SEARCH

SEARCH BY CITATION

Keywords:

  • verbal autopsy;
  • retrospective interview;
  • mortality;
  • morbidity;
  • pneumonia;
  • diarrhoea

Abstract

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

Summary This paper reports the validation of a ‘best-judgement’ standardised questionnaire using guidelines and algorithms developed by an expert working group conducted in Nicaragua between 1995 and 1997. Prospective hospital data, including standardised medical recording of selected signs and symptoms, laboratory and radiographic test results and physician diagnoses were collected for children < 5 years admitted with any serious life-threatening condition in 3 study hospitals. The mothers or caregivers of the children were later traced and interviewed using the ‘best-judgement’ questionnaire. Interviews were completed 1–22 months after admission to hospital for 1115 children (400 who died during the stay in hospital and 715 who were discharged alive). The cause of death or admission to hospital was determined by an expert algorithm applied to hospital data. A similar procedure was used to derive the cause using the answers to questions from interviews. Hospital causes were compared with interview causes and sensitivity and specificity calculated, together with the estimated cause-specific fraction for diarrhoea and pneumonia. Multiple diagnoses were allowed; 378 children in the sample (104 deaths, 274 survivors) had a reference diagnosis of diarrhoeal illness, and 506 (168 deaths, 338 survivors) a reference diagnosis of pneumonia. When results for deaths and survivors in all age groups were combined, the expert algorithms had sensitivity between 86% and 88% and specificity between 81% and 83% for any diarrhoeal illness; and sensitivity between 74% and 87% and specificity between 37% and 72% for pneumonia. Algorithms tested in previous validation studies were also applied to data obtained in this study, and the results are compared. Despite less than perfect sensitivity and specificity, reasonably accurate estimates of the cause-specific mortality and morbidity fractions for diarrhoea were obtained, although the accuracy of estimates in other settings using the same instrument will depend on the true cause-specific fraction in those settings. The algorithms tested for pneumonia did not produce accurate estimates of the cause-specific fraction, and are not recommended for use in community settings.


Introduction

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

Estimates of cause-specific mortality and morbidity obtained using retrospective interviews (often called verbal autopsies in the case of mortality) have many uses including measuring the effectiveness of disease-specific interventions ( Barnish et al. 1993 ; Harries 1993), evaluating the impact of heath care programmes (Dr M. Lucero et al. unpublished observation), and prioritizing health problems for planning purposes ( Fauveau et al. 1990 ; Boerma et al. 1994 ). However, caution is needed in the interpretation of results because several factors can influence the accuracy of estimates, including misclassification, the size of the cause-specific fraction for each disease, and the mix of diseases in the particular setting ( Anker 1997).

If a retrospective interview is to result in an accurate diagnosis, the illness of interest should have a distinct set of signs and symptoms recognizable by the informant ( Chandramohan et al. 1994 ). In general, the more distinct and recognizable the syndrome, the better the performance of the interview in assigning causes. Measles, for example, has generally been more accurately diagnosed by postmortem interview than malaria principally for this reason. Verbal autopsy validation studies for deaths due to measles have found sensitivities between 67% ( Mobley et al. 1996 ) and 98% ( Kalter et al. 1990 ), and specificities of 85% ( Mobley et al. 1996 ) to 99% ( Kalter et al. 1990 ). Results for malaria, on the other hand, have been less promising, with sensitivity and specificity in one validation study at 46% and 89% (Snow et al. 1992), and in another at 55% and 80% ( Todd et al. 1994 ). Previous validation studies have tested algorithms with moderately high sensitivity and specificity for deaths due to diarrhoea ( Table 1) and pneumonia ( Table 2) in the Philippines, Namibia and Bangladesh.

Table 1.  Comparison of interview algorithms for diarrhoea tested in published validation studies with algorithms tested in Nicaragua (deaths < 5-years-old) Thumbnail image of
Table 2.  Comparison of interview algorithms for pneumonia tested in published validation studies with algorithms tested in Nicaragua (deaths < 5-years-old) Thumbnail image of

Few validation studies have included survivors of severe illness. Among these are a recent study in Bangladesh ( Kalter et al. 1999 ) which reported up to 74% sensitivity and 87% specificity for pneumonia in neonates ( Table 2) while another study in the Philippines ( Kalter et al. 1991 ) achieved a sensitivity of 95% and a specificity of 80% for diarrhoea ( Table 1).

The most common method used for assigning causes of death based on interview data has been using a panel of 2 or 3 physicians to review the interview forms independently. Typically, where 2 or more physicians have reached the same diagnosis independently, that cause of death has been assigned ( Alonso et al. 1991 ; Snow et al. 1992 ; Ghana VAST Study Team 1993). Assignment of cause of death by physician review relies to some extent on subjective judgement and results may vary considerably between doctors in the same setting, between settings, and even between the same doctors in the same setting ( Todd et al. 1994 ).

Pre-defined expert algorithms for each disease of interest have sometimes been used to assign a cause of death, either alone ( Becker et al. 1993 ), or in conjunction with physician review ( Bang et al. 1990 ; Kalter et al. 1990 ). An advantage of using algorithms alone is that the derivation of causes of death in different settings can be standardized. There is some doubt, however, about whether algorithms are capable of making diagnoses with validity as high as that obtained by physician review. Chandramohan et al. (1998) reported that physician review produced more accurate results than the application of expert algorithms in a validation study of an adult verbal autopsy. A comparison of the performance of expert algorithms with algorithms derived from VA interview data itself ( Quigley et al. 1996 ) showed that for most causes of death, the data-derived and expert algorithms yielded similar levels of diagnostic accuracy, and in the case of malaria, the data-derived algorithm performed significantly better with an estimated sensitivity of 78% compared to 47% for the expert algorithm. In settings where malaria is present, the validity of retrospective interviews for detecting pneumonia and some other illnesses has been poor, because many signs and symptoms may be indicative of both malaria and other diseases ( Todd et al. 1994 ) and therefore have low specificity. One of the reasons for selecting Nicaragua for this validation study was to remove the effect of malaria; the study was conducted in an area of Nicaragua where malaria is very rare in young children.

Estimates of cause-specific mortality and morbidity obtained by interview are subject to various sources of bias, as well as inconsistencies due to differences in disease mix and methodologies between settings. Selected results of other validation studies for diarrhoea and pneumonia, which have used a variety of questionnaires, methods for assigning cause of death or illness and procedures for defining reference standard diagnoses for comparison, are shown in Tables 1 and 2. There are significant differences in the sensitivities and specificities of interviewing instruments for reaching these diagnoses. Although these may be due partly to cultural factors, they may also be due to differences in the mix of causes, and the methods of deriving reference diagnoses. The lack of standardized questionnaires and methods of diagnosis limits the comparability of results across settings.

As part of a multi-institution study between the World Health Organization, the London School of Hygiene and Tropical Medicine, the Johns Hopkins School of Public Health and the Wellcome-funded Coastal Unit of the Kenya Medical Research Institute we developed a ‘best-judgement’ questionnaire and expert algorithms for deriving diagnoses from signs, symptoms and test results recorded during the stay in hospital (reference diagnoses) and from caregivers' answers to questions during a subsequent interview (interview diagnoses). Validation studies using a similar protocol and questionnaire have been conducted in Nicaragua (this study), Bangladesh ( Kalter et al. 1999 ) and Uganda ( Kalter et al. 1997 ; Kolstad et al. 1997 ), although the results published to date have been only for neonatal causes and anaemia.

Materials and methods

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

Hospital data

In 3 hospitals in Nicaragua (Velez Paiz Hospital, Managua; Manuel de Jesus Rivera Hospital, Managua; and Hilario Sanchez Hospital, Masaya), admissions to paediatric and maternity wards were prospectively screened to identify neonates and children who were severely ill or who died in hospital. The admission of all children < 5 years during the recruitment period (September 1995 to March 1997) was monitored by a study physician in each hospital. The peak season for diarrhoea and pneumonia admissions in Nicaragua is during the rainy season, from about June to about September. The recruitment period included one complete rainy season. All children < 5 admitted to or born in hospital who met any of the inclusion criteria ( Table 3) were eligible for enrolment.

Table 3.  Inclusion criteria Thumbnail image of

The signs and symptoms noted at the time of admission were recorded on an admission summary form by the admitting physician or the study physician. Signs and symptoms noted at any time during the stay in hospital, and the results of all laboratory and X-ray investigations were abstracted by the study physician using a discharge summary form, after the death or discharge of children enrolled in the study. All forms were checked by a physician employed by the project in each hospital and double entered into a customised database.

Nine hundred and fifty-two chest X-rays were taken for 545 children included in the study. The findings for each X-ray film, recorded on the hospital notes by the attending physician, were transcribed by the study physician. Two radiologists (identified as A and B) independently read the transcribed notes and assigned a diagnosis based on these notes alone. The diagnoses for each child by radiologist A were compared to those of radiologist B and the kappa statistic was calculated ( Landis & Koch 1977). The agreement between the two radiologists was substantial, although less than perfect (89% agreement, κ= 0.73). The two radiologists then re-read the X-ray films without re-reading the file notes for a subset of the data (= 78), and made a diagnosis. By this method the level of agreement between the two radiologists for a diagnosis of pneumonia was lower (73% agreement, κ= 0.35). For the purpose of determining the cause of death or admission to hospital by the algorithm which included radiologically confirmed pneumonia ( Table 4; algorithm E), a positive X-ray diagnosis of pneumonia was assigned where both radiologists had given a diagnosis of pneumonia from re-coding the notes, or where both had given a positive diagnosis from re-reading the films. The remaining X-rays were coded as negative for pneumonia.

Table 4.  Algorithms used for deriving reference and interview diagnoses Thumbnail image of

Interview data

The questionnaire was developed by a steering group following an informal meeting of WHO and UNICEF ( WHO/UNICEF 1995) in 1992 and agreed in consultation with other participants in the Child Verbal Autopsy Working Group. The questionnaire contains an open history section, a checklist of major signs and symptoms and modules of more detailed questions when one or more filter question results in a positive response. This ‘best-judgement’ diagnostic questionnaire was revised, translated into Spanish, back- translated into English for verification, and finally pilot-tested in the study area. Pilot-testing included administering the questionnaire to a test group of respondents in order to identify questions that were unclear or poorly understood. The questionnaire was re-drafted and re-tested until we were satisfied that all questions were clearly understood. Interviewers were female university graduates without training in medicine, nursing or health sciences, and the interviews were conducted between March and August 1997. Interviewers were trained by local physicians to explain the meaning of any medical terms in the questionnaire which may not have been well understood by respondents. This included demonstrating various breathing noises, and explaining the meaning of terms such as nasal flaring and chest indrawing. Regular meetings were held with all interviewers and supervisors during the interviewing period to review any difficulties encountered.

Using addresses recorded by project staff at the time of admission to hospital, interviewers attempted to trace the mother of each child enrolled in the study, or another informant who was present during the illness which led to the relevant admission to hospital. If an appropriate informant was not present, up to three further visits were made in an attempt to complete the interview. We interviewed mothers or other informants on 445 deaths including 45 stillbirths (76% of those sought) and on 715 survivors (60% of those sought) after attempting to trace 1777 addresses (586 deaths and 1191 survivors). The higher proportion traced for deaths reflects the additional effort made with the resources available to seek informants for children who died. Retrospective surveys ('verbal autopsies') are most commonly used to measure mortality, and although the focus of this study was mortality, we also included children who had survived a life-threatening illness; partly to increase the sample size and power, and partly to validate the questionnaire for the measurement of severe illness as well as for death. Verbal or written consent was obtained from informants after explaining the purpose of the study and before commencing the interview. 25 traced respondents refused to give an interview (16 deaths, 9 survivors).

Fifteen interviews were excluded from analysis. In 2 cases, the informant was unwilling or unable to give reliable information about the child's illness. In 9 cases, more than one interview was completed for the same child because of multiple admissions to hospital. Duplicate interviews for the same child were excluded, retaining the interview pertaining to the most recent admission. In two cases, the informant was unable to answer whether the child was born alive, and in two cases, the child was reported as having been dead on arrival at the hospital.

Reference and interview diagnoses

An expert panel of physicians, each with many years experience working with childhood illness in developing countries determined standardized algorithms, both for the reference (hospital) diagnoses and for the interview diagnoses. The algorithms used for deriving both diagnoses are shown in Table 4. One gold standard algorithm for each illness of interest was applied to the hospital data. Up to 3 algorithms for each illness of interest were applied to interview data, with the aim of identifying the algorithm which produced the best results. In addition to diarrhoea and pneumonia, algorithms were tested for injury, bacteraemia/ septicaemia, meningitis/encephalitis, low birth weight/ prematurity, severe malnutrition, congenital abnormality, birth trauma, birth asphyxia, and local bacterial infection. Analysis of the results for these other diagnoses is currently in progress. The study therefore included the principal causes of death and severe illness in this setting. Multiple diagnoses (reference and interview) for individual children were allowed resulting in between 0 and 5 reference diagnoses for each child (median 1) and between 0 and 6 interview diagnoses (median 1).

In addition to sensitivity and specificity, we calculated the true cause-specific fraction for diarrhoea and pneumonia using the hospital algorithm, and the estimated cause-specific fraction from each of the interview algorithms. The true cause-specific fraction (CSF) is the number of hospital admissions with an expert algorithm diagnosis of a particular cause, divided by the total number of hospital admissions included in the sample with any diagnosis. The estimated cause-specific fraction is the number of interviews with an expert algorithm diagnosis of a particular cause, divided by the total number of interviews with any diagnosis. Where missing values prevented the application of expert algorithms, both the case with the missing value(s) and the corresponding hospital or interview case were excluded from the calculation of the CSF and other measures of validity for that cause. Neonates (< 28 days), postneonates (28 days to < 5 years), deaths and survivors were analysed separately.

Results

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

At the time of interview, study children had been admitted to hospital between 1 and 30 times (median 1). The interview respondent was the mother in 87% of cases, a grandmother in 5%, and an aunt, the father or another person in the remaining 8%. The mean age of the main informant was 28 years (range 15–66), with a mean of 5 years formal schooling (range 0–15). The mean age of mothers was 26 years (range 15–48). Interviews lasted 7–85 min (mean 22 min) after tracing the main informant. The mean time between admission to hospital and interview was approximately 13 months (range 1–22 months).

Forty-three children discharged from hospital alive had died by the time of the interview. Of these, 11 died within 1 day of discharge from hospital and were classified as hospital deaths because the cause of death was almost certainly that for which the child was being treated in hospital. The remaining 32 were classified as survivors (of that admission to hospital). The median time between discharge and death for these 32 children was 35 days (range 2–589). Where the child was discharged from hospital alive and died subsequently, the informant was asked to answer questions relating to the illness leading to that particular admission to hospital, rather than to the illness leading to death, although in some of these cases the illness is likely to have been the same.

Six (4.3%) neonatal deaths and 98 (43.6%) postneonatal deaths had a reference diagnosis of diarrhoea, while the proportions for survivors were 22 (13.6%) for neonates and 252 (45.8%) for postneonates. The interview algorithms did not distinguish clearly between diarrhoea and dysentery, nor between acute and persistent illness, the disaggregated results being characterized by high specificity and low sensitivity. Combining all diarrhoeal illnesses (deaths and survivors, neonates and postneonates) gave a sensitivity of 86% to 88%, and a specificity of 81% to 83% depending on the algorithm used, rendering acceptably accurate estimates of the CSF. Algorithms D.1 and D.2 gave reasonably accurate estimates (0.41) of the true CSF (0.34) for all groups combined, while algorithm D.3 gave an estimate of 0.42. Sensitivity and specificity were generally higher for neonates than for postneonates.

Thirty-seven (29.1%) of the neonatal deaths and 131 (71.2%) of the postneonatal deaths had a reference diagnosis of pneumonia, while the proportions for survivors were 54 (51.4%) for neonates and 284 (73.8%) for postneonates. Results for the subgroup analyses are shown in Tables 5 and 6. Sensitivity was slightly higher for survivors than for deaths, and the specificity generally slightly lower. There were no consistent differences between the performance of the algorithms for neonates and postneonates. Combining all subgroups for the first interview algorithm ( Table 4, E.1) resulted in moderate sensitivity (74%) and specificity (72%) and a CSF estimate of 0.57 compared to the true CSF of 0.63. The second algorithm for pneumonia ( Table 4, E.2) had very poor specificity (37%) resulting in substantial inaccuracy for the estimate of the CSF (0.78). The poor specificity for algorithm E.2 was consistent across subgroups (deaths and survivors, neonates and postneonates).

Table 5.  Validation results – diarrhoeal illnesses and pneumonia (deaths in hospital) Thumbnail image of
Table 6.  Validation results – diarrhoeal illnesses and pneumonia (survivors of admission to hospital with life-threatening illness) Thumbnail image of

We compared the sensitivity and specificity of algorithms used in previous validation studies with their performance when applied to the data collected in Nicaragua and these results are shown in Tables 1 and 2. For diarrhoea, algorithms from previous studies generally gave lower sensitivity and higher specificity than the expert interview algorithms in our study. For pneumonia, several algorithms from previous studies had much higher specificity than our interview algorithm (E.2), although only cough, dyspnoea and fever had both moderately high sensitivity and specificity (sensitivity 62%, specificity 72%).

Discussion

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

Compared to diagnosis by physician review, our method of deriving diagnoses from expert algorithms has the advantage that definitions can be standardized, thereby improving comparability between settings which use similar validation instruments and algorithms in the future. The results obtained in this study, however, may not be comparable to previous validation studies which have used different methods.

When reporting and aggregating cause of death statistics, usually only a single, underlying cause of death is given, even though most child deaths in developing countries are probably due to multiple causes, and identifying only a single cause of death can lead to misleading results ( Chandramohan et al. 1994 ). Therefore more attention has recently been given to analysing deaths by multiple cause, often described as death with, rather that death of ( Garcia Benavides et al. 1992 ; Marsh et al. 1995 ). We allowed multiple causes of death in this study, with the effect that the CSF for each illness is generally much higher than where a single cause is assigned. The true CSF for each illness of interest is an important determinant of the accuracy of estimates using retrospective interviews ( Anker 1997), and the performance of algorithms will vary substantially when applied to populations with different fractions in the absence of near-perfect sensitivity and specificity. In a study allowing only single causes, the total of fractions due to each cause will add to 100%, while in this multiple cause study, the total is much greater, since each child, on average, had more than one diagnosis. This resulted in very high true CSFs for diarrhoea and especially for pneumonia in this study. This important difference, as well as other differences in the methods used, limits the comparability of our results with previous validation studies, and this should be taken into account when interpreting the comparisons presented in Tables 1 and 2.

In common with all hospital-based validation studies, the disease mix is that of children admitted to hospital and may not therefore be representative of the disease mix in the community. The disease mix represents illnesses in children with symptoms which mothers recognize as serious enough to require urgent treatment in this setting.

The informant's attendance at the hospital with a sick child was likely to influence recall of the illness, and to allow access to medical information of which he or she would not have been aware had the child died without attending hospital. In many cases, and especially when a child died, the mother was informed of the diagnosis by the physician in attendance, and was also normally given a copy of the death certificate which confirmed the final diagnosis. Furthermore, one of our algorithms for diarrhoea (D.2) includes ‘local term for diarrhoea’, which in Nicaragua was ‘diarrea’; and similarly, one of the algorithms for pneumonia (E.1) used ‘local term for pneumonia’ which in Nicaragua was ‘neumonía’. These terms were used because they are well understood in this setting and despite their equivalence to the medical terminology, there were no local alternatives which would carry the same meaning. It is therefore likely that there was some contamination of the informant's reporting for these algorithms by her attendance with the child at hospital. This potential, although unavoidable source of bias has probably resulted in an overestimate of the performance of interview algorithms compared with the likely performance in the community where most children die without attending hospital. This bias is of greater concern for pneumonia than for diarrhoea because an alternative algorithm for diarrhoea based on signs and symptoms (D.1) produced similar results to that for the local term, while there was no expert algorithm for pneumonia based on signs and symptoms which produced equivalent results. The validity of the use of local terms is dependent on local taxonomies of disease and cultural interpretations of those terms when translated into other languages. These difficulties are likely to be more pronounced elsewhere than in the setting for this study.

Another potential bias in a hospital validation study is that the mother may be less able to observe the signs and symptoms of her child with the same accuracy or detail while the child is in hospital as when the child is at home. It is common in Nicaragua for mothers to spend long periods of time at the hospital. However, this was not always the case, for example when the mother was working, or lived a long distance from the hospital. Informant's recall of signs and symptoms is also likely to deteriorate with time after the illness or death of the child. The accuracy of recall of the informant will depend not only on the time interval between the event of interest and the interview, but on the importance of the event for the respondent ( Kroeger 1983; Ross & Vaughan 1986). We might expect therefore, that the history leading to the death of a child will be more accurately remembered, and for longer periods of time, than the history of a nonfatal illness. Despite this, we were unable to detect any consistent differences between the results for children who died and those who survived.

The time between admission to hospital and home interview ranged from 1 to 22 months. This was because all interviews were conducted within a short period (March to August 1997) at the end of the recruitment period for logistic reasons. In future work, we intend to test the effect of different recall periods on the accuracy of estimates obtained from interviews by stratifying the sample by length of recall period and analysing each stratum separately.

Both the reference and the interview algorithms for diarrhoea were based on the caregiver's report, and the validation is therefore partly based on a comparison of the same information provided by the caregiver at two points in time. The combined algorithms for any diarrhoeal illness produced acceptable estimates of the true CSF. The individual algorithms for diarrhoeal illnesses did not perform as well, and the instrument did not distinguish adequately between diarrhoea and dysentery, nor between acute and persistent diarrhoea/dysentery. As a result, we do not recommend that the questionnaire is used to try to distinguish between acute and persistent diarrhoea/dysentery, or between diarrhoea and dysentery. The algorithm for diagnosis of diarrhoea-related causes (frequent liquid or watery or loose stools or local term for diarrhoea of any duration with or without blood in stools) is likely to produce reasonably accurate estimates of cause-specific morbidity or mortality provided the fraction is sufficiently high (say more than 30%), which is likely to be the case in most developing country settings. We would therefore recommend the use of the ‘best judgement’ questionnaire for diarrhoea, although further validation is required in other settings.

Most of these limitations are inherent in any hospital-based study of the validity of a retrospective mortality or morbidity questionnaire. They complicate the assessment of how the questionnaire will actually perform in a general population survey, where few of the illnesses or deaths may have included an admission to hospital. However, it is impractical to design a validation study using subjects who have not been in contact with the health services during their illness, because without this there would be no medical diagnosis to validate the questionnaire results against. Despite the rarity of malaria in young children in this area of Nicaragua, validity for the pneumonia algorithms was poor and the lack of agreement between the radiologists who read the X-ray films (which were the basis for assignment of hospital diagnoses for pneumonia) raises some doubt about the accuracy of the gold standard for pneumonia in this study. It has already been shown that interview diagnosis of pneumonia (and other causes) is problematic in malaria-endemic areas ( Todd et al. 1994 ), mainly because of the mutuality of some signs and symptoms. However, even in this nonmalaria setting, the algorithms applied were not able to estimate pneumonia mortality and morbidity accurately. The specificity of the second algorithm (E.2) for pneumonia was lower (37%) than that found for algorithms tested in previous validation studies (see Table 2), but the first algorithm (E.1) gave sensitivity and specificity comparable to previous studies. Although the specificity of algorithm E.2 was very low both in neonates and post neonates, the agreement in CSFs between the reference and interview diagnoses was poorer for neonates than for postneonates, mainly because of the lower true CSF in this group. The performance of the algorithm E.1 (local term for pneumonia) has only been tested in one other study of morbidity in neonates ( Kalter et al. 1999 ) with results (sensitivity 74%, specificity 87%) similar to ours. However, the local term used in Nicaragua (neumonía) may have particularly good operating characteristics in this setting because of its equivalence to the medical term. We would therefore recommend that algorithm E.1 for pneumonia should not be employed unless the local term for pneumonia can be validated in the community where the question is to be used. We were not able to identify an algorithm based on signs and symptoms which performed with adequate validity in all subgroups, and further work is necessary in order to identify and validate such an algorithm. Diagnosis of pneumonia using retrospective interview is likely to be even more problematic in areas where malaria is present. The effect on the accurate diagnosis of diarrhoea in such areas is likely to be less pronounced.

This study attempted to overcome problems present in some previous validation studies by using a prospective recruitment design and attempting to standardize and optimize the medical diagnosis used as the ‘gold standard’ within the validation. This included standardization of the recording of the child's symptoms and signs on admission and throughout the admission, and the use of standard diagnostic algorithms. Similarly, standard, predefined algorithms were applied to questionnaire data. Furthermore, exactly the same questionnaire and diagnostic criteria, both for the medical diagnoses and the questionnaire diagnoses, have been used in two companion studies in Bangladesh and Uganda conducted by other investigators. When the complete results from these studies become available, it will be possible to compare the performance of the same questionnaire in different settings with substantial differences in culture, language, and the disease mix.

This paper has discussed the use of the questionnaire for estimating morbidity and mortality from pneumonia and diarrhoea only. Further results to be presented shortly will discuss the usefulness of the questionnaire for a range of other childhood illnesses.

Acknowledgements

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References

This study was funded by the Department for International Development (DfID) of the United Kingdom. However, the DfID can accept no responsibility for any information provided or views expressed. The authors also gratefully acknowledge the contributions of the other members of the Child Verbal Autopsy working group: Dr Henry Kalter, Dr Robert Black (both Johns Hopkins University); Martha Anker (World Health Organization, Geneva); Dr Robert Snow (Wellcome Institute, Kenya); Gilly Maude (London School of Hygiene and Tropical Medicine); the study medical supervisors Dr Natalia Zambrana and Dr Martha Rivas; the study physicians Dr Patricia González and Dr Amelia Tijerino; the interviewing team: Nadia Lugo, Rosa Maria Pérez, Aurora Mendoza, Oriana Castillo, Yamileth González, Silvia Avellan, Ivania Haar, Avejaneth Saballos, Maria-Ines Leiva and Carlos Aburto; the data processing staff Martha Castaño, Sergio Baldelomar and Ernesto Rios; the directors and staff of hospitals Manuel de Jesús Rivera (Managua), Velez Paiz (Managua) and Hilario Sanchez (Masaya); Jeffrey Cassel (Director) and Ivania Lopez (Administrator) of the Instituto CentroAmericano de la Salud, Managua, Nicaragua.

References

  1. Top of page
  2. Abstract
  3. Introduction
  4. Materials and methods
  5. Results
  6. Discussion
  7. Acknowledgements
  8. References
  • Alonso PL, Lindsay SW, Armstrong JR et al. (1991) The effect of insecticide-treated bed nets on mortality of Gambian children. Lancet 337, 1499 1502.
  • Anker M (1997) The effect of misclassification error on reported cause-specific mortality fractions from verbal autopsy. International Journal of Epidemiology 26, 1090 1096.
  • Bang AT, Bang RA, Tale O et al. (1990) Reduction in pneumonia mortality and total childhood mortality by means of community-based intervention trial in Gadchiroli, India. Lancet 336, 201 206.
  • Barnish G, Maude GH, Bockarie MJ et al. (1993) Malaria in a rural area of Sierra Leone. I. Initial results. Annals of Tropical Medicine and Parasitology 87, 125 136.
  • Becker SR, Diop F, Thornton JN (1993) Infant and child mortality in two counties of Liberia: results of a survey in 1988 and trends since 1984. International Journal of Epidemiology 22 (Suppl.), S56 S63.
  • Boerma JT, Sommerfelt AE, Van Ginneken JK et al. (1994) Assessment of the quality of health data in DHS-I Surveys: an overview. In: An assessment of the quality of health data in DHS I Surveys. Macro International.
  • Chandramohan D, Maude GH, Rodrigues LC, Hayes RJ (1994) Verbal autopsies for adult deaths: issues in their development and validation. International Journal of Epidemiology 23, 213 222.
  • Chandramohan D, Maude GH, Rodrigues LC, Hayes RJ (1998) Verbal autopsies for adult deaths: their development and validation in a multicentre study. Tropical Medicine and International Health 3, 436 446.
  • Fauveau V, Wojtyniak B, Chakraborty J, Sarder AM, Briend A (1990) The effect of maternal and child health and family planning services on mortality: is prevention enough? British Medical Journal 301, 103 107.
  • Garcia Benavides F, Godoy C, Perez S, Bolumar F (1992) Codificacion multiple de las causas de muerte: de morir ‘por’ a morir ‘con’. Gaceta Sanitaria 6, 53 57.
  • Ghana Vast Study Team (1993) Vitamin A supplementation in northern Ghana: effects on clinic attendances, hospital admissions, and child mortality. Lancet 342, 7 12.
  • Harries AD (1993) Malaria: keeping the mosquitoes at bay. Lancet 342, 506 507.
  • Kalter H, Hossain M, Burnham G et al. (1999) Validation of caregiver interviews to diagnose common causes of severe neonatal illness. Paediatric and Perinatal Epidemiology 13, 99 113.
  • Kalter HD, Burnham G, Kolstad PR et al. (1997) Evaluation of clinical signs to diagnose anaemia in Uganda and Bangladesh, in areas with and without malaria. Bulletin of the World Health Organization 75 (Suppl.), S103 S111.
  • Kalter HD, Gray RH, Black RE, Gultiano SA (1990) Validation of postmortem interviews to ascertain selected causes of death in children. International Journal of Epidemiology 19, 380 386.
  • Kalter HD, Gray RH, Black RE, Gultiano SA (1991) Validation of the diagnosis of childhood morbidity using maternal health interviews. International Journal of Epidemiology 20, 193 198.
  • Kolstad PR, Burnham G, Kalter HD, Kenya Mugisha N, Black RE (1997) The integrated management of childhood illness in western Uganda. Bulletin of the World Health Organization 75 (Suppl.), S77 S85.
  • Kroeger A (1983) Health interview surveys in developing countries: a review of the methods and results. International Journal of Epidemiology 12, 465 481.
  • Landis J & Koch GG (1977) The measurement of observer agreement for categorical data. Biometrics 33, 159 174.
  • Marsh D, Husein K, Lobo M, Shah M, Luby S (1995) Verbal Autopsy in Karachi Slums: comparing single and multiple cause of child deaths. Health Policy and Planning 10, 395 403.
  • Mobley CC, Boerma JT, Titus S et al. (1996) Validation study of a verbal autopsy method for causes of childhood mortality in Namibia. Journal of Tropical Pediatrics 42, 365 369.
  • Osinski P (1994) Cited in WHO/UNICEF. Mesure de la mortalite infantile et juvenile, par cause et toutes causes: memorandum d'une reunion OMS/UNICEF. Bulletin of the World Health Organization 73, 149 156.
  • Quigley MA, Armstrong Schellenberg JRM, Snow RW (1996) Algorithms for verbal autopsies: a validation study in Kenyan children. Bulletin of the World Health Organization 74, 147 154.
  • Ross DA & Vaughan JP (1986) Health interview surveys in developing countries: a methodological review. Studies in Family Planning 17, 78 94.
  • Snow RW, Armstrong JR, Forster D et al. (1992) Childhood deaths in Africa: uses and limitations of verbal autopsies. Lancet 340, 351 355.
  • Todd JE, De Francisco A, O'dempsey TJ, Greenwood BM (1994) The limitations of verbal autopsy in a malaria-endemic region. Annals of Tropical Paediatrics 14, 31 36.
  • WHO/UNICEF (1995) Mesure de la mortalite infantile et juvenile, par cause et toutes causes: memorandum d'une reunion OMS / UNICEF. Bulletin of the World Health Organization 73, 149 156.