Study design
A cross-sectional study was designed to verify registered causes of death in a nationally representative, multistage stratified cluster sample of deaths that occurred in Malaysia during 2013. The study sample frame comprised the dataset of 142,202 deaths in 2013 by age, sex and cause, available from the Malaysian National Department of Statistics [3].
Sample size
The sample size was determined according to the approach presented in the article by Begg et al. [6] to estimate the optimal sample size of deaths required for measurement of specific population-based mortality indicators within defined margins of error. The sample size estimation is based on a statistical model which uses inputs of age-specific death rates and income per capita for the study population, to predict a proportionate distribution of cause-patterns of mortality according to three broad cause groups (communicable diseases, non-communicable diseases, and injuries) by age and sex [7]. Begg et al. estimated sample sizes for three population examples, representing three different strata from the World Health Organization’s (WHO) classification of countries according to national mortality characteristics [8]. According to this stratification, Malaysia belongs to the stratum of countries with ‘low child and low adult’ mortality levels [8]. For this stratum, they estimated that a representative sample of approximately 11,000 deaths would enable the measurement of cause-specific mortality rates according to the three broad cause groups by specific 5-year age-sex groups of interest, within a 15% relative standard error [see Table 4 in [6]]. We chose this estimate of sample size for this study.
As a cluster sample is more efficient from logistical aspects, the primary sampling unit was set as the health district of Malaysia, for which a design effect of 1.25 was applied to inflate this sample size to 13,750 deaths. Finally, the research team expected about 10% dropout from the household enquiry into causes of death due to migration, non-availability of respondents, or refusal for participation. As a result, after factoring in this expected loss to follow up; a total sample size of about 15,000 deaths was estimated for this study.
A multistage stratified cluster sampling strategy was used to select 19 out of the total 144 health districts in Malaysia for this study. All deaths that were registered in the selected districts in 2013 were included in the study, and these amounted to a total of 14,497 deaths. The selected sample of deaths was tested and found adequate for national representativeness in terms of age-sex distribution, proportions of medically certified and non-medically certified deaths, and distributions of registered causes of medically and non-medically certified deaths.
Data collection and processing
For each of the selected districts, the National Registration Department provided a list of all deaths that occurred in the year 2013. For each death, the details of identity, address, reporting institution, and cause of death as determined at registration were given to the District Health Office (DHO) for reinvestigation of the cause of death using two potential approaches. Firstly; all deaths were followed up with a household visit to conduct verbal autopsy (VA) interviews to collect information to ascertain and verify the registered cause(s) of death. Secondly, deaths that had occurred in hospitals were also followed back to review medical records (MR) to verify the registered cause(s) of death. Ethical approval for this study was obtained from the Malaysian Medical Research Ethic Committee with the registration number NMRR-13-1369-18,689. Informed written consent was taken from interviewee before face-to-face interview. Consent was also obtained from the interviewee to review the medical records of the deceased for deaths that had occurred in a medical facility.
For the VA component, details of the development of VA questionnaires, interviewer training and field procedures are provided elsewhere [9]. In brief, international standard VA questionnaires prepared by the WHO [10] and the Population Health Metrics Research Consortium [11] were reviewed and used as a basis to develop a set of questionnaires adapted to the Malaysian context. Following a set program of training, selected verbal autopsy interviewers (attached to the local DHO) were provided with details of the address of the deceased within their health clinic areas. They then visited the home and after obtaining informed consent, conducted a face-to-face interview with a family member of the deceased. VA interviewers were blinded from the cause of death in the registration database, so that the VA interview remained free from bias from this aspect. Completed questionnaires were returned to the DHO, where supervisors reviewed the completed questionnaires for any missing variables or incomplete data. Supervisors also provided field support to staff, where necessary.
All completed questionnaires were submitted to teams of physicians comprising public health specialists and family medicine specialists, for review and assignment of causes of death. Physician reviewers of VA underwent a 3-day training program which covered the principles and practice for assigning causes of death from VA. The training included sessions on diagnostic guidelines for common causes of death, principles of death certification and rules for underlying causes of death, as well as practical exercises and tests of inter-rater reliability to establish consistency between reviewers. Subsequently, each VA questionnaire was reviewed by one physician, who reviewed the questionnaire and assigned causes of death in a format based on the international medical death certificate. In addition, all completed VA questionnaires and death certificates were reviewed by a central team of experienced public health physicians at the Institute for Public Health, Ministry of Health Malaysia, who verified the diagnoses on each VA death certificate, assigned codes from the Tenth revision of the International Classification of Disease and Health Related Problems (ICD-10), and selected the underlying cause of death from VA [4, 12].
As mentioned above, for hospital deaths in the study sample, the second approach involved review of medical records (MR). These deaths were independently followed up for medical record abstraction and review. Lists of deaths from each health institution were prepared, and trained staff reviewed and abstracted information from medical records using a specially designed MR abstraction form. The form included sections to record essential relevant details from the clinical history, findings from physical examination and laboratory and imaging investigations, and details of the clinical course of events culminating in death. Completed MR abstraction forms were reviewed by a member from a panel of medical specialists trained in medical death certification, who certified the cause(s) of death using the standard international medical death certificate. Each form was reviewed by one physician only to determine the cause(s) of death, with the facility to seek a second opinion, if necessary, All completed cause of death certificates (from MR) were reviewed by trained coders, who first assigned ICD-10 codes for each recorded cause of death, and subsequently applied the ICD mortality rules to select the underlying cause for each MR death certificate [4, 12].
Thus, each death in the study sample was assigned a cause of death in the vital registration (VR) data, an underlying cause from the VA investigation, and for those deaths that had occurred in health facilities, an underlying cause from the medical record (MR). The underlying causes coded according to the 3-character ICD code from VR, VA, and MR were then aggregated to the WHO Global Burden of Disease categories [13] for descriptive and comparative analyses.
Data analysis
The data analysis was done for both arms of the study. For non-hospital deaths, the VA diagnoses were compared with the causes recorded at registration (VR diagnoses), to understand the degree of the reliability of the registration data. Patterns of misclassification between VA and VR diagnoses were assessed, along with the net misclassifications for specific cause categories. These misclassification patterns were analyzed in terms of overall changes in cause-specific mortality proportions for each cause, as determined from the VA study. Eventually, the VA cause-specific mortality fractions for the sample of non-medically certified deaths were adopted for estimating the overall cause-specific patterns for non-medically certified deaths in Malaysia.
Similar analyses were also conducted for the sample of medically certified deaths, for which the diagnoses of underlying causes of death from the MR review were compared with the registration diagnoses. In this analysis, the validity of the registration diagnoses for each cause was assessed in terms of sensitivity, specificity and positive predictive value, using the medical record review diagnoses as the reference standard (data not shown). The net misclassifications for specific cause categories were analysed in terms of overall changes in cause-specific mortality proportions that resulted from the MR review. The MR derived cause-specific mortality fractions were adopted for estimating the overall cause-specific patterns for medically certified deaths in Malaysia.
Cause specific mortality estimates
The revised cause-specific mortality proportions for each cause in the hospital death and home death components of the study sample were first applied to the total number of registered deaths in each component separately for males and females, to generate the total estimated number of deaths from each cause. Subsequently, these total numbers of deaths from each cause for each sex were weighted according to the age distributions for each cause as observed in the registration data, separately for hospital and home deaths. This step resulted in the corrected numbers of deaths by age, sex and cause for the hospital and home deaths. These revised death numbers were then summed across the two components to derive the preliminary national mortality estimates by age, sex and cause. Subsequently, standard redistribution algorithms used in the WHO Burden of Disease methodology were applied to redistribute the remaining deaths with ill-defined causes, as well as those with codes from non-specific cardiovascular diseases and cancers without mention of primary site [14].
After corrections for ill-defined and non-specific codes, the resultant estimates were first converted to proportionate distributions by cause for each sex-age category (i.e. the proportionate distribution of causes for deaths in males aged 5–9 years, males age 10–14 years and so on). Subsequently, these proportionate distributions by cause for each sex-age category were fitted to the total national registered numbers of deaths for the corresponding sex-age category, to generate the resultant final estimates of deaths in 2013 by age, sex and cause. These final estimates were analyzed to derive the rank order and magnitude of leading causes of death (for all ages together) by sex at the national level. These rank order estimates were compared with the rank orders from the vital registration data, to understand the influence of the field research on the overall rank structure and proportional magnitude of cause-specific mortality in males and females separately.
Directly standardized mortality rates
We also used our study results to estimate direct age standardized death rates (ASDR) for specific causes in males and females. We used national population estimates by age and sex for Malaysia for 2013, obtained from the National Statistics Office [3], to first derive the observed death rates. These were based on the actual deaths we recorded in our sample, and the number of individuals in the local denominator population for each age group scaled down to our sample size. We then used the WHO 2000 population standard [15] as the reference to derive direct ASDRs for each cause by sex. Finally, we used the methodology published by the Association of Public Health Observatories, Public Health England [16] to compute the 95% confidence intervals for the ASDRs The empirically derived Malaysian cause-specific ASDRs by sex for 2013 from our study were compared to imputed estimates of ASDRs derived for Malaysia in 2013 by the Global Burden of Disease 2016 study, conducted by the Institute of Health Metrics and Evaluation, USA [17].