Institute of Health Systems

     

Training
Burden Of Disease
AP Burden of Disease Study
AP Health State Valuation Study
Cause of Death
BOD Training Programmes
DFHS
Indoor Air Pollution
Health System Performance Assessment
Health Sector Reform 
Health Informatics
Public Health Databases
Health System-Fact Sheets
Publications
Library
About IHS
Home

 

IHS Mission & Goals:
Groom Skills,
Gather Evidence and
Generate Knowledge for people's health.

To Improve the Efficacy,
Quality & Equity
of Health Systems.

 

Cause of Death    

 

   
  

General design features of a verbal autopsy system:

The verbal autopsy method has been studied and applied in many parts of the world. For example the demographic surveillance system (DSS) in Matlab, Bangladesh (Nahar et al 1985; Zimicki, 1990); assessment of child mortality in Latin America (Puffer and Serrano, 1973); monitoring endemic diseases in West Africa (Bradley and Gilles, 1984; Greenwood et al, 1987) in Kenya (Omondi-Odhiambo et al, 1984), Namibia (Mobley and Ties, 1996); in Phillipines (Kalter et al, 1990) and in India (Bang et al 1992; Awasthi and Pande, 1998). Much of the VA related work, however, remains unpublished. For example the WHO-UNICEF (1994) memorandum on measurement of cause-specific mortality in children cites many unpublished sources.

 

The current knowledge base on feasibility and validity of VA is largely restricted to childhood mortality. The WHO-UNICEF memorandum, cited above, summarizes results of validation studies and has tabulated sensitivity and specificity of VA for detecting major causes of childhood death. In addition, the memorandum contains expert opinion about use of VA for investigation of causes of childhood death. This memorandum was the result of an internal consultation in December 1992 in which experts engaged in research and implementation of VA participated. Bang et al (1992) have used consensus development techniques to synthesize expert opinion on diagnostic criteria for identification of causes of childhood deaths. They have developed questionnaires incorporating local terminologies in their study area (Gadchiroli, Maharashtra) to generate the required information by verbal autopsy to satisfy the coding algorithm.

 

Studies about the validity of VA in identifying causes of adult death have been undertaken recently (Garenne and Fontaine, 1989; LSHTM, 1993). Garene and Fontaine (1989) have reported their experience in Senegal. The London School of Hygiene and Tropical Medicine (LSHTM) workshop (1993) on verbal autopsy tools for adult deaths was conducted on the eve of a study in sub Saharan Africa. Proceedings of this workshop, cited above, documents a consensus of expert opinion about VA for adult deaths. The World Bank working paper by Hayes et al. (1989), is another summary and source of expert opinion. Chandramohan et al (1994) have published discussions at the LSHTM verbal autopsy workshop and have summarized all VA-based studies published upto mid-1993. Certain general design features are the key to wide applicability, efficiency and validity of data generated by a VA based cause of death reporting system. Over the years, some degree of consensus on major design issues has been achieved. I have drawn upon these sources in order, to critically examine the extent to which SCD-Rural meets the criteria of a good VA-based system.

Top 

The structured questionnaires of the SCD-Rural system are systematically examined for each of the conditions included in the non-medical list, in the light of available research results on verbal autopsy. SCD-Rural seems to satisfy most of these criteria except that of reporting multiple causes of death. However, assigning multiple causes of death creates problems for aggregation and reporting of deaths by cause. Manton and Stallard (1984) analyzed multiple cause of death patterns in the USA. Although their preferred suggestion is to use patterns of failure as the basis of analysis, it may not be a feasible alternative considering the small sample sizes inherent in verbal autopsy-based statistics. To the extent that certain deaths are assigned to a combination of causes, there will be reduction in number of deaths reported under the respective component causes (LSHTM, 1993). A compromise may be to restrict the number of multiple causes of death to a manageable number and develop algorithms to distribute these to their component causes. Manton and Stallard’s (1984) study suggests that recording upto three multiple causes would include more than two third of deaths. Choosing the top three most probable causes contributing to death may help improve the accuracy of estimates and keep it manageable.

 

The trade off between the open-ended interview and the structured questionnaire needs further elaboration. Although an open-ended interview format allows for the pursuit of unusual diagnostic clues not covered by a structured questionnaire, it requires more skilled interviewers. For example, comparatively lower assignments to unknown category have been achieved with physicians acting as interviewers (Greenwood et al 1987). Open-ended interviews and coding of cause based on the judgment of the interviewer reduces the inter-regional and inter-temporal comparability of cause of death statistics.

 

Since SCD-Rural satisfies most of the general design criteria for VA, does it follow that the statistics generated by it would automatically be valid? Not necessarily. Validity of classification of deaths to particular causes will depend on characteristics of the cause of death per se, as also the content of the questionnaire and algorithm used for specific disease entities the latter two of which are discussed in the next sub section.

  Top 

Validity of SCD-Rural disease-specific algorithms:       

As the WHO-UNICEF (1994) consultation noted, VA is suitable only for causes that have clear and unambiguous set of symptoms at the time of death. The symptoms and signs chosen to code deaths due to a particular disease should result in most of the deaths truly due to the concerned cause to be coded as such (sensitivity) and exclude other causes that may have related symptoms. In addition, the choice of symptoms and signs must be parsimonious to reduce interviewer and interviewee fatigue. At the very least, questions and coding algorithms should have face and content validity. In other words, they should be based on expert judgment about their usefulness in identifying and excluding specific causes. In addition, validity with respect to a criterion will be desirable. The validity of an instrument is assessed by comparing its result with some reference standard. Thus the choice of a reference standard is the key to empirical validation of VA algorithms.

 

The gold standard reference for assignment of cause of death has been the autopsy. This is not a practicable solution to validate VA, since the later alternative to medical certification of cause of death is considered only in areas with scarce medical facilities. Two other references have been proposed (LSHTM, 1993) and used, namely (a) hospital diagnosis and (b) clinical diagnosis. To validate VA with respect to hospital diagnoses, deaths in a community are coded using the VA instrument under testing. If the deceased happened to have been hospitalized, the medical records from hospital are retrieved. The reference cause of death is assigned on the basis of the person's medical record in hospital. Alternatively patients discharged from a hospital may be followed up after a lapse of time and deaths if any may be coded by VA. The hospital-based reference diagnosis and the VA-based code are then compared. Selection bias is a major shortcoming of hospitals based reference. The LSHTM workshop discussed possible ways of reducing selection bias. An example of hospital diagnosis-based reference is the study in Kenya by Snow et al (1992). In this study hospital diagnosis was used as a reference to check validity of cause of death coded by physicians from verbal autopsy data. On the other hand, clinical diagnosis in the community has less of a selection bias. This would require a lot of medical manpower, which may not be available in an area for which VA is considered. It may, however, be possible to temporarily mobilize physicians for purposes of a validation study, since methodological lessons learnt from it would be useful for wider application. Kalter et al (1990) used physician diagnosis as the reference to estimate validity of different verbal autopsy based algorithms. Zimicki (1990) compared interviews by lay reporters with in-depth interview by physicians.

 

Sometimes empirical validity of VA tools are assessed indirectly by checking consistency of VA-based statistics with known epidemiological patterns. One approach has used known efficacy of vaccination to reduce mortality due to concerned disease. Validity of a VA tool measuring mortality due to that disease may be indirectly inferred from the time trend of estimates generated by it and vaccination coverage. Stephens (1990) studied measles-related morbidity and mortality data collected by nonmedical field interviewers in a rural area in Senegal. Data on measles incidence and cause-specific mortality was aggregated by hamlets. Stephens examined if the movement of measles epidemic from hamlet to hamlet implied by the verbal autopsy data was consistent with known epidemiologic pattern of measles and vaccination coverage in respective hamlets.

 

SCD-Rural algorithm organizes all causes, at the highest level, into ten modules based on obvious age-sex-major symptom complex. The solution to the first round of questions about applicability of these modules leads the interviewer into the detailed questions under that module. It will be fairly obvious to determine if the death was due to, say, accidents and injuries (SCD module-1), maternity (module-2) or if it was of an infant less than one year old. There is a problem about the last module on senility. There are no further expansions of causes under senility. Criteria for inclusion under senility is that the person was extremely old and apparently not sick? The person should be above 60 years and none of the specific causes in SCD list be traced. The age criteria of more than 60 years would tend to put more deaths under this category.

 

SCD (Rural) cause groups, availability of expert opinion or validity information on each cause and concordance of SCD questions with expert opinion.

Category1 Not Available Availability and Concordance2
Fevers (3) Influenza, Typhoid Malaria
Digestive disorders (6) Food poisoning, Peptic ulcer, Acute abdomen Gastroenteritis (diarrhoea), Cholera, Dysentery
Coughs (5)   Tuberculosis of lungs, Bronchitis, Asthma, Pneumonia, Whooping cough
CNS disorders (3)   Stroke, Meningitis, Convulsions
Diseases of the Circulatory System (3) Anaemia Congestive heart failure, Ischaemic heart disease
Other clear symptoms (13) Cirrhosis and chronic liver diseases, Chicken pox, Leprosy, Poliomyelitis, Mental disease, Diabetes, Hyperplasia of prostate, Uraemia, Obstructed hernia Jaundice, Measles, Tetanus, Cancer
Infant deaths (6)   Prematurity, Congenital malformation, Birth injury, Respiratory infection of the new born, Cord infection (Neonatal tetanus), Diarrhoea of the new born

Figures in parentheses are the number of conditions within the group. For 12 causes under Accident and Injuries, expert opinion is not available for an individual or specific cause. However, there is a general agreement that these causes are obvious to lay reporters and hence verbal autopsy is considered to accurately assign deaths due to these causes. For similar reasons, the six causes under maternal deaths is not shown. Senility and other residual codes are not shown.

Expert opinion and SCD questions for the underlined causes of death listed under this column do not agree.

The SCD structured questions and checklist were compared with currently available expert opinion or validity information for respective causes of death in the SCD non medical list. Altogether there are 57 specific causes in the SCD non- medical list, excluding the residual categories. Accidents and injuries account for 12 of these. Consensus about validity of VA to code deaths due to accidents and injuries is quite strong, since most of these are easily recognized by lay persons. Discussions of VA on accidents and injuries are not available in the literature. So is the case with deaths due to maternal causes, under which SCD non medical list contains 7 causes. Excluding these 19 causes under accidents, injuries and maternal deaths, there are 38 specific codes in the rest of the SCD non- medical list. At least some expert opinion or validity information is available for 24 out of these 38 causes.

  Top 

Previous             Next  

  

| AP State Profile | Public Health Database | IHS Resource |  Careers | Contact us |