nova página do texto(beta)
Inglês (pdf)
Artigo em XML
Referências do artigo
Tradução automática
Enviar este artigo por email
Citado por SciELO 
Similares em
SciELO 
Permalink
Introduction
In Mexico, the Subsistema Epidemiológico y Estadístico de Defunciones (SEED) and the Instituto Nacional de Estadística y Geografía (Inegi) independently collect and code death certificates. These death registries were created for official statistics and mortality surveillance. Their usefulness for the linkage of individual records to external databases is unclear. As public health research capacity is strengthened in Mexico, understanding how these databases can be used for vital status assessment is essential. Following a previously established strategy,1,2 we sought to estimate the sensitivity and specificity of these databases for the identification of vital status comparing them to an epidemiologic cohort with active follow-up using a publicly available record-linkage tool.
Materials and methods
SEED
SEED is Mexico’s mortality surveillance tool. Until 2014, standardized coders in health districts coded death certificates using International Classification of Diseases (ICD)-10 codes and manually attributed the underlying cause of death.3 The database was updated continuously.4 We accessed 2006-2014 databases after approval by the Instituto Nacional de Salud Pública (INSP) (CI-249-2016-1396).
Inegi
Inegi generates Mexico’s official death statistics and processes death certificates independently of SEED. Information from death certificates is coded at regional offices using an adapted version of the Mortality Medical Data System for automated data entry and classification based on ICD-10 codes. Inegi’s central office validates the databases periodically.5 We obtained on-site access at Inegi to death records between 2006 and 2014.
Mexican Teachers’ Cohort (MTC)
The MTC is a prospective cohort of 115 314 female teachers established in 2006-2008 with a follow-up response of 83% for the 2011-2014 questionnaire cycle.6 Deaths were identified yearly through data linkage to human resource databases from education authorities, a pension fund database, and next-of-kin reports. As of December 31, 2014, we had identified 581 deaths and randomly selected a sample of 575 participants known to be alive (i.e., answered follow-up questionnaire and were not reported dead). At the time, the national identifier (Clave Única de Registro de Población, CURP) were available for 70% of study participants.
Mortality linkage
We used Centers for Disease Control and Prevention’s (CDC) probabilistic record linkage software Registry Plus™ Link Plus (Version 2.0) to search for deaths and women known to be alive.7 A probability score for record pairs was generated based on the probability that the matched records belonged to the same person. We hoped to identify the ideal probability score cut-point by using 5, 7.5, and 10 (recommended range is 7-10). For SEED, we used female sex as a matching variable along with: CURP (on average 12.8% of records), names (name and two last names), and CURP plus names. For Inegi, we used female sex and names (CURP were unavailable). The linkage software used a phonetic algorithm known to accommodate Hispanic names.8 Potential matches were manually confirmed using date and state of birth. Minor mismatches on a single field were allowed. Two independent reviewers conducted the manual assessment for 2010 and reached perfect reproducibility. See supplemental materials for more information on our linkage process.9 We estimated the sensitivity and specificity (and 95% confidence intervals; 95%CI) of SEED (CURP, names, CURP + names), Inegi (names), and SEED (names) plus Inegi (names) to identify deaths and women known to be alive. We sought to identify a process to minimize the manual review burden. We assessed discrepancies in the underlying cause of death by comparing deaths found on both databases. Major discrepancies were a difference in the first digit of the ICD-10 or cancer site. We subclassified discrepancies in adjudication when there was no difference between the multiple causes and in coding when multiple causes´ codes differed.
Results
We found 509 out of 581 deaths: 305 appeared on both databases, while 72 were not identified by either. Fifty of these unidentified deaths (69%) occurred in the two most recent years (2013 and 2014). Three participants known to be alive were found in both databases.
Using the most inclusive probability matching score, the SEED achieved the highest sensitivity when using names (84.9% [95%CI: 81.7, 87.7]), although this method had the most records to review. When using only CURP, we achieved a low sensitivity (due to missingness), and when using CURP + names, the number of potential matches, relative to names, decreased without affecting the sensitivity substantially. In contrast, Inegi had a sensitivity of 51.1% (95%CI: 47.0, 55.3, 118 840 matched records). Using both databases, a sensitivity of 87.6% (95%CI: 84.7, 90.2) and specificity of 99.3% (95%CI: 98.2, 99.8) were achieved by reviewing 224 645 matched records. We found that using 7.5 and 10 probability scores dramatically reduced the number of records for review (38 944 and 8 595, respectively) while keeping the sensitivity above 80% (table I). Among the 305 records identified in both registries, the underlying cause of death was discordant in 18.3% (n=56), but only 10.8% (n=33) were major discrepancies (mainly occurring in cancer). Overall, we found that potential errors in adjudication and coding were roughly similar in frequency (adjudication 30; coding 26).
Table I Specificity, sensitivity and possible matched revised by search method. Mexico, 2006-2014*
|
SEED |
Inegi |
SEED (N)+ Inegi (N) |
|||
|
CURP |
Name |
Name + CURP |
Name |
||
|
Cutoff score 5 | |||||
|
Sensitivity (95%CI) |
6.5 (4.7,8.9) |
84.9 (81.7,87.7) |
84.7 (81.7,87.7) |
51.1 (47.0,55.3) |
87.6 (84.7,90.2) |
|
Specificity (95%CI) |
99.3 (98.2,99.8) |
99.3 (98.2,99.8) |
99.3 (98.2,99.8) |
99.5 (98.5,99.9) |
99.3 (98.2,99.8) |
|
Potential matches |
42 |
105 805 |
50 345 |
118 840 |
224 645 |
|
Cutoff score 7.5 | |||||
|
Sensitivity (95%CI) |
0.0 |
81.8 (77.7,84.2) |
81.1 (77.7,84.2) |
49.6 (45.3,53.7) |
84.5 (81.3,87.4) |
|
Specificity (95%CI) |
100.0 |
99.3 (98.2,99.8) |
99.3 (98.2,99.8) |
99.7 (98.8,100) |
99.3 (98.2,99.8) |
|
Potential matches |
0 |
18 370 |
7 985 |
20 574 |
38 944 |
|
Cutoff score 10 | |||||
|
Sensitivity (95%CI) |
0.0 |
78.8 (75.3,82.1) |
76.9 (73.3,80.3) |
47.9 (43.7,52.0) |
81.6 (78.2,84.7) |
|
Specificity (95%CI) |
100.0 |
99.3 (98.2,99.8) |
99.3 (98.2,99.8) |
99.7 (98.8,100) |
99.3 (98.2,99.8) |
|
Potential matches |
0 |
3 167 |
1 941 |
5 428 |
8 595 |
* All search methods included gender variable
SEED: Subsistema Epidemiológico y Estadístico de Defunciones
Inegi: Instituto Nacional de Estadística y Geografía
CURP: Clave Única de Registro de Población
We conducted a stepwise method aimed at reducing the manual record review burden where matched records were sequentially removed. We first used CURP as the matching variable in SEED and identified 38 deaths (table II). After removing these deaths, we repeated the matching process and matched-record removal using SEED (CURP + names), SEED (names only), and Inegi (names) sequentially. This method identified the same death as our previous method with an important reduction in manual record review (from 224 645 records to 23 412).
Table II Stepwise matching strategy. Mexico, 2006-2014
|
Steps |
Deceased participants searched |
Deceased participants found |
Potential matches found |
|
1. SEED, CURP |
581 |
38 |
42 |
|
2. SEED, CURP + names |
543 |
441 |
7 868 |
|
3. SEED, names only |
102 |
4 |
5 989 |
|
4. Inegi, names |
98 |
26 |
9 513 |
|
Total |
509 |
23 412 |
SEED: Subsistema Epidemiológico y Estadístico de Defunciones
Inegi: Instituto Nacional de Estadística y Geografía
CURP: Clave Única de Registro de Población
Discussion
Linkage of Mexican mortality registries using a publicly available probability record matching tool may be useful to determine vital status in epidemiologic cohorts. Strategies to increase the efficiency of manual record review can be implemented.
The sensitivity of SEED and Inegi for cohort mortality follow-up approached 90%, which is somewhat lower than the 97-98% observed in the U.S National Death Index.1,2 However, this is probably an underestimate. Most unidentified deaths were most likely due to a mortality reporting lag. Also, we expect the sensitivity to have increased after 2014 because usage of the national identifier has increased with each year. SEED was significantly better at identifying deaths relative to Inegi. This was expected since the latter’s collection of death certificate data probably does not emphasize including identifiers because these are unnecessary for national statistics. Major discrepancies between registries were among the lower range reported in the literature and are consistent with prior work.9,10,11
Our study has limitations.10,11,12 We assumed as gold standard for vital status, employer and pension fund manager information, and next-of-kin reports. While 95% of participants considered alive answered a follow-up questionnaire between 2011 and 2014, three participants found in both databases were misclassified as alive by the gold standard. Our analysis assumes that the probabilistic record linkage tool is adequate for Mexico. While this tool accommodates Hispanic names, this and other database characteristics may have affected our capacity to identify some deaths. Finally, our study only included middle-aged women and results may not be fully transportable to children or men.
Conclusion
Our study provides initial evidence that national mortality databases can be used for mortality follow-up with reasonable use of human resources. SEED performs better than Inegi, but when possible, these registries should be used jointly. Our results require confirmation in other Mexican prospective studies that include different populations and age groups.
