In this study we set out to develop a reliable method that may in future enable us to access the necessary data for the establishment of a micro health insurance in low income rural communities where data would not be available otherwise. The objective before us was to find a way to overcome the two constraints associated with data sourcing through household survey, namely, the cost and time required. The Illness Mapping method we describe here seems to meet this objective. The information given in Table 7 illustrates the advantage of the Illness Mapping method in terms of human resources and time required, which obviously translate into differences in costs (e.g. salaries, travel, accommodation etc.).
The design of an insurance product requires estimates of the prevalence/incidence of the events covered by the insurance. Our previous studies showed that: (i) the incidence of illness episodes, and prevalence of hospitalizations and delivery is strongly context-dependent and varies across locations even in the same country  making it necessary to obtain local data. (ii) Prospective clients of health insurance in rural India are exposed to hardship financing not only in cases of hospitalizations but also in cases of outpatient treatment and in deliveries . In fact, this is even more pronounced in case of chronic illnesses . (iii) When expressing their priorities regarding benefits that should be covered by insurance, prospective clients expressed a clear wish to include both inpatient and outpatient benefits [34, 37]. It is thus clear that the information obtained through Illness Mapping regarding the prevalence/incidence of prioritized cost generating events is essential for the design and pricing of context-relevant health insurance.
We followed a strategy of soliciting local information from groups rather than from individuals. We were inspired by group techniques, assuming that the small cosmos of a village community could be captured through harvesting the knowledge that is readily available to its inhabitants free of charge. Having failed to find a ready-made suitable method in the published literature, we opted to utilize a combination of established methods and adapt them to our settings. Group approaches such as the Delphi and NGT have been used successfully and with high accuracy for business forecasting as well as for public policy [38, 39]. We adopted the criteria for resourcing quantitative information from qualitative non-interacting groups such as Delphi and NGT [22, 26], and modified those to take account of the advantages of interactive group situations in which the discussions are moderated and facilitated rather than left to chance (as often happens in exploratory brainstorming groups or focus groups ). Such structured group methods are based on the principle of collective intelligence , or group intelligence that emerges through managed consensus decision making .
Our method was based on small group discussions with people who were marginally literate and numerate, but nonetheless experts or valid representatives of their village communities. They were chosen (with the help of our partner NGO staff who had prior access to the village) for their social attributes and their knowledge of households in their own neighbourhood in the village. In each village we carefully identified such participants and facilitated their interaction to obtain estimates for the prevalence of illness for the entire village. Other key contacts in the village such as teachers, village head, and health workers could also be recruited to provide similar information if there were no prior links with the village.
We organized gender homogenous groups in each village to ensure that both men and women would be able to express themselves freely. We thought that women, who are usually caregivers, might be more familiar with illnesses than men. However we found no statistical difference between the estimates given by men’s and women’s groups. We found it more difficult to assemble men’s groups as men were usually away during the day. From this experience we infer that Illness Mapping could be extracted from interactions with either gender of respondents, and that women’s groups are likely to be easier to assemble than men.
Our method had to be adjusted to the field reality of low literacy which meant that written consensus and voting was not the best option and so we employed a strategy which involved everyone in a sequential and structured interaction. Our structure emerged from the motivation, explanations, and facilitation techniques that we used to encourage accurate recall and steer discussions towards final consensus.
We examined the potential of our new Illness Mapping method by comparing the results obtained with those derived through a household survey. We compared three parameters which are important for implementation of micro health insurance: (i) prevalence of illness for acute and chronic illnesses, both of which entail cost implications which can be much higher in the case of chronic illnesses , (ii) incidence of hospitalization, as this cost is included in most health insurance programmes, and (iii) incidence of deliveries, especially hospital deliveries. We found very good agreement between the two methods on incidence of deliveries, and good agreement on prevalence of illnesses (in the last one month) and on prevalence of acute and chronic illnesses, as well as on the share of deliveries in hospital.
We obtained a lower estimate of incidence of hospitalization from the Illness Mapping than from the household survey (1.1% (±0.4) from the first source versus 2.6% from the second source). This discrepancy could be the result of two types of memory effects that can lead to erroneous reporting by respondents: errors of omission and of telescoping . While omission means forgetting or omitting to report an episode entirely, telescoping works in the opposite direction, i.e. the respondent remembers and reports an event as having occurred more recently than it actually had. The telescoping effect increases the total number of events reported in a given period. It has also been found that telescoping may be greater in face to face interviews as the presence of an interviewer and the face to face interaction may prod the respondent to give “too much rather than too little information” . It is possible that the telescoping effect may have resulted in an overestimation of hospitalizations in our household survey. In contrast, hospitalizations may have been underestimated in the Illness Mapping method as the group members may have only been aware of the longer duration hospitalizations in their communities and those due to major procedures such as surgeries. They may have omitted the shorter and less severe hospitalizations. This view is supported by prior evidence that longer duration stays and surgeries are more positively associated with recall than other hospitalizations . We do not have a definitive basis to determine which of these estimates is more pronounced, and only actual utilization data could indicate which estimate is the more accurate prediction.
Data obtained either from Illness Mapping or from a household survey would usually be treated by insurers with some reserve, as both methods are less reliable than actual claims data over a long period of time. The Illness Mapping did not, a-priori, show any difference on this count relative to the data obtained from the household survey. In insurance business, it is therefore common practice to include a safety loading in premium calculations, to account for errors in assumptions or inaccuracy of estimates.
The main advantage of the Illness Mapping method is that it is cheaper and faster to operate, and could replace a household survey for estimating morbidity and healthcare utilization, especially where local data is needed but not readily available. While we have tested this method in rural settings in India, we have no reason to think that it could not be equally effective in urban settings (e.g. slums), or in other countries. The estimates about morbidity and healthcare utilization are of course essential not only for insurance purposes, but also for health policy choices more generally. Limitations of this method include the need to establish good contacts with the study communities in order to identify the most suitable community experts. Secondly, high quality group facilitation is essential, by facilitators that must speak the local language and understand the local social settings (and probably be local). Finally, as the estimates obtained by both methods are predictive, one powerful way to evaluate the robustness of the estimates obtained would be to examine both Illness Mapping data and household survey data against actual claims data. Such a follow-up examination is needed to validate the accuracy of the Illness Mapping as a generally applicable alternative to household surveys for the data in question.