Verifying causes of death in Thailand: rationale and methods for empirical investigation

A brief overview of civil registration and vital statistics in Thailand helps place the research study into context. Birth and death registration in Thailand commenced in 1916, supplemented by the introduction of the household register in 1956, a copy of which is issued to the household as proof of civil status [9]. In 1991, a revised Civil Registration Act nominated the Bureau of Registration Administration, Ministry of Interior, as the central agency responsible for civil registration through a network of offices at local, district, municipality, and provincial levels and stipulated the requirement for death notification within 24 hours [10]. Causes of deaths in hospitals are recorded using a Thai version of the standard International Form of Medical Certificate of Causes of Death, with an additional column in which the certifying physician records one cause in Thai language to be used for registration purposes. It is this cause that is entered in the national registration electronic database. For unnatural (i.e., injury) deaths, causes are certified following forensic investigation by a physician using a similar process. For deaths occurring outside hospitals (about 65% of all deaths), local registrars inquire about the cause of death from family members and also seek documentary evidence, wherever possible, from previous hospitalization or medical attention during the illness preceding death. Local registrars then record the reported cause of death in Thai. After computerization, a complete extract of the database (with a single cause in Thai for each death) is transferred to the Ministry of Public Health, where the causes are coded according to ICD-10 and tabulated by age, sex, and cause for subsequent dissemination and analysis [5].

Thailand is one of the few Asian countries that have submitted data on a regular basis to the World Health Organization since 1950 [2]. A recent assessment of mortality statistics in Thailand identified that the sound legal framework and institutional capacity for implementation are strengths that result in regular and timely data [5]. Computerization of civil registration over the past two decades has vastly improved efficiency of registration services as well as compilation of statistics. However, mortality data are limited by the incompleteness of death registration and more so by the high proportion of ill-defined causes of death [5]. These limitations also preclude the application of detailed technical criteria to assess additional aspects of data quality in terms of validity and reliability [11].

In recent years, a gradual increase in the number of registered deaths has been noted, from 264,350 in 1991 to 395,374 in 2005 [12]. This increase is attributable to various factors, including the 1991 revisions to the legal framework, administrative reforms to registration procedures, and computerized compilation of data, all of which came into effect over the period up to 1996 [10]. However, the current level of completeness of adult death registration remains a subject of ongoing research, with the application of different demographic methods yielding completeness estimates ranging from 80% to 95% [13, 15]. While there is general consensus regarding the extent of under-registration of child deaths (approximately 50%) [5, 10, 13, 16], the overall completeness of death registration in Thailand remains uncertain.

In terms of recorded causes of death, about 40% are routinely coded to symptoms, signs, and ill-defined conditions [17]. The bulk of these ill-defined deaths occur outside health facilities, although in-hospital cause-of-death attribution is also problematic, arising from the process of recording the single cause of death for registration in Thai language. Hence, the validity of registered causes of death is questionable and greatly limits their public health utility. A study conducted by Boonthai et al found low validity of diagnoses on death certificates issued in a sample of hospital deaths during the period 1967-1984 [18]. Choprapawon and colleagues later conducted a detailed verification of 47,632 deaths that occurred in 15 provinces in Thailand during 1997-1999, using a combination of verbal autopsy interviews and medical record review, as applicable [10]. Their findings indicated poor validity of registered causes for deaths in hospitals as well as for deaths occurring at home.

In view of the existing problems with cause-of-death ascertainment, further reforms to the Thai death registration system were pilot-tested in 18 provinces from 2001 to 2003. Direct support from the medical profession was sought to certify causes for nonhospital deaths, using information from previous contact with health services and/or limited verbal autopsy interviews [10]. However, potential medico-legal implications from attributing causes without having attended to the deceased restricted the implementation of this reform by the medical profession [5].

Given these circumstances, the Thai Ministry of Public Health proposed a long-term strategy to improve cause-of-death statistics using two approaches [19]. The ministry proposed that for deaths in hospitals, capacity should be strengthened to improve accuracy of medical certification of cause of death as well as selection and coding of underlying causes of death according to ICD principles. For home deaths, local health personnel should be trained in the use of available medical records and verbal autopsy procedures to improve accuracy in the recording of causes of death at registration. As an interim measure, the Thai Ministry of Health identified the need to conduct research studies to verify causes in a national sample of registered deaths every five years and to use this research to derive periodic estimates of cause-specific mortality patterns to monitor health status and inform health policy and planning in Thailand [5, 19]. This set of articles describes the methods and findings from such research on a sample of deaths that occurred in 2005. This study forms the basis for a broader study to estimate the relative cost-effectiveness of various intervention choices in Thailand for which corrected age-, sex-, and cause-specific death rates were required to more reliably estimate the burden of disease applicable to each intervention [20].

The overall goal of the study was to derive the best estimate of cause-specific mortality patterns in Thailand for 2005. Specific objectives were to:

In addition, it was anticipated that the research activity would also accomplish a longer-term outcome, namely building capacity among Thai health professionals (physicians, paramedical staff, biostatisticians, and epidemiologists) to critically assess vital statistics data and improve the quality of causes of death recorded at registration in Thailand.

A cross-sectional study was designed to verify causes of death for a nationally representative, multistage stratified cluster sample of deaths that occurred in Thailand during 2005. The sample was drawn from the mortality database maintained by the Bureau of Policy and Strategy, Ministry of Public Health [12]. The sampling unit was a registered death of a Thai citizen, identified by assigned national identification number, who was a permanent resident in one of the sample provinces included in the study. The cause or causes for each study death were investigated through verbal autopsy (VA). For study deaths that had occurred in a hospital, relevant medical records were accessed and reviewed to derive reference diagnoses. These reference diagnoses were used to validate and correct registered causes in the study sample, as well as to assess the validation characteristics of VA procedures for individual causes of death. For deaths that had occurred at home, VA diagnoses were used to estimate cause-specific mortality patterns, with subsequent adjustments of biases for individual causes of death, as observed from the VA validation component. Proportionate mortality distributions for health facility deaths in the study sample were used to estimate mortality patterns for all deaths in hospitals in Thailand (about 140,000 deaths each year). Similarly, VA-based proportionate mortality distributions for deaths outside hospitals were used to estimate mortality patterns for such deaths for 2005 (about 254,000 deaths). The overall proportionate (by cause) mortality estimates from the two settings were then applied to the national estimate of total mortality derived from demographic analyses to obtain corrected national cause-specific mortality estimates by age and sex for Thailand in 2005.

Several considerations influence sample size estimation for a study of this nature. While statistical considerations are paramount when computing sample size, the ultimate choice of such parameters is guided by available financial and human resources as well as time constraints. Also, epidemiological considerations influence the inclusion of important causes of death in the study. Such logistical and epidemiological considerations usually limit the number of causes of interest to a selection of leading causes of death.

In this study, generalizability of findings is essential to accomplish the overall objective to derive cause-specific mortality estimates for Thailand. Intuitively, a close fit between the proportionate distributions of causes of death in the registration data and the study sample would support generalizability. The 20 leading causes of death in the Thai registration data (based on the WHO Mortality Tabulation List 1 [21]) accounted for more than 85% of all deaths, including 38.2% from the leading registered cause of death--"symptoms, signs, and ill-defined conditions." Therefore, we chose a sample of deaths that matched the proportionate distribution for the 20 leading causes in the registration data.

The proportion of deaths from the 21st leading cause of death in the Thai registration data was 0.0118. We assumed that the proportionate distribution by cause in a random sample of deaths adequate to measure the proportion of deaths from this 21st leading cause (within defined statistical parameters of error and significance) is likely to match the proportionate distribution for the 20 leading causes of death in the sampling frame. Therefore, we used this proportion for the 21st leading cause to compute the study sample size (n) as follows:

where π = 0.0118, within a selected margin of error d = 0.0025%, at the 95% level of confidence. This suggested that a random sample of 7,168 deaths would be required to accurately measure the proportion of the 21st leading cause in registration data within the given margin of error and level of confidence, based on prior information as to the proportion of interest. Further, since this sample was designed to be implemented in clusters for logistical and operational reasons, a design effect of 1.4 was applied, yielding an overall sample size of just over 10,000 deaths (10,035) [22, 23].

The nationally representative study sample was selected from the national death registration database using a multistage stratified clustered approach. At the first stage, Thailand was stratified into four broad regions - Northeast, North, Central, and South--as well as Bangkok, with the total sample of 10,000 deaths being distributed across regions according to the proportion of deaths from each region in the national death registration data. Each of the four regional samples was inflated by 15%, and by 50% for Bangkok, to account for potential losses to follow up, as suggested by the pilot study for this research and earlier research in Bangkok [10]. Subsequently, in each of the four broad regions, provinces were ordered according to numbers of registered deaths in 2005 and divided into two strata at the 50^th percentile. One province was randomly selected from each stratum, leading to two study provinces from each of the four broad regions. The adjusted (inflated) regional sample was then distributed between the two provinces proportionate to the number of deaths registered in 2005. Figure 1 shows the geographical distribution of the study provinces in Thailand.

In each study province, districts were ranked according to the number of registered deaths in 2005 and similarly divided into two strata at the 50^th percentile. In order to ensure representation of predominantly urban or rural communities as well as differential access to health facilities, districts within the provincial study sample were selected according to probability proportionate to size. In each stratum, a district was randomly selected at first, and the stratum-specific sample was proportionately allocated to the selected district according to probability proportionate to the number of deaths registered in 2005. Additional districts were selected one at a time with similar allocation of the study sample, cumulating the sample across selected districts until the province- stratum sample was attained. Within each selected district, the study deaths were randomly selected without replacement from all deaths registered during 2005.

For Bangkok, the study sample was distributed across the inner-, middle-, and outer-concentric zones of the Bangkok metropolitan area, and one district was randomly selected from each zone for inclusion in the study. The district-specific samples were determined according to probability proportionate to size, and within each district, the study sample was selected without replacement. Based on this strategy, a total of 11,984 deaths were selected into the study, distributed across 25 districts located in eight provinces from among the four broad regions of Thailand, and three districts from Bangkok.

For each sampled death, relevant information on the age, sex, and identity of the deceased, date of death, as well as the address and the underlying cause of death recorded at registration (the VR cause of death) was extracted from the national death registration database for 2005. Subsequently, an initial visit was undertaken to the households of each of the deceased, to:

In a subsample of 2,232 hospital deaths, an audit procedure was conducted to assess the quality of information recorded on medical certificates of cause of death. Detailed data collection procedures, relevant training support, data processing, and management protocols are described in the separate articles for each arm of the study, along with key findings and their implications [6, 8]. A detailed quality-control mechanism was implemented to ensure accuracy in the process of selection and coding of underlying causes of death (Figure 2). Measures were instituted to detect and correct errors in this process from a range of potential sources, and summary indices were developed to assess the overall quality of cause-of-death certification and ICD coding from either medical records review [6] or verbal autopsy questionnaires [7].

Underlying causes of death were derived for each death in the sample and classified according to the 10th revision of the International Classification of Diseases and Health Related Problems (ICD-10). These data were then aggregated to the ICD-10 Mortality Tabulation List 1 consisting of 103 cause categories, and all subsequent descriptive and comparative analyses were conducted using these aggregated data.

Figure 3 describes the different elements of data analyses pertaining to each arm of the study and the process used to derive cause-specific mortality estimates for Thailand using the study findings. For deaths in hospitals, the validation characteristics (concordance, sensitivity, and positive predictive value) of registration diagnoses were derived for leading causes of death, using the MR diagnoses as the reference standard. The study findings were subsequently applied to estimate cause-specific mortality patterns for all hospital deaths in the registration data. In a subsample of hospital deaths for which a VA diagnosis was also available, validation characteristics were derived for VA methods (the VA validation study) [6, 7].

Finally, mortality estimates from the two arms of the study were summed and adjusted to fit the overall numbers of deaths by age and sex estimated from demographic analysis to derive detailed cause-specific mortality estimates for Thailand in 2005 [8].