In order to function well as part of routine systems, the forms had to be easy to use by interviewers with varying degrees of clinical skills and knowledge. They were also used to record relevant contextual information (such as use of health facilities in the period before death and data on risk factors). Additionally, the forms made use of any documentary evidence available from the household of the deceased that might aid in determining the probable cause of death. Lastly, the forms had to provide physicians with enough data to produce internationally comparable mortality statistics based on ICD coding guidelines, and be amenable to developing data-derived algorithms to determine the probable cause of death [41].

All short core VA forms referred to in this article are available online. They include the forms for perinatal events and neonatal deaths (Figure S1), deaths in post-neonatal children under age 5 (Figure S2), and deaths among persons aged 5 y and above (Figure S3). Each form follows the same basic structure: identifying information about the deceased (including age, sex, and place of death), cause of death according to respondent, short narrative history, symptom duration checklist, health services used in the period before death, and any medical evidence available at the household, including whether a health worker informed the respondent of the cause of death.

A section on the condition of the mother during and after pregnancy and birth is included on the neonatal form. Deaths to women of reproductive age, and maternal deaths in particular, are addressed in a subsection of the form for deaths over age 5. The questions contained in the symptom duration checklist are generally arranged by anatomical system. They are intended to provide strong support for a positive diagnosis of probable cause of death, and the confident exclusion of differential diagnoses.

Table 2 compares the VA form used at some sites that are members of the INDEPTH Network [42], the VA form used by the Indian Sample Registration System (Indian SRS) [15–18], and a VA form from the WHO [12], and notes key areas of difference in content. VA interview forms tailored to specific age groups generally use the same standard cut-off for neonatal mortality (i.e., death before 28 d), although there is some variation in the age range for the application of forms for post-neonatal child deaths [17,42–44]. For reasons of cost and ease of implementation, the layout and length of the proposed core forms were limited to no more than two A4-sized pages. In this they are similar to forms used in the Indian SRS, and much briefer than most other VA forms presented in the literature [12,13,41].

Aside from the differences noted in Table 2, the use of long “open history” sections in other forms is another major difference between the proposed core forms and other published VA tools. The Indian SRS form for adult deaths, for example, relies almost exclusively on narrative histories of the events preceding death to provide evidence about the cause of death. If administered as a clinical history, these sections can provide relevant information to physicians who assign probable causes of death. The short core forms in Figures S1–S3 allow for brief narrative histories, but emphasize a “symptom duration checklist” approach for use in cause-of-death attribution.

Experience in implementing VA procedures suggests that lengthy clinical history sections cannot be standardized and vary substantially depending on the clinical skills and medical training of the interviewer. In addition, interviewers may introduce bias into data collection by recording histories that neatly fit into known or familiar disease descriptions or are based on the interviewers' initial impressions of the likely cause of death. Therefore, a core symptom duration checklist may be more systematic and to produce a more complete inventory of the signs and symptoms before death than would a heavier reliance on open histories.

A VA interview is conducted similarly to any confidential health-related interview, with the added consideration that the subject matter concerns a topic that could hardly be more distressing—the recent death of a family or household member. This, in part, speaks to the need to enroll respected community members in areas where VA will be implemented to help build local awareness and acceptance of what is, generally speaking, a new and unfamiliar mechanism of collecting health information.

There is a range of opinion about whether medical training should be a preferred qualification for VA interviewers or whether educated but non-medically trained persons are more suitable. Local experience will determine the optimal solution. Training should include discussion of symptoms and their description in local languages. In addition, a clear understanding of how live births and stillbirths can be accurately differentiated using appropriate terminology is important. Ideally, the interview should happen as soon as possible after a death with due consideration to culturally appropriate mourning periods. All questions on the VA form (aside from the appropriate skips) must be asked of the respondent regardless of the opinion of the interviewer as to their relevance. Quality assurance should be performed routinely. If feasible, re-interview of a 10%–15% sample of VAs would offer a strict standard. However, given the sensitive nature of VA, it may be sufficient to verify for a similar proportion of deaths that (a) the death indeed occurred and (b) the VA interview in fact took place at the household of the deceased, with an appropriate respondent. Interviewer retraining and supportive supervision are probably the most important components of quality assurance for VA.

As with any survey instrument intended for cross-cultural and cross-linguistic application, care must be taken in translation into local languages and field testing so that all questions are understood by respondents in the way they are intended [22]. Part of the cultural validation of VA should include observations of VA by a medical anthropologist or sociologist, and interviews with community members to ensure accurate understanding of terms used in the VA form. VA interviewers should be informed about any areas of potential confusion due to local or colloquial expressions. This will help ensure that in addition to building community rapport for administering VA, culturally appropriate and sensitive terms, idioms, and expressions are used in interviews without sacrificing precision and cross-comparability of results. A balance must be struck between clinically precise terminology, which can be confusing or even offensive in the context of a VA interview, and colloquial expressions or local terms that might impede accurate cause of death attribution and ICD coding.

It is also important to consider linguistic and cultural issues in implementing proposed core VA forms in very different settings. The experience of transferring these procedures from Tanzania to China has been instructive in this regard. It is felt that both the brevity of the interviews and the efforts expended in establishing rapport contributed to attaining response rates in both settings of over 90%. Minimal, though important, modifications were required to translate specific questions and variables from the original Tanzanian forms into the Chinese context. For instance, the question “Was [the deceased] breathless on lying flat?” employed in Tanzania was not clearly interpreted in China, and a question on “breathlessness interfering with sleep” was substituted. Chinese interviewers readily adopted the protocols, and physician reviewers in both sites were able to certify causes of death using an international death certificate. Finally, statistics could be compiled from both countries according to the proposed tabulation list, yielding internationally comparable data.

Because VA may serve as the best or even sole evidence on cause of death in many settings, establishing international comparability is important [4–6,26,45]. The lack of standard interview forms, cause-of-death categories, and coding practices has hindered attempts to synthesize results from various applications of VA to assessing precise causes of child mortality [4,6,45]. Therefore, it is recommended that the causes of death as determined through physician review of VA be recorded using a four-line death certificate (i.e., underlying, immediate, associated, and contributory causes). Subsequently, a physician or medical recorder should select and code the underlying cause to the core 3 character code using standard ICD-10 rules.

In certain cases it may be possible to code to the fourth digit. On the other hand, ICD-10 rules may frequently preclude the use of certain three-digit codes in the VA context. For example, ICD codes starting with B50, B51, and B52 refer to malaria. Use of these codes requires both confirmation of parasite infection and identification of the malaria species. It is unlikely that evidence of such confirmation would be available at the household level. In such instances there are usually three-digit codes available for use (in this case either “B53 other parasitologically confirmed malaria” or “B54 unspecified malaria—clinically diagnosed malaria without parasitological confirmation”) that would not affect the outcome of tabulating and reporting VA data, or the main public-health interpretations and policy implications of the tabulated data.

Physicians who review the completed VA forms usually require training in cause-of-death certification using ICD rules and international death certificates. An explanation of the structure and content of the ICD classification, and of the rules for selection and coding of the underlying cause of death, is also necessary to ensure uniformity of data across different coders. Details of coding guidelines and criteria, manuals, and options for organizing VA coding can be obtained from the authors.

The issue of reporting single versus multiple causes of death in VAs, particularly for children, has been addressed extensively. Most sources recommend or employ multiple cause-of-death attribution in children without providing a single underlying cause [43,44,46–48], and at least one source does so for adults [13]. We recommend the use of standard death certificates for all ages in accordance with ICD convention ([32], p. 31). This enables recording, coding, and analysis of multiple causes of death while retaining comparability of mortality data based on the tabulation of a single underlying cause, as prescribed by ICD. Although ICD does recommend a specially designed death certificate for perinatal deaths ([32], p. 90), few countries have implemented it. For the present, therefore, perinatal deaths (and stillbirths, if desired) may be recorded, together with neonatal deaths, on a conventional death certificate.

The ICD recommends two “condensed” tabulation lists for mortality reporting [33]. These lists contain many causes that can be accurately identified only with specific diagnostic or clinical information. In the case of VA, the smaller list of causes presented in Table 1 is more appropriate. The ICD specifically sanctions the development of such tailored tabulation lists [32].

There were 25 causes in China and 26 causes in Tanzania (for at least one age group) for which there were at least five deaths in both the VA and the medical records. Table 3 shows summary results for causes of death for which at least five deaths were validated, sensitivity was greater than 50%, and the relative difference in the cause-specific mortality fraction (CSMF) in the VA (CSMF_VA) and medical record (CMFS_MR) was equal to or less than 20%. These criteria are based on threshold values for sensitivity and CSMF suggested for assessing accuracy of adult VA [49]. The comparison of CSMFs was based on the relative difference in the proportion of deaths due to cause X in the medical records dataset (CSMF_MR) from the proportion of deaths due to the same cause in the verbal autopsy dataset (CSMF_VA). We calculated sensitivities and specificities for all causes using conventional two-by-two table analysis, although results are displayed only for those causes reaching the threshold sensitivity. For the over-five age group, data were available for 1,912 deaths in 42 cause-of-death categories from Tanzania, and 2,029 deaths in 37 categories from rural areas in China. Of these, 140 deaths from Tanzania and 170 deaths from China were coded to either “all other specified diseases” or “undetermined”; these are excluded from the comparison.

Of the 20 causes of death listed in Table 3, three (VA-02 tuberculosis, VA-27 cerebrovascular diseases, and VA-50 transport accidents) met the threshold criteria in both countries. Six causes reached the threshold in Tanzania only, and 11 causes met the threshold levels in China only. At the upper bound of the 95% confidence level for sensitivity, seven additional causes reach the threshold in Tanzania (VA-09 remainder of infectious and parasitic diseases, VA-15 malignant neoplasm of liver, VA-19 other neoplasms, VA-21 diabetes mellitus, VA-25 hypertensive diseases, VA-38 other maternal causes, and VA-57 all other external causes).

Significantly, for six causes (VA-09 remainder of infectious and parasitic diseases, VA-17 malignant neoplasm of breast, VA-21 diabetes mellitus, VA-25 hypertensive diseases, VA-29 asthma/chronic obstructive pulmonary disease, and VA-51 falls) the relative difference in sensitivity of VA was less than 25%. It should be born in mind that because of the small samples for certain causes of death it was not possible in the studies to validate all the causes contained in Table 1.