Prospective disease and death surveillance can be used to determine disease-specific CFRs. However, because highly effective interventions now exist to prevent cases and deaths, it is ethically unacceptable to follow measles outbreaks without offering vaccination to affected communities and optimal case management, including vitamin A supplementation, to case-patients. Further, because this standard of care may not be routinely available, the CFR from such a prospective study may not reflect the true background rate. As a result, the true risk of death may best be studied retrospectively by conducting a cross-sectional survey to determine the number of measles cases that occurred in a defined area during a pre-determined period, and the number of associated deaths.

In conducting a retrospective study of this type, a critical issue is defining the recall period. Guides for the conduct of retrospective mortality surveys and their strengths and weakness have been described in the literature [29,20]. The recall period must be long enough such that sufficient cases and deaths occur for a precise CFR calculation, but on the other hand, the longer the recall period, the greater the likelihood of recall bias. Because measles is a disease found primarily in epidemics, identifying an adequate number of measles cases to determine measles CFR with some precision generally requires selecting the recall period to fall within the epidemic. Particularly in settings where measles is highly seasonal, this may restrict the period of interest to only several months. The survey must then be conducted toward the end of or very shortly after the epidemic season. Furthermore, determining measles CFRs retrospectively requires identification of all cases of measles within a household, as well as all deaths occurring in measles case-patients within 30 days of rash-onset. (The definition of a measles case is discussed in the forthcoming section). These requirements presuppose detailed recall, but permit focus on a short period (e.g., 3-6 months) prior to the survey.

To accurately determine CFR, both the number of measles cases and associated deaths occurring during the recall period are needed. Literature on all-cause mortality shows that asking in aggregate about deaths occurring in the past 12-24 months results in 30%-40% under-reporting of deaths [30,31]. In addition, differential reporting by both age and gender [32] occurs, with the deaths most frequently omitted variously reported as those among young children [33] and those among children aged 5-15 years [30]. These findings on omission of childhood deaths have led to recommendations in the literature that mothers or other household women serve as respondents [34]. An additional difficulty concerns ascertaining the exact age of household members. Births and deaths are not routinely recorded and age is almost always imprecise. Further, rather than reporting deaths in aggregate, household enumeration should be conducted in order to permit every individual present in the household at the beginning of the recall period, or who joined the household during the recall period, to be accounted for [35,36]. We were unable to find literature on the accuracy of retrospective determination of measles cases. However, one might assume that steps recommended to minimize omission of deaths might also minimize the omission of cases.

A further challenge in conducting retrospective studies of measles is ensuring that the illness studied is indeed measles. Misclassification of rash-fever illness will result in inaccurate CFR estimates. Ensuring that the disease is measles is straightforward if cases have been laboratory confirmed. However, when this is not the case, it is necessary to rely upon history.

Although measles has distinctive clinical features, it can be mistaken for illness due to other causes, most frequently rubella. Thus, measles should be confirmed by collecting serum for testing between days 3 and 28 after rash onset [37]. The implementation of a global measles mortality strategy has led to improved surveillance for measles and the development of a global measles laboratory network [38,39], making laboratory confirmation of measles cases more widely available. However, development of laboratory capacity has been focused in countries that have initiated accelerated measles control, and, even in these settings, laboratories rapidly become overwhelmed during outbreaks [40]. As a result, current WHO recommendations are to limit laboratory confirmation to 5 - 10 specimens per outbreak, although the geographic area implicated (e.g., community, district) is undefined [41]. This recommendation takes into account the fact that the positive predictive value of the clinical case definition for measles increases greatly in the context of an outbreak [42].

The usual WHO laboratory protocol specifies testing for rubella if serum from a suspected measles case tests negative for measles. This has led to the recognition that some outbreaks thought clinically to be due to measles were in fact attributable to rubella, and has also led to confirmation of measles and rubella outbreaks occurring simultaneously in the same communities. In these circumstances, it is difficult to determine by history alone which rash-fever cases were due to measles and which to rubella, thus rendering accurate determination of disease-specific CFRs unfeasible.

In summary, in most measles-endemic settings it is rare that case-patients have laboratory confirmation of measles infection, although it is usually possible to test whether the rash-fever illness circulating is measles. Verifying measles infection in the actual case-patients that we wish to study - i.e., those that have survived 30 days after rash onset and those that have died - remains virtually impossible, as these are patients in whom it is too late to perform confirmatory laboratory testing.

As most measles cases in endemic settings are never subjected to laboratory testing, in retrospective studies history must be used to identify cases. Although a vast literature on verbal autopsies exists, these studies have focused primarily on use and validation of verbal autopsy methodology [43,44]. Verbal autopsy methodology has been extensively used in retrospective mortality studies addressing measles. Two approaches have been used: a clinical algorithm and the local term for measles. The algorithm using age ≥ 120 days, fever ≥ 3 days, and rash to identify measles cases has shown sensitivity ranging from 67% - 98%, and specificity ranging from 85% - 99% [45-47]. WHO recommends the use of either the algorithm or the local term [48]. In studies that compared this algorithm to the use of the local term for measles, the local term was more sensitive and more specific [49].

In order to obtain a precise estimate of CFR, it is important to include an adequate number of measles cases in the sample. This number will depend upon the precision desired, as well as the pre-study estimate of CFR (Table 2). Estimates of probable CFRs for different regions have been published previously and can be used as a guide for determining sample size [10]. The level of precision desired will depend on the ultimate use of the CFR estimate, but should be as precise as possible. Ideally, the sample should be selected from a sampling frame of all case-patients resident in the geographic area of interest with rash onset during the period of interest. However, in the absence of a highly sensitive and specific surveillance system, such a sampling frame does not exist. Because measles tends to be highly clustered in both place and time, a sample drawn from the general population may result in a very large number of individuals being surveyed before an adequate number of measles case-patients is identified.

In addition to estimating CFR, one would like to assess risk factors for increased or decreased CFR among the population studied.

We conducted only retrospective studies for the reasons mentioned previously. In order to best identify all cases and deaths, we used a household census approach, conducting a complete household census of those resident in the household at the beginning of the recall period and using this list to identify those who developed rash illness and died within thirty days of rash-onset. Whenever possible, we limited respondents to mothers or, if they were unavailable, other women within the household. Furthermore, we have limited the recall period to 3-12 months prior to the date of the survey, with the start and end of the period corresponding to a major local festival or event. All studies were conducted at the end of the epidemic season or shortly after.

In one study [21] we had the survey supervisors revisit households with reported measles deaths to confirm the numbers of measles cases and deaths. In addition, one of the primary investigators conducted follow-up household visits to ascertain whether the number of measles cases and deaths detected at the first visit matched those at the second visit, prioritizing households in which the questionnaire responses were unclear or incomplete. These follow-up visits comprised approximately 10% of households.

Despite increasing the likelihood that measles cases and deaths will be accurately detected, this approach has its own difficulties: household enumeration is time-consuming, and respondents may be suspicious of the motives for enumeration. This is particularly challenging when conducting studies in refugee or displaced populations, in which respondents may infer that benefits are tied to their response.

To maximize the likelihood that the rash-fever illness we are studying is truly measles, we have opted to confirm virologically cases in the same community if the outbreak is still ongoing or, if necessary, in a neighboring community to which the community being studied has epidemiologic links. At times, we have also chosen to restrict the communities that we consider for study to those that have had laboratory confirmation of circulating measles [21,22,28]. We have also encountered co-circulating rubella and measles epidemics [28]. In these situations, because of the difficulty in differentiating measles and rubella on clinical grounds, a retrospective study to determine measles CFR may not be feasible.

We initially asked about measles infection using questions extracted from a standardized verbal autopsy questionnaire. However, when we used the algorithm [42-44], we doubted our results because they suggested isolated cases of measles in a poorly vaccinated population, a finding inconsistent with the epidemiology of the disease. We have since chosen to ask about measles infection by using the local term for the disease and then verifying that cases thus identified met the standard clinical case definition for measles, i.e., fever, rash and at least one of the following: cough, coryza and conjunctivitis, in case the term for measles in the local language covers other rash fever illnesses. We recommend this latter approach.

One of the greatest challenges has been identifying a method to select a representative sample of cases. Ultimately, we have chosen to select randomly communities (or other administrative units) from which several measles cases have been reported, and to conduct comprehensive, active, house-to-house case-searches within these communities. With this approach, all case-patients in communities eligible for study have an equal probability of selection, and findings, including the CFR, can be generalized to the population of case-patients in the communities studied. A confidence interval that accounts for clustering can be calculated. This approach may be criticized for failing to include sporadic cases reported to be measles. As indicated earlier, clustering of cases may lead to increased CFR through increased intensity of viral exposure; focusing solely on clustered cases could theoretically lead to an overestimate of true CFR. However, this is unlikely to have a major influence, because the vast majority of cases occur in outbreak settings and so even if sporadic cases were in fact true measles cases, they contribute relatively little to the total burden of disease. Furthermore, sporadic cases reported to be measles and meeting clinical criteria are far less likely to be true measles cases than clinically or laboratory-confirmed measles cases in laboratory-confirmed outbreak settings, because the positive predictive value of a case definition is much lower when the prevalence of measles is low [52].

We have generally felt that the logistical difficulties of conducting a community-wide nutritional survey outweighed the potential benefits of doing so. We have chosen to rely instead upon data from the most recent Multiple Indicator Cluster Survey [53] or Demographic and Health Survey [54] to give an indication of probable community nutritional status, while understanding that these data may be neither geographically nor temporally specific to the outbreak under investigation. These surveys are internationally recognized, use standardized methodologies, permit cross-country comparisons, and frequently provide the only information available on country-level nutritional status. If there is no compelling reason to suspect rapid change in nutritional status, estimates from past surveys may give a rough idea of the importance of malnutrition in contributing to elevated CFRs.

We have tried to assess receipt of vitamin A by showing respondents a vitamin A capsule and asking whether the case-patient received similar medication. Although we have also asked about number of doses received, we have questioned respondents' ability to recall such detail, particularly several months after the fact. In one instance, receipt of vitamin A did not show any impact on measles mortality; leading us to question the reliability of the responses we received [21].

Because of the complexity and additional length of including evaluation of measles case management in our surveys, and because our focus was primarily on determining CFRs rather than evaluating risk factors for death, we chose not to evaluate measles case management in our surveys.

We ensured strong supervision of survey workers with a supervisor responsible for no more than two or three teams. Conducting complete enumeration of households as well as travelling to randomly-selected but distant locations to conduct studies may add to financial costs. However, we consider these approaches to be critical to the rigor of the studies. Studies using convenience sampling or only considering hospitalized cases may be less costly to conduct, but could lead to biased CFR estimates.