions to reduce tobacco use have the promise to reach a wider audience with minimal cost and fewer resources. The current meta-analytic review provides unequivocal support for the efficacy of text messaging interventions for smoking abstinence. Future research should be directed to understanding for whom and under what circumstances text messaging interventions are optimized, and the duration of the effects.