• No results found

Linkage studies aiming to match deaths captured in CRVS systems with deaths from other data sources have been conducted successfully in a number of countries. These studies were conducted with diverse purposes, including to assess the feasibility and quality of mortality data linkage among different data sources; verify the number of deaths or completeness of death reporting in a specified period and space; identify the optimal choice of personal identifiers for linkage; identify participant or population characteristics that compromise complete linkage and the sources of such bias; assess the reliability of cause-of-death data on death certificates; consider the implications of study findings for civil registration mortality statistics; as well as combinations of these reasons. Selected studies, with purposes similar to the purpose of the linkage study for this thesis (Section 4.5), are referred to below.

4.3.1 Assessing the feasibility, quality and optimal identifiers of linking CRVS and other mortality data

Studies to assess the feasibility, success, or quality of data linkage between CRVS and other data sources include that of Roos et al, linking data from the Canadian Mortality Data Base, Manitoba Health Services Commission (MHSC), and the Manitoba Cancer Foundation. A combination of family registration number, sex, birth year, and initial was used to match the records, with 96% of MHSC records being matched with Canadian Mortality Data Base records. The study showed that the record linkage rate improved significantly with improvements in data quality over time.198 Newman and Brown linked patient data from a hospital database with death data from the California Department of Health Services (Deaths of California residents) and from the United States (USA) Social Security Administration

99 (USA Deaths), using both unique (social security number) and non-unique identifiers (last name, first name, date of birth, middle initial, race, county of residence and sex). Linkage between civil registration and hospital data was conducted successfully, linking 97% of all in- hospital deaths, and 99% of deaths of patients with social security numbers. The authors reported linkage challenges among infants (< 1 year) due to more cases with missing identifiers (e.g. name and social security number) compared to cases involving older persons.178

Linking California state-wide hospital discharge records with a database of death certificates for all persons dying in California and California residents who die elsewhere, Zingmond and co-authors successfully linked 95% of in-hospital death records which contained social security numbers to the death records. Subsequent calculation of linkage accuracy showed an accuracy rate of 99%.182 Herrchen and colleagues successfully linked vital statistics birth data with hospital discharge data in California, USA. Using zip code, hospital code, date of birth, gender, payment source for hospital stay, neonate’s ethnicity, and type of delivery, a very high percentage, 98.6%, of the births were matched with patient records in the hospital discharge data.180

For residents of the Calgary Health Region of Alberta in Canada, Li et al linked death records from (a) the Vital Statistics Registry and the Alberta Health Care Insurance Plan Registry, and then from (b) the Vital Statistic Registry and hospital discharge data. Testing different combinations of identifiers, the investigators found that the combination of surname, sex and date of birth achieved the best linkage rate, i.e. 88% and 93%, respectively, for (a) and (b). The identical recording, or not, of residents’ unique lifetime Personal Health Number in the data sources used, enabled an assessment of the correct linkage rate in the study, achieving 97% and 99%, respectively, for (a) and (b). As in the work of Newman and Brown,178 Li and colleagues found many infant records with missing or inaccurate variables, and subsequently excluded infants in this study.184

4.3.2 Assessing numbers of deaths and completeness of death registration, and identifying sources of bias via record linkage

In Bohol province in the Philippines, Carter et al collected death records from civil registration offices, hospitals and health centres, and parish churches for the years 2002 to 2007, and manually matched the records. With a capture-recapture analysis, a total number of deaths was estimated. Completeness of death registration in the civil registration system was estimated at

100 72%, and it was found that 5 – 10% of total deaths were not reported under any system.186 In Tonga, Hufanga et al obtained and manually matched death records from the Civil Registry and the Ministry of Health, the latter receiving death records from the Health Information System and the Reproductive Health System. Using a capture-recapture analysis, the authors were able to estimate the total numbers of deaths for Tonga for the years 2001 – 2004 and 2005 – 2009, and estimated completeness at 98% and 88% for these periods, respectively.185 In both these studies, completeness of death recording varied by age, with child deaths less likely than adult deaths to be recorded.185,186

In their study of Calgary residents in Canada, linking the Vital Statistics Registry and the Alberta Health Care Insurance Plan Registry, Li et al found that the linkage rate varied by age. Over the four-year study period the linkage rate was consistently lower for the age groups 10 – 19 and 20 – 29 years, and considerably lower for the age group 1-9 years.184 The work of Zingmond and others showed that, when linking hospital discharge data containing a social security number with death certificate data in California, the infants and older adults were more likely than those 1 to 64 years old to have unlinked records; unlinked record rates were two to three times higher for older females compared to older males among those 65 years or older; and Hispanics and non-Hispanic Blacks were consistently more likely than non-Hispanic Whites or Asians to have unlinked records.182

In their review, Bohensky et al identified a number of participant characteristics, such as age, gender, ethnicity, socio-economic and health status, that were associated with sub-optimal linkage which influences the linkage rate.204 For example, nine studies in their review indicated that people in lower socio-economic groupings and with lower levels of education were less likely to have matched records, mainly due to lower consent rates for data linkage and less complete data for people of lower socio-economic status.204 Seven studies showed that people in minority ethnic groups had lower linkage rates. Reasons include a lower likelihood of having a social security number recorded, minority groups being treated at health facilities with less complete data, and lower rates of consent to data linkage.204 These findings demonstrate that participant characteristics can influence the completeness of data linkage, which may result in systematic biases in reported outcomes such as the relative risk of mortality by age, ethnicity or socio-economic status. This highlights the need for researchers and decision-makers to consider such biases when interpreting and reporting linkage results.204,226,227

101 4.3.3 Assessing the reliability of cause-of-death data in mortality data sources through

record linkage

Linking data from death certificates of reproductive-age women with foetal death and live birth records from the Vital Statistics Administration in Maryland, USA, and reviewing medical records for evidence of current or recent pregnancy at the time of death, Horon found that only 62% of maternal deaths were recorded on death certificates, implying a substantial under- estimation of maternal deaths when death certificates alone were used.179 Amin and others linked cases of hepatitis B and hepatitis C notified to the New South Wales Health Department Notifiable Diseases Database (NDD) to deaths in the Australian National Death Index (NDI). Probabilistic linkage was performed for the years 1990 - 2002. For hepatitis B and hepatitis C deaths, cancers (38%) and external causes (28%) were among the most frequently misclassified underlying causes of death reported on death certificates, pointing to considerable under- reporting of hepatitis B and C.228

In two research papers,229,230 Johansson and Westerling report the linkage of death certificate data for 1995 from the Swedish National Cause of Death Register with hospital discharge data from the Swedish Hospital Discharge Register. The authors found that the last main hospital diagnosis and underlying cause of death agreed in less than half (46%) of 69,818 cases, but indicated that cause-of-death information on hospital discharge records and death certificates supplemented each other well, and that the linkage was useful in that conditions absent from one data source may be found in the other.229,230 In a similar exercise, Klijs et al linked death certificate cause-of-death data from the Dutch Cause of Death Registration and hospital discharge data from the National Medical Registration, using a unique personal identification number.181 They aimed to assess the extent to which registered causes of death represent morbidity at the end of life, and concluded that, except for cancers, registered causes poorly represented diseases recorded during hospitalisation. Similar to the work of Johansson and Westerling,229,230 Klijs et al181 compared underlying causes from death certificates with main

causes on hospital discharge records. As the definition and purpose of underlying and main cause on these two types of records do not concur, the low levels of agreement were to be

anticipated.

Mühlhauser et al compared multiple data sources, including interviews with family physicians and relatives, and autopsy, physician and hospital records, related to deaths from a study cohort of 3,674 patients with Type 1 diabetes who were treated with insulin before age 31 and

102 monitored for a mean of 10 years. The authors concluded that death certificates were not reliable sources of information for Type I diabetes. Substantial misclassification occurred to conditions such as cardiac and cerebrovascular disease, and diabetes was mentioned on only 71% of death certificates.231 Findings of unreliable diabetes diagnoses on civil registration death certificates were also reported by Pattaraarchachai et al232 in Thailand, finding about six times fewer deaths from diabetes as a registered underlying cause among Thai hospital deaths, compared to the number of deaths with diabetes that were ascertained via medical record review.232 For non-hospital deaths in Thailand too, Polprasert et al found that diabetes was under-recorded as an underlying cause of death in civil registration data compared to verbal autopsy data.233