Factors associated with recruitment, surveillance participation, and retention in an observational study of pregnant women and influenza

Prospective, community-based studies of the epidemiology of infectious diseases and the effectiveness of treatment and prevention strategies including vaccination are rare, in large part due to their high costs and many logistical challenges [1, 2]. Yet, observational studies are essential for understanding the clinical presentation of infectious diseases, transmission dynamics, and the effectiveness of interventions, in particular for populations at high-risk of morbidity and mortality such as pregnant women. However, there is a dearth of information on how best to conduct cohort and active surveillance studies among pregnant women [3‐5].

This lack of information was especially evident to us in 2009 as we planned a two-year observational study of influenza vaccine effectiveness (IVE) in preventing laboratory-confirmed influenza illness during pregnancy [5]. Although we successfully completed the study and have previously reported IVE [6] and other findings on the epidemiology and prevention of influenza during pregnancy [7‐10], many of the planning assumptions we made about cohort recruitment and surveillance participation [5] were proven wrong. Therefore, after completing the main study objectives, we returned to take a closer look at how pregnant women responded to our study requests and consider possible sources of selection and information bias.

This report describes the results of our recruitment efforts and the subsequent participation of pregnant women in study activities, including a 6-month post-delivery follow-up interview. The main project consisted of three sub-studies. In both years, we recruited pregnant women following medical encounters for acute respiratory illness (ARI) during influenza season; the women in these sub-studies are referred to as medical enrollees (year one) and medical enrollees (year two). Because of uncertainty regarding where and how often pregnant women sought medical care for ARI, we also recruited pregnant women prospectively into a cohort prior to the initiation of influenza season in year one; this also gave us an opportunity to examine a broader range of illness severity. Participants in this part of the study are referred to as cohort enrollees (year one).

Because our study took place within two integrated healthcare systems with longitudinal electronic medical record (EMR) data, we had the unique opportunity to identify the demographic and health characteristics of women who were eligible for our study (as either a medical or cohort enrollee) but could not be reached for recruitment, as well as women who were contacted but refused participation. Specifically, we examined possible indications of selection bias, such as systematic differences in age, study site, race, ethnicity, underlying health status, and medical visits. Given the focus of the original study on IVE, we were particularly interested in whether receipt of influenza vaccination differed between participants and non-participants and between the cohort vs. medical enrollee sub-studies. We also examined the associations between these factors and participation in weekly surveillance (for cohort enrollees in year one) and completion of a follow-up interview (among all enrollees in both years).

We had greater difficulty enrolling cohort participants and conducting active surveillance than we expected [5, 16]. We only reached half of the potential participants we tried to contact by telephone in year one, and only 1 in 5 of the women we invited agreed to join the cohort. Participants clearly differed from non-participants. We found statistically significant differences between enrollees, refusers, and women never reached on all ten of the characteristics we could obtain from medical records. One in 5 cohort enrollees never started their weekly surveillance reports, and among those who did participate, reports were only completed for about half of the weeks. Nonetheless, retention among all enrollees was very high (92%) for a 6-month follow-up interview. Here too, we noted statistically significant associations between surveillance participation, retention, and most of the socio-demographic and health characteristics we examined.

Race was a consistent predictor in all of our analyses. Non-white women were less likely to be reached and enroll in our cohort; once in the cohort, non-white women were less likely to start active surveillance, completed fewer weekly reports, and were less likely to complete the 6-month follow-up interview. This finding is consistent with previous observations of lower rates of completed contacts and research consent by non-white adults [17‐19]. However, in year two, when participation did not involve weekly surveillance, non-white women who we reached and invited to participate as a medical or control enrollees were more likely to consent to participate than white women. The difference we observed between years is consistent with reviews that note research participation is not always lower among non-white adults and may depend on study type and tasks [19, 20].

Although the differences we noted were statistically significant, many were of relatively small magnitude. Certainly, statistical differences in participation among subgroups do not necessarily imply that the study findings are biased. In our IVE study, we were most concerned about potential confounders: characteristics associated with both the exposure of interest (vaccination) and the outcome (laboratory-confirmed influenza illness). We previously reported that only gestational age at enrollment was associated with both the likelihood of vaccination and influenza outcomes in our final study sample [6]. However, we noted here that women who were not reached and those who refused cohort participation were less likely to receive the influenza vaccine during pregnancy and were in somewhat better health (as indicated by fewer chronic conditions, pregnancy complications, and medical encounters for ARI). Enrolling a sample that was somewhat less healthy than other pregnant women in the community would not necessarily change the IVE we observed; indeed, there is little evidence that socio-demographic factors (other than age) or underlying health conditions modify IVE [21‐23]. Nonetheless, enrolling a cohort with high vaccination coverage likely contributed to the fact that we observed much fewer influenza illnesses than we expected [5], and systematic differences in race and health may have biased our results in unpredictable ways [24].

Enrolling women following medical encounters for ARI proved to be an efficient way to increase the total number of influenza positive cases in our study. The IVE literature typically assumes that ARI patients differ from their full local population in their propensity for medical utilization and vaccination [21, 25]. Indeed, the test-negative design, which compares vaccination rates among influenza positive patients with that of influenza negative patients, is believed to minimize potential biases associated with healthcare seeking behavior by focusing on people who have sought medical care for similar symptoms [21, 25]. However, in study year one, we found that influenza vaccine uptake among medical enrollees was similar to cohort refusers and the large population of pregnant women never reached. In contrast, our cohort participants stood out as having very different influenza vaccine uptake; pregnant women in our cohort were 2-fold more likely to receive the influenza vaccine than medical enrollees even after adjusting for socio-demographic and health characteristics, indicating a higher levels of vaccination compliance among women who participated in the cohort study. This suggests that the medical enrollees in our previously published IVE estimates [6] may have been more representative of the source population of pregnant women in these healthcare systems than we assumed.

Enrolling women following medical encounters also proved to be an effective way in year one of our study to reach and enroll women who were non-white, Hispanic, in poorer health, and had children at home. In year two, non-white women were also more represented among the medical enrollees. This may have been due in part to the fact that we were able to reach a more diverse population of women when many were at home sick. Although we asked the cohort and medical enrollees to perform similar activities (including written informed consent, an enrollment interview, respiratory specimen collection), the commitment asked of medical enrollees was more immediate and clear, since they often completed the bulk of study tasks on the same day. In retrospect, we recognize that the scope of tasks involved in our cohort study, the uncertainty of what would be required (depending on if they became ill or not), and the extended time commitment, may have biased participation toward women who were more educated, had higher subjective social status, had fewer stressful events in the past year, already had health concerns, and were more likely to have future pregnancy and ARI issues.

Among the strengths of our study is the detailed report of participation rates which start with the denominator of all presumed eligible potential participants, as recommended by the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines [26], but rarely published as part of cohort study results. Because our study took place within integrated healthcare systems, we were also able to describe summary counts of the demographic and health characteristics of those never reached and cohort refusers; these characteristics are usually unavailable when studies exclude individual-level data on those who refuse research consent. Indeed, greater clarity in reporting on potential selection biases is essential to interpreting previous and future findings regarding the possible benefits of vaccines for pregnant women and their infants [24]. In our study, the association of participant engagement with poor health and influenza vaccination status would have been missed had we focused only on consented enrollees, since the differences we noted were only apparent earlier in the steps of contacting and recruiting participants.

Our study also has at least five limitations. First, we do not have information on the specific reasons why some women refused participation or did not complete surveillance tasks. Second, the extent to which our findings will generalize to other study settings and designs may be limited by the features of our study. For example, our study benefited from being able to offer small financial incentives for participation, from conducting home visits rather than requiring special trips to research facilities, and from having regularly updated contact and medical care information within the integrated healthcare systems. Third, we did not have access to complete information on Hispanic ethnicity or racial categories beyond white vs. non-white for all potential participants, which limited our ability to describe and understand how these factors influenced participation. Fourth, we noted differences in participation between study sites, which may have been due to different recruitment and staffing methods, population differences in race and ethnicity, or other reasons. In year one, study site one used local staff (with local telephone numbers) for recruitment and had higher enrollment and surveillance participation rates than site two which relied on an external contractor. Interestingly, in year two, when both sites used local staff for recruitment, the percentage of contacted women who enrolled was higher at site two. Finally, our surveillance methods did not include the use of short message service (SMS) texts, which have recently been demonstrated to be an effective way to conduct health monitoring during pregnancy (e.g., [27, 28]).

In conclusion, we observed systematic differences in socio-demographic and health characteristics across different levels of participant engagement and in comparing cohort vs. medical enrollees. The indications of selection bias we observed are a reminder that having a representative sample and a high participation rate remains critically important for ensuring valid conclusions in observational studies. Faced with a growing need to conduct observational studies of pregnant women due to new disease threats [29, 30] and new opportunities for immunization [31‐35] and treatments [36] during pregnancy, more methodological research and innovation in conducting prospective observational studies in this population is needed [16], especially when extended participant engagement and ongoing surveillance are required.