Cigarette smoking remains the leading cause of preventable disease and death in the United States [1]. Although the prevalence of smoking has been declining for decades, smoking rates have declined more slowly in socioeconomically disadvantaged groups [1,2]. One such group facing unique challenges in quitting smoking is people experiencing food insecurity or a lack of consistent access to sufficient food to lead a healthy, active life [3]. A person’s level of food security is largely determined by purchasing power, which is mostly derived from wages, government transfers (eg, Social Security payments), and household composition [4], and money spent on cigarettes becomes unavailable to be spent on essential goods, including food. Although food insecurity is associated with both financial strain and psychological distress, it is also an independent predictor of smoking status and intensity and likely impedes smoking cessation success via multiple mechanisms, such as the reinforcement of smoking via nicotine’s appetite suppressant effects and negative affect reduction [5]. Thus, there is a critical need to directly target individuals experiencing food insecurity with evidence-based smoking cessation interventions. In this population, successful long-term smoking cessation might be especially beneficial, as it may lead not only to abstinence-based health improvements but also to potential increases in food security due to both eliminating spending on cigarettes and reducing spending on smoking-related health care costs.

Recruiting special populations (eg, patients with cancer) to smoking cessation intervention trials can be very challenging, with eligibility percentages among those contacted sometimes in the low single digits [6]. Given these and other obstacles to obtaining desired sample sizes, researchers targeting various special populations have often used innovative recruitment strategies. For example, a recent feasibility trial for a text-based smoking cessation intervention for sexual and gender minorities yielded 79 participants following the posting of approximately 700 flyers in a variety of public establishments in the local catchment area, 1500 emails sent to potentially eligible participants on ResearchMatch (who were overwhelmingly from outside the catchment area), and the posting of advertisements on Facebook, Instagram, and Craigslist. The screening yield rate for each modality was below 6%, although the enrollment rate among those screened was approximately 50% [7].

Latino Americans are another special population that can be difficult to successfully reach and enroll in smoking cessation randomized controlled trial (RCTs), necessitating multiple types of recruitment approaches. For example, a secondary analysis of data from Decídetexto, a mobile health smoking cessation RCT for Latino Americans conducted across 4 states, compared the effectiveness of outbound, direct recruitment strategies, in which study staff contacted potential participants, to inbound, mass recruitment strategies (eg, flyers and television, newspaper, or Facebook advertisements), in which interested individuals had to contact study personnel to be screened [8]. Although most enrollees were recruited via “high effort” direct strategies (eg, in-person community outreach or personal phone calls to patients in registries), individuals recruited through either mass recruitment or “low effort” direct strategies (eg, emailing or texting patients in registries) were more likely to screen eligible and to enroll in the trial. In terms of personal characteristics, mass-recruited enrollees differed considerably from enrollees recruited via direct strategies, including having lower socioeconomic status, being less likely to have health insurance, and being less acculturated (eg, being more likely to primarily speak Spanish).

A contrasting pattern emerged from a smoking cessation RCT conducted by our team for another special population, cervical cancer and precancer survivors [9,10]. Enrollees recruited online (vs enrollees recruited from clinics at a National Cancer Institute–designated cancer center) had higher levels of education and health literacy but did not differ in age, race, household income, smoking behaviors, or motivation to quit smoking [9]. Thus, online recruitment strategies (eg, social media advertisements) cannot be presumed to consistently yield more- or less-disadvantaged members of special populations than in-person strategies.

To date, few studies have specifically addressed the process of recruiting people who smoke and are experiencing food insecurity to smoking cessation RCTs. One notable exception, however, was a pilot study that supported the feasibility of in-person recruitment of such individuals at food pantries to a smoking cessation treatment study [11]. The study entailed 22 research staff visits distributed across 4 food pantries in the Greater Cleveland region. During each site visit, the research staff set up an outreach table and remained there for between 1.5 and 3.5 hours, during which approximately 28% of the clients (173 of an estimated 628 total pantry visitors) approached their table. Of them, 31% (54/173) expressed interest in information on quitting smoking, and of those interested in information on quitting, 78% (42/54) were able to be contacted and screened for eligibility via phone. Ultimately, 18% (31/173) of eligible individuals gave consent to have their contact information sent to the Ohio Tobacco Quitline (ie, 5% [31/628] of individuals served by the food pantries during recruitment) [11]. Thus, the recruitment yield was low.

With the limited exceptions of the studies described earlier that recruited Latino Americans and cervical cancer or precancer survivors [8,9], few studies have examined (1) potential sociodemographic, psychosocial, or smoking-related differences among individuals recruited to smoking cessation trials using in-person versus online recruitment methods or (2) differences in recruitment yield and enrollment outcomes associated with these 2 methods [6,9]. Indeed, to our knowledge, the current secondary analysis represents the first such comparison using data from a smoking cessation RCT for people experiencing food insecurity, a vulnerable special population for whom quitting is especially urgent.

The primary objective of this secondary analysis was to compare baseline sociodemographic, smoking-related, and psychosocial characteristics of participants recruited via in-person versus online methods to a National Cancer Institute–funded smoking cessation trial evaluating the efficacy of a mobile health intervention for individuals experiencing food insecurity, as such differences may have clinically meaningful implications for trial design and outcome interpretation. A secondary objective was to describe screening, eligibility, and enrollment rates among individuals recruited in person versus online.

The study underlying this secondary analysis is a smoking cessation RCT evaluating the efficacy of a fully automated smartphone-delivered intervention (automated treatment [AT]) versus connections to state quitlines (standard treatment [ST]).

Individuals reporting current smoking and recent receipt of food assistance (eg, food bank or food pantry or electronic benefits transfer [EBT] or food stamps) were recruited in person (March 2022 to February 2024) and online (February 2024 to June 2024). All assessments were conducted online via REDCap (Research Electronic Data Capture), and participants were followed for 12 months [12]. Additional procedural information has been published previously [13].

The AT intervention was fully automated after enrollment, whereas each ST participant’s contact information was provided to the appropriate state quitline, and the quitline proactively contacted the participant. All participants received a 10-week supply of combination nicotine replacement therapy (ie, nicotine patches and lozenges).

The AT participants used a unique link sent to their email address to download the treatment app. Data collected on the app were stored via a backend database meeting all Moffitt Cancer Center data security policies and procedures. Participants in the ST condition were connected to state quitlines that were free of charge to callers. All participants were assigned unique identification numbers for use in data analyses and for data transfer, in the event of a request. Only study identification numbers were used in analysis files.

Recruitment groups were compared on categorical variables using chi-squared and Fisher exact tests and on continuous variables using independent samples t tests. Effect sizes were gauged using Cohen d for continuous variables and either Cramer V or odds ratios for categorical variables. CIs for Cohen d were created using the noncentral t distribution method, creating exact interval estimates for the standardized mean differences [15], while 95% CIs for Cramer V, which is bounded between 0 and 1 and whose magnitude can be interpreted similarly to a Pearson correlation coefficient [16,17], were constructed using a nonparametric percentile bootstrapping procedure with 10,000 resamples [18-20]. The 95% CIs for the odds ratios were calculated using the standard, natural logarithm transformation–based formula [21].

To explore whether differences in self-reported food insecurity between the in-person and online groups might have been a consequence of basic demographic differences between the groups, a hierarchical linear regression analysis was performed. Little’s MCAR (missing completely at random) test was used to ascertain whether the data were MCAR [22]. Although there were no missing data on in-person versus online group membership, nor on food insecurity, cases with missing data on any of the predictors present in one or both models tested were deleted listwise, which resulted in a nominal 5% reduction of the analytic sample size from 502 to 475. To identify model covariates, univariate demographic predictors of either food insecurity (P<.10) or recruitment modality (P<.05) were identified using simple logistic (categorical variables) or linear (continuous variables) regression analyses [9]. (Candidate covariates included age, ethnicity or race, sex, sexual orientation, relationship status, educational attainment, health insurance coverage, household income, and employment status.) Next, variables meeting those criteria were entered as a set to estimate a covariates-only model (model 1). Before model inclusion, categorical variables were dichotomized (eg, income was approximately median split to form less than US $20,000 per year and US $20,000 or more per year categories) to prevent estimation problems (eg, coefficient instability and inflated SEs) [23]. The next step entailed adding recruitment modality to the model as a predictor (model 2). Differences in model fit were examined using an F test; a significant difference would indicate an improvement in model fit, meaning that there was an association between food insecurity and recruitment modality even after accounting for the effects of relevant demographic covariates. The 95% CIs for the proportion of variance in self-reported food insecurity explained by models 1 and 2 (ie, model R²) were based on a noncentral F distribution and created using R package MBESS’ function “ci.R2” with parameter Random.Predictors set to “FALSE,” as the model predictors were fixed, rather than random [15,24]. To generate the increment in R² between models 1 and 2 (ie, ΔR²), nonparametric bootstrapping with 10,000 resamples was used. An inspection of the histogram of the bootstrapped ΔR² estimates revealed slight positive skew in the distribution, and so both percentile and bias corrected and accelerated (BCa) bootstrap CIs were calculated and reported [20] (for details, see Multimedia Appendix 1).

To compare differences in enrollment patterns between participants recruited in person versus online, the flow of potential participants between prescreening and full enrollment in the trial was presented, keeping the in-person and online recruitment denominators (and thus percentages) as comparable as possible within the constraints of the available data. Prescreening numbers were defined as the number of people assessed for eligibility after being approached at food pantry sites and the number of people who were assessed for eligibility after clicking online advertisements. Regardless of whether they were recruited in person or online, potential participants who were not excluded at the prescreening stage completed the study’s eligibility questionnaire.

Potential participants who prescreened eligible (eg, aged ≥18 y, self-reported smoking, and receiving food assistance) were considered “unreachable” after not being able to be contacted following approximately 2 weeks of staff outreach attempts using a structured outreach schedule. For the first week following prescreening, 3 REDCap notifications were sent via email and text (using Twilio integration) to remind potential participants of the screening survey, and study staff attempted contact via phone, text, or email 3 to 5 times. The second week consisted of 3 contact attempts approximately every other day. A similar procedure was followed at other steps of the enrollment process. For example, a participant who consented to participate but did not complete the baseline questionnaire was considered unreachable after 2 weeks of contact attempts.

Next, those who screened eligible and were not excluded (either because they were unreachable or because they declined to participate) completed the informed consent process. The onboarding completion rate was defined as the number of people who enrolled in the trial divided by the number of people who completed the consent process. Enrollment rates for those who both consented to participate and subsequently completed the preenrollment baseline questionnaire were calculated by dividing the number of people who completed the baseline questionnaire by the number of people who enrolled in the trial.

To provide a fuller sense of the intensity of recruitment efforts by recruitment modality, time (in days) from screening completion to enrollment was reported, along with the number of contact attempts made prior to enrollment for online recruits. Owing to the positive skew in the distributions of these values, medians and IQRs were reported in addition to means and SDs.

Online (vs in person) recruits reported greater food insecurity (mean 4.5, SD 1.9 vs mean 3.0, SD 2.3; d=0.74, 95% CI 0.55-0.93; t₅₀₀=7.93; P<.001). As illustrated in Figure 1, 64% (n=208) recruited online (vs in person: n=66, 37%) were classified as having very high food insecurity (odds ratio [OR] 3.04, 95% CI 2.05-4.53; χ²₁=34.1; P<.001). Online recruits were also younger (45.7 vs 52.9 y; d=−0.70, 95% CI −0.89 to −0.51; t₅₀₀=−7.48; P<.001) and less likely to be Hispanic or Latino (n=22, 7% vs n=39, 22%; OR 0.25, 95% CI 0.14-0.47; χ²₁=24.9; P<.001). Participants recruited online were more likely to be female (n=250, 77% vs n=117, 66%; OR 1.76, 95% CI 1.15-2.69; χ²₁=7.6; P=.006), non-Hispanic White (n=249, 77% vs n=88, 50%; OR 3.35, 95% CI 2.26-5.06; χ²₁=38.3; P<.001), and LGBT+ (n=57, 18% vs n=11, 6%; OR 3.20, 95% CI 1.60-6.96; χ²₁=12.5; P<.001). Moreover, online recruits were more likely to be educated beyond high school or a general educational development certificate (n=225, 69% vs n=87, 49%; OR 2.37, 95% CI 1.60-3.53; χ²₁=20.7; P<.001) and to have annual household income of US $20,000 or more (n=150, 46% vs n=64, 36%; OR 1.52, 95% CI 1.03-2.26; χ²₁=4.8; P=.03).

Online recruits reported lower motivation to quit smoking (Contemplation Ladder; mean 7.2, SD 2.4 vs mean 8.0, SD 2.8; d=−0.315, 95% CI −0.499 to −0.131; t₅₀₀=3.38; P<.001), smoking cessation self-efficacy (mean 20.5, SD 8.0 vs mean 23.2, SD 8.6; d=−0.318, 95% CI −0.502 to −0.134; t₄₉₉=3.40; P<.001), and subjective social status (mean 4.6, SD 2.0 vs mean 5.9, SD 2.2; d=−0.63, 95% CI −0.82 to −0.44; t₄₉₈=−6.75; P<.001). They also reported more financial strain (mean 17.9, SD 6.3 vs mean 16.2, SD 6.6; d=0.27, 95% CI 0.09-0.46; t₄₈₆=2.89; P=.004), depressive symptoms (mean 8.6, SD 6.3 vs mean 7.4, SD 6.1; d=0.19, 95% CI 0.01-0.38; t₄₉₈=2.07; P=.04), and loneliness (mean 6.0, SD 2.1 vs mean 5.2, SD 2.0; d=0.40, 95% CI 0.22-0.59; t₄₉₉=4.33; P<.001), as well as lower resilience (mean 19.5, SD 5.1 vs mean 20.5, SD 4.3; d=−0.22, 95% CI −0.40 to −0.03; t₄₉₃=−2.30; P=.02).

Results were mixed with respect to substance use. Online recruits reported lower levels of past-year alcohol misuse (ie, binge drinking or heavy drinking, per National Institute on Alcohol Abuse and Alcoholism guidelines [27]; n=88, 27% vs n=66, 37%; OR 0.63, 95% CI 0.42-0.95; χ²₁=5.3; P=.02). However, online participants were more likely to report having used one or more recreational drugs on a monthly or more frequent basis in the past year (n=74, 23% versus n=17, 10%; OR 2.78, 95% CI 1.56-5.22; χ²₁=13.5; P<.001) and to report recreational cannabis use in particular within the past 30 days (n=80, 25% vs n=27, 15%; OR 1.82, 95% CI 1.10-3.07; χ²₁=6.1; P=.01).

Further results (eg, on tobacco product use) are available in Table 1. Notably, using the Benjamini-Hochberg procedure [28,29] to control the false discovery rate (or FDR) by limiting the proportion of the significant analyses in Table 1 due to false positives to 0.05, only 3 (11%) of the 28 (out of 39 total) significance tests became nonsignificant. Specifically, differences between recruitment groups on typical cigarette type (regular vs menthol), age of smoking initiation, and depressive symptoms had P values <.05 but were nonsignificant according to the FDR-adjusted results.

No adverse events were reported, nor were any stopping rules triggered during the trial.

Little’s MCAR test indicated that the data used in the exploratory hierarchical linear regression analysis were not MCAR (χ²₃₂=52.9; P=.01). To evaluate the sensitivity of the results to the handling of missing data, multiple imputation by chained equations was conducted, generating 20 imputed datasets using predictive mean matching [30]. The pooled regression results using the multiply-imputed datasets did not differ substantively on predictor significance or effect magnitude from those obtained using listwise deletion, and conclusions were unchanged. Thus, for parsimony and ease of interpretation, results presented below are based on the complete-case analyses.

Exploratory hierarchical linear regression analysis results (Table 2) indicated that, as a set, the demographic covariates explained a significant proportion of the variance in self-reported food insecurity (R²=0.098, 95% CI 0.040-0.136; F_7,467=7.21; P<.001). When recruitment modality was added as a predictor, the resulting model (ie, model 2) remained significant (R²=0.150, 95% CI 0.081-0.194; F_8,466=10.29; P<.001); the improvement in model fit was significant (F_1,466=28.86; P<.001). Critically, recruitment modality accounted for approximately an additional 50% of the variability in self-reported food insecurity above and beyond the demographic factors alone (ΔR²=0.053; percentile bootstrapped 95% CI 0.019-0.100; BCa bootstrapped 95% CI 0.020-0.101). Briefly, in model 2, age (b=−0.03, 95% CI −0.05 to −0.01; P=.003) and having an annual household income ≥$20,000 (b=−0.52, 95% CI −0.91 to −0.14; P=.008) were significantly associated with lesser food insecurity. In contrast, female sex (b=0.49, 95% CI 0.06-0.92; P=.03) and being recruited online (vs in person; b=1.22, 95% CI 0.78-1.67; P<.001) were significantly associated with greater food insecurity.

This secondary analysis examined differences in baseline sociodemographic, psychosocial, and smoking-related characteristics of participants recruited in person and online to a smoking cessation RCT for people with food insecurity and compared differences in recruitment pace and enrollment throughput across in-person versus online recruitment methods. Results indicated that participants recruited in person at food distribution events in West Central Florida differed in important and sometimes paradoxical ways from those recruited online via social media. Given our use of these 2 different modalities at different points in the study, a secondary objective was to examine the relative speed of in-person versus online recruitment. Following relatively slow in-person recruitment, switching to an online, national recruitment strategy was associated with a substantially higher enrollment rate; however, other concurrent differences between the strategies and their consequences, such as how in-person and online recruits became aware of the study (ie, face-to-face interactions with study personnel vs online via advertisements on social media or internet search results) and geographical reach (regional vs national) are important to consider.

To our knowledge, this study provided the most comprehensive comparison of the personal characteristics of participants recruited via different modalities to a smoking cessation RCT for a special population of any kind. The comparison revealed noteworthy sociodemographic similarities between groups, including relationship status, living in a household with another person who smoked, and employment status. However, there were also important group differences. In-person recruits had lower household income and educational attainment. They were also significantly older, more ethnically diverse, and more likely to be male. Despite having lower socioeconomic status, participants recruited at in-person food distribution events reported lower levels of loneliness, greater resilience, fewer depressive symptoms, higher subjective social status, and lower levels of perceived financial strain (despite having lower incomes). These protective characteristics [31,32] may have reflected greater engagement in their communities [33] and receiving services from a trusted community organization (ie, Feeding Tampa Bay) [34]. They could also represent self-selection into receiving on-site food assistance services by people who already felt a relatively strong connection to their communities [35]. That said, it remains possible that regular in-person engagement with community-based resources may have engendered a sense of belonging and support [36]. Furthermore, the additional resources to address food insecurity provided by food banks (above and beyond all other possible sources of food aid, such as EBT or food stamps) may have resulted in the in-person recruits having more food available on average than the individuals recruited online, potentially helping to explain the considerable gap in food insecurity between the recruitment groups [37].

It is also notable that relative to online recruits, in-person recruits scored more favorably on a variety of characteristics associated with smoking cessation success. These included considerably higher levels of food security [5], lesser financial strain [38], higher subjective social status [39], better psychosocial functioning [40,41], and greater readiness to quit smoking [42]. An important caveat is that although in-person recruits were less likely to report using recreational drugs, they were more likely to endorse alcohol misuse and more likely to consume noncigarette tobacco products (eg, cigars). In keeping with vaping’s greater frequency among younger cohorts [2], however, in-person recruits reported both having tried vaping and current vaping less frequently than online recruits. Finally, the exploratory hierarchical linear regression analysis suggested that the observed difference in food security between the groups was not solely a byproduct of contrasting basic demographic characteristics. The results were consistent with the possibility that differences in behavioral and psychological characteristics between the groups might help explain the sizable difference in food insecurity observed between the groups. Although the analyses could not establish causal pathways, the in-person and online recruits nonetheless differed considerably in terms of risk and protective factors found to predict successful smoking cessation in previous studies.

On average, individuals recruited online may have been more technologically adept, comfortable with digital platforms, and willing to engage with smartphone-based care than individuals recruited in person, who had not proactively self-selected into screening for participation via a personal electronic device but rather had been approached at a food bank and invited to complete the screener via tablet [43,44]. Comfort and skill with technology might not only be relevant to recruitment outcomes but also might have downstream consequences for intervention uptake, follow-up completion, and cessation outcomes [45,46].

Taken together, findings suggest that a 2-pronged recruitment approach involving both in-person and online methods may be needed to effectively target the wide spectrum of individuals seeking food assistance. It is possible that differences might not have been observed if online recruitment had been conducted only in the same geographic area that in-person recruitment took place [47]. Thus, online recruitment may not always serve as a complete substitute for “boots on the ground” approaches, depending on researchers’ goals and priorities. Furthermore, results from the pilot trial on smoking cessation treatment recruitment at food pantries in Greater Cleveland suggest that in-person recruitment at food distribution events resulted in a slightly higher proportion of people seeking food assistance who enrolled. Nonetheless, the enrollee yield rate in that study was still very modest [11]. Given the divergence in the personal profiles of the in-person and online recruits, it could be beneficial to use stratified or adaptive recruitment methods to improve representation of groups of interest recruited via a given method. For example, recruitment modality quotas could be prespecified or online advertisement spending reallocated toward underrepresented subgroups in real time.

Intriguingly, although participants recruited in person and online were equally likely to screen eligible, the probability of enrolling in the study was much higher for individuals recruited online, suggesting a need for future research designed to carefully evaluate differences in recruitment yield by recruitment strategy. For example, the relatively private, impersonal online approach might encourage some otherwise reluctant individuals to enroll [48], while others might more effectively be persuaded to enroll via in-person interactions with study personnel [49]. Learning which approach is likely to be more effective for different groups of people who smoke could improve both recruitment speed and the reach of future smoking cessation trials and programs.

First, it should be noted that in-person approaches and online advertisement clicks represent different stages of study exposure and intent to participate (staff-initiated initial contact vs self-selected engagement). Thus, comparisons of the percentages of individuals recruited via in-person versus online modalities who enrolled should be contextualized in terms of those differences, although the conversion percentages become more directly comparable later in the enrollment process (eg, enrollment rates among those who provided consent to participate and completed the baseline questionnaire). Moreover, possible geography-based differences between in-person and online recruits included overall rurality, state and local smoking ordinances, and the prevalence of smoking in participants’ immediate environments [47]. Furthermore, online and in-person recruitment occurred during different periods, which might have influenced both receptiveness and observed differences. Future researchers could avoid related concerns by recruiting via both modalities simultaneously and, if feasible and desired, ensuring that online and in-person participants live in the same region (eg, a cancer center’s catchment region).

Contextual priming effects (eg, answering questions in the food bank environment) [50], social desirability bias (eg, wanting to please study staff) [51], and differential modes of survey administration (eg, using personal phones vs tablets provided by study staff to fill out questionnaires) were also possible [43,44]. Nevertheless, mode of survey administration was not systematically recorded, and some participants filled out the baseline questionnaire while at the food bank (albeit in a private area); thus, contextual priming, social desirability, and survey administration mode–based effects cannot be conclusively ruled out as alternative explanations for observed psychological differences between recruitment groups. Future researchers could record device type and administration context (eg, on-site with staff administration, on-site without staff administration, or offsite) to directly test for such effects.

Finally, given the large sample size involved and the number of comparisons conducted, some significant differences might have been of limited practical importance [52] or due to random chance [53]. On the other hand, key differences between the groups (eg, age) had effect size CIs indicating that the magnitude of the group difference was medium to large, in addition to having very low P values, and the FDR-adjusted analyses for Table 1 suggested that nearly all significant results were not chance occurrences resulting from having conducted multiple comparisons.

Prior to this secondary analysis, differences in the baseline characteristics of participants enrolled in person versus online, and differences in recruitment pace across methods, had not been examined in a smoking cessation RCT for people with food insecurity. Nonetheless, it is important to stress that this study is a descriptive comparison within a specific trial context, rather than strong evidence for the comparative advantage of one recruitment modality over another. Regardless, using both recruitment strategies when conducting smoking cessation intervention trials for people with food insecurity may broaden reach, although, as seen in this RCT, the pace of enrollment might differ considerably between the 2 approaches. Additionally, the use of both online and in-person strategies might yield a sample that more closely reflects the underlying population of individuals who smoke and are food insecure. This study expands the RCT recruitment strategy evidence base by providing concrete results to inform researchers designing and conducting smoking cessation trials for special populations, particularly people with food insecurity. Furthermore, examining whether treatment effects and follow-up assessment completion status are influenced by recruitment modality will be important to understanding the potential substantive implications of recruiting from this population via online versus in-person methods.