The present analysis was performed on data from participants in the NutriNet-Santé Study enrolled from May 10, 2009 through December 31, 2011. The rationale, design, and implementation of the study has been reported elsewhere [20]. Briefly, the NutriNet-Santé Study is an ongoing, large, Web-based prospective cohort study launched in France in May 2009. Its primary aim is to investigate the associations between nutritional factors and health outcomes and to elucidate the role of various determinants (eg, demographic, socioeconomic, cultural, and cognitive) of dietary patterns and nutritional status.

In the NutriNet-Santé Study, adult participants aged 18 years and older are recruited from the general population on a voluntary basis. All data are collected through a dedicated website [21] via adapted questionnaires using a secure and user-friendly HyperText Markup Language (HTML) interface. Individuals who fill out all baseline questionnaires (pertaining to sociodemographics, dietary behaviors, physical activity, anthropometrics, lifestyle, and health status) are included in the cohort [20]. All baseline questionnaires were first pilot-tested and compared with the traditional modalities (paper versions or dietitian interviews) [22-24]. Health events are monitored via questionnaires about hospitalizations and medication use as well as via a linkage with the national vital statistics database. The study was approved by the Institutional Review Board (IRB) of the French Institute for Health and Medical Research (IRB Inserm no: 0000388FWA00005831) and the Comité National Informatique et Liberté (CNIL no: 908450 and 909216).

Recruitment is scheduled for 5 years with an additional follow-up planned at 10 years. The call for volunteers is based on a vast, biannual, multimedia campaign (including print media, Internet websites and networks, and television and radio broadcasts).

At the launch of the study, the first campaign contained general information about the study and was carried out under the auspices of the Health Minister. Subsequent campaigns have been built around specific findings from the study (eg, calls for reduction in salt consumption, increase in fiber intake, adherence to nutritional guidelines). Before each campaign, a press conference is organized and a press release is widely disseminated within a 72-hour period. Aspects related to academic endeavor, public health interest, scientific progress, confidentiality, and convenience regarding participation (ie, ≤20 min each month) are emphasized.

Additional dissemination strategies have also been used, such as short message service (SMS) text messaging for the initial mass media campaign especially (n=400,000 recipients), leaflets/flyers (n=50,000 recipients), and a nonpaid advertising display. For example, the SMS text messages were sent by a French mobile service provider to a random sample of their customers. Nonpaid advertising displays included postings in doctor’s offices, health centers, worksites, public transportation, and on billboards, as well as video clips shown in post offices.

The baseline set of questionnaires includes information about date of birth, sex, area of residence, education, employment status, household composition, occupational category, and income. At 1-month follow-up, participants are asked to provide information about how they heard about the NutriNet-Santé Study and are given 18 response options including “I don’t remember” or “other.” Each participant is allowed to select only 1 response to this question. For the present analyses, these responses have been grouped into the following 9 recruitment sources: (1) general newscast or a health program on television, (2) radio newscast, (3) newspaper article, (4) Internet website or network, (5) SMS text message sent by a service provider, (6) advice from a friend/family member/health care provider, (7) leaflet/flyer, (8) nonpaid advertising display, and (9) other (including “I don’t remember”).

For the present analysis, we focused on participants included in the NutriNet-Santé Study between May 10, 2009 and December 31, 2011. From the 104,020 participants meeting that criterion, we selected 88,238 individuals who provided information at 1-month follow-up on how they had learned about the study.

Income per household unit was first calculated and then divided by the number of household members (eg, 1 unit for the first adult in the household, 0.5 units for the other persons aged 14 years or older, and 0.3 units for children under 14 years of age) [25]. The occupation category for retired and unemployed people was defined as the last job held.

Sociodemographic characteristics of the sample, including age (<30 years, 30-45 years, 45-60 years, ≥60 years), education (primary/secondary, postsecondary <bachelor’s degree, ≥bachelor’s degree), current employment status (employed, unemployed, homemaker, retired, student), occupational category (never employed, farmer, manual worker, employees/medium-skilled labor, intermediate professions/skilled labor, self-employed, managerial/professional staff), type of area of residence (rural, urban <200,000 inhabitants, urban ≥200,000 inhabitants), and monthly household income (<€1200, €1200-€2700, €2700-€3700, >€3700 per household unit), are presented in a frequency/percent format for the entire sample and by recruitment source.

To better understand the selected sample, we compared the characteristics of included and excluded NutriNet-Santé Study participants using chi-square tests and Student t tests, as appropriate.

Crude and adjusted associations between the sociodemographic characteristics and the recruitment source were estimated using polytomous logistic regression (reference=television). Polytomous logistic regression generalizes the binary logistic regression model, allowing the dependent (outcome) variable to have more than 2 categories [26]. The multivariate model included the following covariates: age, sex, education, employment status, occupational category, area of residence, and monthly household income. Additionally, the model was adjusted for the interval between the most recent press release and the completion of the first follow-up questionnaire that included information about the recruitment source. Odds ratios (OR) and 95% confidence intervals (CI) are reported. We also performed a sensitivity analysis after exclusion of participants for which the most recent press release occurred between baseline and the first follow-up questionnaire. Tests of statistical significance were 2-sided and the type I error was set at 5%. Statistical analyses were performed using SAS software (version 9.2, SAS Institute Inc, Cary, NC, USA).

We examined how the sociodemographic profiles of participants in a large, Internet-based cohort differed according to the recruitment source used. The use of a number of different dissemination channels allowed the recruitment of a sizeable and diverse cohort. As expected, most participants enrolled after hearing about the study on television because this medium has the widest reach. In particular, television announcements permitted the recruitment of a larger proportion of members of population subgroups (eg, those belonging to lower socioeconomic strata) that are not typically well represented in population-based epidemiologic research. However, the elderly were more likely to be informed about the cohort through channels other than television suggesting that television did not necessarily represent the best information medium for all population subgroups. In turn, radio, newspapers, Internet, and personal advice also proved to be substantial means of disseminating information about this epidemiologic study to encourage participation.

Overall, the recruitment of participants from a wide range of sociodemographic backgrounds provides the study with a broader range of exposures and confounders than does a more homogeneous sample. This is of important concern when the overarching aim is to estimate associations between certain exposures and health/disease outcomes, as in the NutriNet-Santé Study [1,13,27,28].

Among the initial efforts to recruit Web-based cohorts, in examining the literature we identified 1 study calling for participation in a cohort of smokers intending to quit, by using existing Internet panels intended for market research [12]. Another study recruited nurses and midwives through targeted email distribution [14]. A birth cohort study employed hospital advertising [15] and a pregnancy planning study made recruitment announcements on a health-related website [13]. To the best of our knowledge, this is the first study pertaining to the quantitative description of participant profiles according to the recruitment source.

The main limitation of this analysis was the lack of information on participation/refusal rates because the call for participation was not delivered to a predefined and exhaustive list of randomly selected individuals. In turn, the results need to be interpreted in light of the fact that the actual effectiveness of the media channels cannot be estimated. Second, misclassification of recruitment source is also possible because the question about information source was asked 1 month after baseline. Additional misclassification about the recruitment source cannot be ruled out because participants may have been exposed to several information sources. For example, if a press release occurred during the interval between baseline and the completion of the first follow-up questionnaire, this may have led to an overreporting of the related information channels. However, the sensitivity analysis performed after removing these participants did not substantially modify the findings.

Concerns have been raised regarding the selection bias potentially inherent in cohorts followed via the Internet. Web users may present particular sociodemographic profiles, especially regarding sex, age, and education [3,10,29]. Our findings suggest that vast media campaigns coupled with expanding Internet access permit reaching individuals from a wide range of sociodemographic backgrounds.

As compared with national estimates [30-32], the NutriNet-Santé Study sample included proportionally more women (77.60% vs 52%) and individuals of relatively high socioeconomic status (managerial/professional staff excluding retired people: 31.64% vs 9.3% nationally; postsecondary education: 67.92% vs 24.3% nationally). This is consistent with existing knowledge regarding the characteristics of participants in volunteer-based studies dealing with health and nutrition [1]. Any additional selection bias is likely negligible given the wide range of recruitment channels used and the widespread access to the Internet in France, with a current penetration rate of 77% [33]. For example, 27% of the Web users in France are older than 50 years, which is relatively close to the respective proportion of individuals in that age range in the NutriNet-Santé Study (36%).

As suggested in the literature [28], the response rate and the related potential lack of representativeness, as observed in our study, is not critical in an etiologic context. However, this remains a key issue when descriptive information is provided. In the present study, 1 of the most important aspects pertained to the sizeable and heterogeneous cohort, including a wide range of sociodemographically diverse profiles, thus ensuring material variability in exposure factors. The use of various recruitment channels allowed meeting this objective by including specific subgroups of the population, such as those older than 60 years, those with low income, and/or with low levels of formal education.

It has also been recently postulated that in the context of a Web-based cohort of volunteers, the generalizability of etiologic findings depends on whether the studied associations might differ between those with and without Internet access, which seems unlikely [13].

Along with being geographically unrestricted, Web-based prospective cohorts present with many logistic and cost advantages compared to traditional samples [3,4,11,22,34,35]. Furthermore, use of the Internet entails lower social desirability effects, which could also facilitate the recruitment of participants exhibiting unhealthy behaviors or those with socially undesirable/stigmatizing backgrounds.

In conclusion, the present findings fill gaps in current knowledge about cohort recruitment by providing new insights about the sociodemographic profiles of adult volunteers. Our Web-based cohort study uses recurrent mass media campaigns and numerous public outlets for encouraging participation. The various information channels allow the inclusion of individuals with diverse sociodemographic profiles. As expected, television appeared to be the most promising channel for reaching potential participants representing a wide range of sociodemographic backgrounds, including subgroups that are usually difficult to involve in long-term epidemiologic studies. Our findings could inform future Web-based studies regarding the development of operative targeted or general population recruitment strategies.