One of the most challenging issues in the design of case–control studies is the selection of an optimal comparison group that minimizes selection bias, while allowing cost-efficient and feasible recruitment. The overall goal of control selection is to obtain a comparison group that is representative of the source population of cases, with two basic rules: controls should be selected from the same source population where the cases come from and should be selected independently of their exposure status [1].

A commonly accepted method to select controls is random digit dialing (RDD) by which households in the target area are randomly called to identify potential eligible participants. However, new technologies such as the increasing use of cell phones, answering machines, and caller ID in people’s homes, which allows them to screen calls, have contributed to lower response rates to RDD methods [2,3]. Furthermore, area codes are no longer accurately representing area of residence [1]. Therefore, we can no longer assume that the use of RDD for control identification results in a random sample of the general population or that it will avoid selection bias in case–control studies. In fact, studies have shown that the use of RDD leads to lower participation rates [4-6] and that participants recruited through RDD tend to be more educated than the general population [4,7,8].

Despite well-documented health disparities for African Americans, their recruitment into research studies has been traditionally challenging [9]. Moorman et al. [10] encouraged the sharing of successes and pitfalls in recruiting attempts among epidemiologists to improve recruiting and perhaps to help develop alternative methods of selecting controls. In this manuscript, we describe our experience recruiting community controls and the methods used. We also compare characteristics of community controls to those identified through RDD, and discuss potential selection bias and representativeness of the two populations.

The distributions of selected socio-demographic characteristics and known or suspected breast cancer risk factors for community and RDD controls are shown in Table 1. As expected, community controls were, on average, less educated, had lower income, more likely to be single, and have Medicaid or Medicare as a form of insurance than RDD controls. They were also slightly more likely to be uninsured, but the number of women without health insurance was small in both groups. Community controls were more likely to be obese, defined as a body mass index (BMI) greater than 30 kg/m²;

this was particularly true for premenopausal women (54.2% vs. 39.5%, P=0.04). While the percentage of nulliparous women in the two groups was similar, community controls tended to have more children and have their first child at a younger age; they were also less likely to breastfeed their children. When we compared FFQ data from the two control groups (Table 2), we found no significant differences in food consumption between the two controls groups.

Because of our concern that controls identified through RDD may have higher education and income than the general population, we compared our study population with census data to evaluate the representativeness of both control groups. As shown in Table 3, the combined community and RDD controls appeared to represent the general population better than RDD or community controls alone, at least with respect to education, income, and marital status, characteristics that were readily available in census data. When we compared the percentage of obese women in our study with that reported in NHANES for AA women (Figure 1), the same was true. Compared to NHANES data for Black women, the percentage of obese women was lower for RDD and higher for community controls, suggesting that the combination of the two control groups may be a better comparison group.

We also evaluated the potential differences in risk estimates when using only RDD controls and when including community controls as well. As shown in Table 4, risk estimates tended to be of similar magnitude including and excluding community controls for reproductive variables, and their inclusion also improved the precision of the estimates. For obesity, risk estimates were in the expected direction consistent with the literature only when including community controls.

Recruitment of controls remains a challenge in case–control studies. Because the goal is to obtain a comparison group that is representative of the source population of cases [1], RDD has been historically considered the best method to recruit controls in population-based studies. However, recent changes in technology have made this method less efficient and concerns have been raised about potential biases introduced by RDD sampling [6]. Minorities are underrepresented in research studies in general, but there is particularly a special need for more data on cancer in

minority populations [9]. This is particularly true for breast cancer, for which evidence in AA women is limited [17,18]. Therefore, exploring new methods to improve minority recruitment, and for control selection, is particularly important in breast cancer studies of AA women.

Because community recruitment uses non-random methods to identify controls, we were concerned that the use of this control group may introduce selection bias. We compared socio-demographic, reproductive, and lifestyle factors between controls recruited through RDD and community-based efforts, and found differences in some of these factors. For example, we found that RDD controls had higher socioeconomic status (higher education and income levels), which is consistent with what has been reported by others [7,8]. Importantly, education, income, and marital status distributions among the community controls and among the combined RDD and community controls mirrored the general population distributions for AA women according to NJ census data more closely than RDD controls alone. Furthermore, comparing case and control characteristics, the addition of the community controls did not substantially affect risk estimates, while providing more statistical power. The fact that the group of combined controls tended to be more representative of the source population lends more confidence to the estimates computed including the community controls. To further assess potential selection bias in analyses of data, we routinely compare results including and excluding community controls. If we were to find differences in estimates, we would present results both including and excluding community controls so that the reader could draw their own conclusions. Similar to the results we presented in this manuscript, results do tend to be in the same direction, while estimates are more precise given the larger population when including the community controls.

As noted earlier, one of the rules for an appropriate comparison group in case–control studies is that the controls are selected from the same source population as the cases [1]. We were reassured that the population that we were covering with RDD recruitment, case recruitment, and community-based recruitment came from the same source population when we received several calls from RDD controls who had actively or passively refused

and later changed their mind and wanted to participate in the study after seeing it advertised in their community, as well as from cases who refused earlier or were just diagnosed and contacted us to participate even before the NJ State Cancer Registry received their diagnoses. Another rule is that controls are selected independently of their exposure status [1], which should not be a concern in our community-based efforts, as all women meeting our study criteria were invited to participate and those criteria were unrelated to the exposures of interest.

Although investigators are able to calculate response rates for cases if the study is population based, or from hospitals where all pathology reports are available to identify eligible cases, finding a true denominator for the majority of eligible controls is very difficult. For RDD, the numbers of those who are called through RDD, say that they fit the eligibility criteria, and agree or do not agree to be contacted by the interviewer or study staff can be calculated. But does this really represent the true response rate? It is likely that a proportion of households do not answer the telephone when it is a number they do not recognize, or if they do answer, do not answer truthfully if there is a person in the household who meets the eligibility criteria. Thus, the ‘response’ rate is likely misleading. One also cannot document response rates from community participants because the true denominator is unknown. Perhaps it is time to reconsider issues of response rates for controls in this era when there is much less openness to unsolicited telephone calls, or willingness to participate in research studies. Particularly for minority participants, the best approach may be community events where there can be education and outreach, with attention to the populations from which these groups arise, and if they are likely to represent the same population as cases.

The problems with finding the ideal control group may appear to make case–control studies less attractive, with attention turning to cohort studies. However, for studies of rare cancers, or in special populations, case–control studies may be the best way to find answers to pressing questions. More importantly, case–control studies may allow for more in-depth translational research, with capabilities to rapidly identify cases, and to obtain tumor tissue blocks and medical record and treatment information in an easier, more timely fashion. As we learn more and more about the heterogeneity of various cancers, having the ability to examine tumor tissues for molecular characteristics will allow for better classification of disease and therefore, clearer associations regarding etiologic factors.

There are limited data evaluating alternative methods to recruit minorities in epidemiologic studies. In agreement with our study, Cabral et al. [19] found RDD to be extremely labor intensive and not efficient to recruit AA and Latino controls in a case–control study of lung cancer, and explored community-based methods, similar to ours. Similar to our experience, they found the community-based control group to be more representative of the San Francisco Bay Area AA population than the RDD control group or the HCFA (Health Care Financing Administration) control group which was limited to controls older than 65 years.

Several barriers to minority participation in research studies have been identified including historical mistrust, economic, socio-cultural, knowledge and accessibility issues [9]. Strategies that have been proposed to increase minority recruitment in research studies include involving members of the target population in recruiting efforts [9,20], addressing misperceptions that potential participants may have about research [21], using face-to-face outreach and recruitment methods [21,22], offering monetary incentives, showing respect for individuals, appreciating cultural differences, and “giving back” to the community [23]. We adopted all these strategies in our community-based recruitment efforts, and our approach proved to be better for recruiting minority populations than RDD or other methods. Community-based recruitment was also cost-efficient, as it allowed us to use our existing resources and personnel, rather than paying an external firm to identify RDD controls. Moreover, potential participants approached us with interest in participating (rather than some commercial firm contacting them about the study), due to which they were much more likely to actually do the interview.

In our experience, community-based recruitment was an efficient and feasible method to recruit minorities, especially when combined with RDD recruitment. However, with the growing issues regarding success of telephone contact for participation in studies, the utility of RDD will need to be monitored going forward, and may prove too expensive and less efficient for optimal recruitment of controls. Although successful for us, particularly in the AA community, where churches and community groups tend to be important, we are aware that community-based recruitment may not work in certain populations (for example, in rural underdeveloped areas) or in multi-site studies where standardization of community strategies would not be possible. Therefore, recruitment strategies should be tailored to the target population, while taking into account resources.

In our study, community-based recruitment helped us stay in touch and “give back” to the community from which we were recruiting, while supplementing the RDD controls to achieve a comparison group more representative of the general target population. We learned that for community efforts to achieve an unbiased comparison group, strategies have to be tailored to the community being recruited. Therefore, partnering with community agencies, cancer support groups, community members interested in the topic, and particularly with churches and church leaders, is crucial for the efforts to be successful. Furthermore, recruiting efforts should occur in several fronts (churches, health events, clinics) so that different demographics are represented, and in a variety of locations representing the target population. Importantly, characteristics of the participants and responses to interview data should be closely monitored in ‘real time’, so that the appropriateness of the controls in relation to cases in the study can be frequently appraised and adjusted, if necessary. From our experience, however, community-based recruitment appears to be a cost-effective method to recruit AA controls in case–control studies.

AA: African ancestry; CI: Confidence interval; EA: European ancestry; NJ: New Jersey; NYC: New York City; OR: Odds ratio; RDD: Random digit dialing; WCHS: Women’s Circle of Health Study

The authors declare that they have no competing interest.

EB and CA designed the study. EB and CA and GC directed its implementation. EB and UC designed the analytic strategy. UC obtained NJ Census data. UC and GZ conducted statistical analyses. SM provided expertise in the interpretation of results. EB wrote the first draft of the manuscript and all co-authors provided substantive comments and editorial review. All authors read and approved the final manuscript.

The pre-publication history for this paper can be accessed here:

http://www.biomedcentral.com/1471-2288/13/71/prepub