Timely and low-cost outbreak investigation thanks to panel controls

The majority of panellists responded to the online questionnaire in less than a week, ensuring a higher response rate than traditional methods such as random digit dialling [Reference Schonlau9, Reference Craig13]. The recruitment of panellists as controls also required far less staff and financial resources than other approaches, as it amounts to sending an email rather than conducting numerous face-to-face or telephone interviews [Reference Baker12, Reference Hays, Liu and Kapteyn14]. In the event of an actual food or waterborne outbreak, the HuGO panel offers a more timely and cost-effective control recruitment and analysis, which would lead to timelier public health actions to limit additional cases (e.g. trace-back investigations, recall of products). It is nevertheless important to consider that the constitution of the HaBIDS panel, from which the HuGO panel stems, did require substantial resources. Indeed, the 2.379 panellists were recruited thanks to the dedication of a 1–2 researchers, who drew 16.000 people from the registries of Lower Saxony and send them invitation letters per post therefore reaching an overall response rate of 8.9% [Reference Rübsamen26].

Suitability of panel controls for outbreak investigation

The study assessed the suitability of panel controls for case–control studies. Ideally, one would be confident using panellists as controls if these were not more biased than controls used in current best practices. Controls are used to estimate the prevalence of exposures in the population that gave rise to the cases. They are expected to be free of the outcome of interest, representative of the population at risk of the outcome and selected independently of the exposure of interest [Reference Grimes and Schulz28]. In the situation of an outbreak investigation, case–control studies aim to estimate a particular exposure-disease association, while appropriately controlling for confounding and avoiding other biases.

Selection bias is a particular problem inherent in case–control studies, where it gives rise to non-comparability between cases and controls, threatening the generalisability of the study results [Reference Waldram2]. Our study is affected by a sampling bias, as HuGO panellists volunteered to participate in the study. First, HuGO panellists differ from the general population of Lower Saxony in terms of sociodemographic factors such as age, sex and education level. We reduced this bias by matching the frequencies between historical cases and panellists on sex and age, therefore controlling for the possible confounding effects of both variables. Additionally, controls recruited through traditional methods are also rarely representative of the general population [Reference Waldram2]. It is largely accepted that a lack of representativeness on socio-demographic characteristics does not hamper scientific inference [Reference Richiardi, Pizzi and Pearce29, Reference Rothman, Gallacher and Hatch30]. Valid scientific inference is achieved if the confounders are controlled for, and there is no reason to believe that control of confounding can be more easily achieved in a randomly selected control group that in a panel group [Reference Richiardi, Pizzi and Pearce29]. Second HuGO panellists are significantly more educated than the general population of Lower Saxony. As studies have shown that a higher education status is associated with healthier eating habits [31], it is likely that HuGO panellists may be more health-conscious than the general population [Reference Erens7, Reference Baker12].

Using the HuGO panellists as controls decreased the likelihood of information bias as compared to using traditional controls, resulting in a greater validity of the food exposure information. First, there is little recall bias as panellists responded within a few days to the questionnaire. As traditional controls are often recruited with days or weeks of delay after the outbreak occurs, they have more difficulties to remember their food exposures at a particular time point. Panellists would then be particularly useful when investigating outbreaks caused by uncommon food exposures, as they are oftentimes more forgotten than common food exposures [Reference Inns32].

Additionally, online surveys are less likely to suffer from social desirability bias as other data collection methods, such as phone or face-to-face interviews [Reference Rübsamen26]. Panellists are less likely to underreport ‘bad’ food items, and over report ‘good’ food items than traditional controls. They would therefore also be particularly adequate to investigate sensitive exposures, benefitting from the survey anonymity [Reference Baker12, Reference Rübsamen26].

Reliability of control panels to identify outbreak sources

The study explored the circumstances in which using panellists would be reliable to investigate food and waterborne outbreaks. Indeed, investigators expect consistent conclusions regarding the likely outbreak source, regardless of the study design they use for the investigation.

In the present study, three of four outbreak sources could be identified using panellists as controls (studies B, C and D). The differences in the magnitude of the effect estimates (odds ratio) between historical and HuGO studies did not affect the ability of panellists to successfully identify the outbreak sources. As odds ratios measuring associations between contaminated food items and the disease are usually very high during food and waterborne outbreaks [Reference Inns32], we argue that they are not significantly affected by differences in exposure proportions between panellists and traditional controls.

The study shows that in outbreak A, the association between the consumption of tap water and gastrointestinal disease could not be identified with a statistical significance by using panellists as controls. We hypothesise that, in this particular situation, panellists did not have the chance of being exposed to the very localised outbreak source (contamination of water supplies with surface water due to heavy rainfall), and were thus not representative of the population at risk of the outcome. They did not live in the same area as the cases, and were not questioned at the particular time during which the waterborne outbreak occurred. This result furthers the argument that the reliability of panellists is likely to be higher when used as controls in the investigation of region- or nationwide outbreaks (like outbreaks B, C and D) as compared to very local outbreaks (like outbreak A). On the other hand, historical controls were likely to have changed their drinking habits, and drank significantly less tap water than usual, as they were interviewed around 14 days after the first cases developed symptoms. At that time, information related to the contaminated tap water was already communicated in the news, with a recommendation to rather drink bottled water.

Strengths and limitations

The main strength of this study is that it is a proof of concept for a promising method to recruit controls for case–control studies. Panellists can be asked specific questions about particular products they ate, ways of cooking, places they shopped at etc., as soon as an outbreak occurs, enabling a quick investigation. The present study participates in the efforts to provide local public health professionals with innovative methods to empower them in the conduction of outbreak investigations [33]. The study also assesses likely bias of using panellists as controls in the context of outbreak investigations. It supports the findings of other studies that the risk of bias must be assessed anew when a study is conducted to investigate an outbreak [Reference Mook11]. Yet overall, the risk of bias when using the HuGO panel should be smaller in situations where the cases seem to resemble the general population (with a tendency of higher education because this is what the panel is composed of). Another strength of the study is the oversampling of the control group. This allows the selection of a subset of more valid controls to match the frequencies of cases in terms of sex and age, thereby addressing confounding and some of the bias introduced by using panellists as controls. Such an approach is here possible given that there is no additional cost per additional questionnaire.

The main limitation of the study concerns the historical studies. Panellists could only be used when population-based controls were used in the historical studies, but not in specific settings (e.g. outbreaks during a party or at a restaurant). This strongly reduced the choice of historical studies for this analysis. The available studies had limited information on some key exposures and multivariable analyses could therefore not always be conducted. Additionally, the panel is affected by a selection bias. As this cannot be controlled via study design, it is therefore important to consider whether the measure of effect for a particular exposure may have been due to such an inherent sampling issue. In the situation of a prospective outbreak investigation, socio-demographic data could easily be collected among cases and panellists, and a weighting method could minimise this bias (e.g. frequency matching, propensity scores, quotas) [Reference Craig13].

Perspectives

The study provides encouraging results and warrant further exploration to prove the validity of using panellists as controls in case–control studies. First, we will recruit more panellists (including children), in order to increase the sociodemographic diversity of the panel and its suitability to investigate the upcoming outbreaks. Then, as soon as a likely food or waterborne outbreak occurs in Lower Saxony, we will conduct two parallel case–control studies in which controls would be selected either through a traditional method or from the HuGO panel. The results of both studies will be compared. Propensity score matching will be used to reduce selection bias [Reference Natilli, Romano and Leone34, Reference Randolph35].

Finally, in order to assess in which scenarios or for which hypothesised food and water exposures or behaviours using panellists as controls might be more appropriate than traditional controls, any selection bias introduced by using the panel need to be better understood. A prospect is to compare the reported food and water exposures and behaviours of panellists with that of probability samples from population-based food exposure surveys, such as with the German food exposure survey [Reference Heuer36, 37] or other sources of such data for which selection biases are minimised or previously characterised [Reference Rosner38].