Analytical Studies: Methods and References
The 1996 CanCHEC: Canadian Census Health and Environment Cohort Profile

Acknowledgements

The authors wish to acknowledge the contributions of Hélène Lamadeleine who assisted with manual review of census records, and of Jacques Dubois and Barry Zaid, who assisted in accessing census microfiche files for the manual validation.

Abstract

This paper describes the creation of the 1996 Canadian Census Health and Environment Cohort (CanCHEC)—3.57 million respondents to the census long-form questionnaire who were retrospectively followed for mortality and mobility for 16.6 years from 1996 to 2012. The 1996 CanCHEC was limited to census respondents who were aged 19 or older on Census Day (May 14, 1996), were residents of Canada, were not residents of institutions, and had filed an income tax return. These respondents were linked to death records from the Canadian Mortality Database or to the T1 Personal Master File, and to a postal code history from a variety of sources. This is the third in a set of CanCHECs that, when combined, make it possible to examine mortality trends and environmental exposures by socioeconomic characteristics over three census cycles and 21 years of census, tax, and mortality data. This report describes linkage methodologies, validation and bias assessment, and the characteristics of the 1996 CanCHEC. Representativeness of the 1996 CanCHEC relative to the adult population of Canada is also assessed.

Keywords: Age-standardized mortality rate, census, cohort, data linkage, mortality, survival

1 Introduction

Since 2008, a number of population censuses have been linked to administrative health data (Bushnik et al. 2016; Rotermann et al. 2015; Withrow et al. 2017; Tjepkema, Wilkins and Long 2013) and to financial data (Jeon and Pohl 2016). These linked datasets have been instrumental in examining health inequalities and have been used in environmental health research (Crouse et al. 2012; Weichenthal et al. 2016).

In 2008, the 1991 Census was linked to 10 years of death records to study mortality by socioeconomic characteristics (Wilkins et al. 2008). In 2012, the dataset was enhanced by the addition of information about place of residence and cancer outcomes (Peters et al. 2013), and was named the Canadian Census Health and Environment Cohort (CanCHEC). This dataset was particularly useful for environmental health research because of its long follow-up period (20 years), large size (2.6 million people), approximate representativeness (it was census-based), broad geographic coverage, and information about where people lived over a 30-year period (via postal codes from 1981 to 2011). The same linkage methodology was applied to the 2001 Census to create the 2001 CanCHEC (Pinault et al. 2016).

With similar methodology, the 1996 CanCHEC was created by linking the 1996 Census (long form) to mortality and tax files, which makes it possible to examine mortality by ethnocultural and socioeconomic factors (as measured on the 1996 Census) and associations between environmental exposures and health. In conjunction with the 1991 and 2001 CanCHECs, temporal trends in mortality can be traced by socioeconomic characteristics. Environmental hazards (for example, air pollution) and neighbourhood characteristics (for example, green space) can be integrated into estimates of exposure because of the inclusion of annual residential location.

This report describes the deterministic and probabilistic linkage methodologies used to create the 1996 CanCHEC as well as the validation methods and bias assessment. It also describes the 1996 CanCHEC (or the Cohort) by demographic and socioeconomic factors, and mortality rates. Approval for the record linkage was provided by the Executive Management Board at Statistics Canada in 2016 (Record Linkage no. 047-2016) (Statistics Canada n.d.a).

2 Data

2.1 1996 Census of Population

The 1996 Census of Population was conducted on May 14, 1996. Information about residents of enumerated households was collected with two questionnaires: the short-form questionnaire (80% of households) and long-form questionnaire (typically, 20% of households, except on Indian reserves and remote/northern areas, where 100% of households receive the long-form questionnaire). The short-form questionnaire gathered basic information: name, relationship to “person 1” (head of household), date of birth, sex, legal marital status, common-law status, and first language learned in childhood. The long-form questionnaire contained 55 questions, including the 7 listed above, on topics such as labour force activity, income, education, activity limitations, citizenship, housing, and ethnic origin.

Names were not used for analytical or linkage purposes, as digitization of the 1996 Census questionnaires did not include names of respondents. To perform the linkage with names would require a time- and labour-intensive review of the census questionnaires, which have been preserved in microfiche format. However, names of respondents were used to manually validate a sample of links and compute a rate of false positives (Subsection 3.1).

The census questionnaire was sent to approximately 98% of households; in the other 2% (typically in remote and northern areas, on most Indian reserves, and in special core areas of major cities), the questionnaire was completed by interview at the household. Some remote northern areas had been enumerated before Census Day, in February and March 1996.

The estimated population undercoverage rate (people not enumerated) of the 1996 Census was 3.18% (Statistics Canada 1999a). Undercoverage was higher for males (3.89%, versus 2.49% for females). Higher undercoverage rates by age were among 20- to 24-year-olds (men 9.48%; women 6.45%) and 25- to 34-year-olds (men 7.74%; women 3.94%). A total of 78 Indian reserves or Indian settlements were incompletely enumerated and not included in 1996 Census products, and so were not available for linkage.

2.2 Derived Record Depository

The Derived Record Depository (DRD) is a national relational database created by linking selected Statistics Canada source index files (for example, vital statistics registration records for births and deaths, tax records, and immigrant data) to produce a list of unique individuals. Each of these files linked only once to the DRD, and each individual in the DRD is assigned a unique Social Data Linkage Environment (SDLE) identifier (Statistics Canada n.d.b; St-Jean 2016). For the 1996 CanCHEC, version 4 of the DRD was used. To ensure consistent methodology with the 1991 and 2001 CanCHECs, a tax file was built from individual variable-specific tables prepared for the SDLE.

Historical residential postal codes were extracted from the DRD. These postal codes come from Canada Revenue Agency sources, such as the T1 Personal Master File, Canada Child Tax Benefit files, and dependant registries from 1981 to 2012. The postal codes reported on the T1 Personal Master File reflect the mailing address for that tax year, and are assumed to be broadly representative of where a person resides.

Mortality records that were previously linked to the DRD were deterministically linked to the Cohort using the DRD identifier. Deaths captured in the Canadian Mortality Database (CMDB) (1970 to 2012) were probabilistically linked to the DRD using standard linkage methodology (St-Jean 2016). The linkage rate was estimated at 95% to 98% for deaths that occurred between 1996 and 2012. Deaths reported in the CMDB included underlying cause of death and were coded using the version of the International Classification of Diseases (ICD) in effect at the time of death. Deaths that occurred from 1996 to 1999 were coded according to the Ninth Revision of the ICD (ICD-9); deaths that occurred from 2000 to 2012 were coded according to the Tenth Revision (ICD-10). Deaths reported via tax filings did not include data on cause of death. For comparability in this paper, deaths previously coded by ICD-9 and ICD-10 were grouped according to the Global Burden of Disease coding (Mathers, Lòpez and Murray 2006).

3 Methods

3.1 Linkage of 1996 Census to the Derived Record Depository

Persons enumerated by the 1996 Census (respondents to both the long-form and short-form questionnaires, referred to herein as “respondents”) were linked to the DRD using deterministic and probabilistic methodology.

Data on sex and birth date were obtained from the DRD version 4. Records were retained if they had a birth date before Census Day, and either a death date after Census Day or no reported death date. Records missing a birth date or postal code were dropped. If sex was missing, two alternate records with options of both male and female were generated in place of the original record. Postal code history from the DRD address table was linked to the Cohort with a deterministic linkage.

In the first of four sequential groups of the deterministic linkage process, the keys used to link the census and the DRD were sex, date of birth, 1996 postal code, and marital status (names were not electronically available for the 1996 Census). The second group was identical to the first, except that it did not consider marital status. The third group used sex, date of birth, and 1995 postal code. Finally, the fourth group used sex, date of birth, and the closest available postal code before 1995 or after 1996 in cases where the 1995 and 1996 postal codes were both missing.

In each group, only records with unique keys on both the DRD and census files were retained in the Cohort; duplicates and non-matches were considered in subsequent linkage groups. Respondents to the census short-form questionnaire were included at this stage to reduce the number of records to be considered for probabilistic linkage and to reduce the number of false-positive matches, because records from the DRD linked to records of the short-form questionnaire at the deterministic step were not eligible for linkage in the subsequent probabilistic step.

Respondents to the census long-form questionnaire who were not matched during the deterministic linkage and who did not reside in an institutional collective dwelling were considered for probabilistic linkage. Several groups were formed, using linkage weights derived from a probabilistic linkage methodology influenced by the Fellegi-Sunter theory of record linkage (Fellegi and Sunter 1969). The following variables were used for probabilistic linkage: birth date, spousal birth date (if applicable), sex, marital status, postal code(s), and rural/urban status based on postal code. A total of six probabilistic linkage groups were considered for inclusion in the Cohort.

To determine the linkage threshold and the groups that would be included in the Cohort, linkage accuracy was estimated for all potential linkage groups by verifying a random sample of linked records against the original scanned census questionnaires, which contained respondents’ names. Respondents for whom the name and birth dates were a match (allowing for minor spelling differences) to those on tax files were considered a successful match. Groups for which 90% of records were not successfully matched were excluded from the Cohort. Five out of the six groups were included in the final probabilistic linkage.

To be considered in scope, individuals had to meet the following criteria: they were enumerated by the 1996 Census long form; they were aged 19 or older on Census Day (May 14, 1996); they were enumerated as residents of Canada (overseas nationals were excluded); they were not institutional residents; and they had filed an income tax return.

3.2 Re-weighting and bootstrap weights

Sampling weights were created for the 1996 CanCHEC to (1) ensure that the Cohort represents the target population (non-institutional population aged 19 or older in 1996); and (2) reduce bias (for example, due to missed links between census and DRD). Weights for the census long-form questionnaire were adjusted by model parameters on the probability of linking to the DRD. These sampling weights were then calibrated according to the raking method (Kalton and Flores-Cervantes 2003) and incorporated trimming to prevent negative or excessively large weights (Izrael, Battaglia and Frankel 2009).

Generalized bootstrap weights were derived from the calibrated sampling weights using the Poisson bootstrap method proposed by Beaumont and Patak (2012). The same raking and trimming macro that was used to calibrate the sampling weights (after adjustment for missed links) was applied to each of the 500 bootstrap replicates. These weights were finally adjusted to take into account the high sampling fraction of the census, using a method proposed by Beaumont and Charest (2012).

3.3 Mortality analysis

Mortality was tracked for 16.6 years from May 14, 1996 (Census Day), to December 31, 2012. Based on records of death, person-years-at-risk (PYAR) were calculated for each respondent. Respondents who lived to the end of follow-up were assigned a PYAR of 16.6 years; those who died during follow-up were given a PYAR reflective of the length of time between Census Day and their death. PYAR was calculated starting at the beginning of the day on Census Day; that is, if someone died on May 14, 1996, their PYAR was equivalent to one day (1/365.25 = 0.00273 years at risk). To examine the comparability of the Cohort to the Canadian population, the years that each Cohort member survived during follow-up were compared with survival rates in the Canadian population over the same period.

Age-standardized mortality rates (ASMRs) and 95% confidence intervals were calculated using the PYAR value and the PROC RATIO procedure in SUDAAN version 11.0.1 (Research Triangle Institute 2013). The procedure used the balanced repeated replication design with calibrated sampling weights and calibrated bootstrap weights. ASMRs were calculated for population-standardized five-year age groups at Cohort inception (Table 1) for a variety of demographic and socioeconomic characteristics. Rate ratios (RRs) and 95% confidence intervals were calculated using ASMR standard errors that were calculated from PROC RATIO and the methodology of Carriere and Roos (1997).

3.4 Analytical file

The analytical file contains nearly all the variables in the census long-form questionnaire, including marital status, household composition, languages, labour force activity, income, education, physical activity limitations, housing, Aboriginal identity, immigration, and ethnic origin. However, variables that were the result of write-in responses were removed. Details about the census variables are available elsewhere (Statistics Canada 1999b).

Variables were derived for the analytical file to simplify categorization of key concepts and comparisons between cohorts. For example, highest level of education was grouped into four categories: less than secondary graduation, secondary graduation, postsecondary diploma, and university degree. Documentation for the derived variables is available in the 1996 CanCHEC Technical Report (Christidis and Tjepkema 2017).

One of the derived variables ranked the population by income adequacy quintiles and deciles. For each economic family or unattached individual, total pre-tax, post-transfer income from all sources was pooled across all family members, and the ratio of total income to the Statistics Canada low-income cut-off (LICO) for the applicable family size and community size group was calculated (Statistics Canada 1999b). Thus, all members of an economic family were assigned the same LICO ratio, which was calculated for all non-institutionalized persons (the in-scope population), including people living on Indian reserves. This population was then ranked according to the LICO ratio, and quintiles and deciles were constructed, nationally and within each census metropolitan area or census agglomeration, or rural and small town area. The reason for constructing the quantiles within each area was to account for regional differences in housing costs, which are not reflected in the LICO, and to base comparisons across areas on comparable percentages of the population in each quantile.

4 Results

4.1 Cohort

The Cohort was formed in several steps (Figure 1), first by combining the deterministic links of respondents to the census long-form questionnaire who did not reside in institutional collective dwellings with the probabilistic links. Of the original 4,706,075 respondents to the 1996 Census long-form questionnaire, 4,389,835 (93.3%) were “in-scope.” The number linked to tax records was 3,566,775 (81.3% of the in-scope population); 99.5% of records were estimated to be true links (less than 0.5% false positive error rate). The majority of these Cohort members (93.5%) were linked via deterministic linkage, for which 99.5% of records were considered to be true links (based on a manual review of 266 records). Fewer Cohort members (6.5%) were linked through the probabilistic linkage, for which 99.4% of links were considered to be true links (based on a manual review of 450 records). After mortality records were linked to the Cohort, 90 respondents were removed because of ambiguous death records, resulting in a final Cohort of 3,566,685 census respondents.

Description for Figure 1

The title of Figure 1 is “Flow chart of the creation of the in-scope census population and the final 1996 Canadian Census Health and Environment Cohort, showing respondent exclusions.”

This diagram presents the number of respondents to the 1996 Census long-form questionnaire that were considered in-scope in the census (n=4,706,075), followed by the number of non-institutional residents (n=4,394,850) from which the number of institutional residents (n=311,230) was excluded. From the remaining respondents, those who did not live in Canada were excluded (n=5,015), resulting in a group that is defined as the eligible in-scope census population (n=4,389,835). From this eligible in-scope census population, the population that was not linked to income tax returns was excluded (n=823,055), leaving a population that linked to income tax returns (n=3,566,775). From this population, ambiguous death records were excluded (n=90). The resulting population is the final 1996 Canadian Census Health and Environment Cohort (n=3,566,685).

The sources of the figure are “Statistics Canada, Derived Record Depository and authors’ calculations based on data from the 1996 Census of Population.”

Overall, 81.3% of the in-scope population was linked to the DRD—3,566,685 non-institutional individuals aged 19 or older (Tables 2 and 3). Linkage rates were below 70% among 19- to 24-year-olds (69% women, 67% men) and women aged 85 or older (66%). Men and women who had moved in the last year had low linkage rates (62%, 66%), as did men who were not married or in a common-law relationship (66%).

4.2 Weighting

The characteristics of the weighted in-scope respondents (using original census sampling weights) and weighted Cohort respondents (using calibrated weights for missed links) were compared to assess how well the Cohort sampling weights reflect the target population and reduce bias (Table 4). Proportions were estimated by selected characteristics and compared between the Cohort and the original census (using a ratio measure). Ratios close to 1.00 indicate that the Cohort sampling weights adequately addressed differences in the in-scope and Cohort groups. After re-weighting, the Cohort slightly underrepresented unmarried men and women (0.98, 0.99), men in urban areas (0.99), men who moved in the year before the census (0.98), and unemployed women (0.99). The Cohort slightly overrepresented men and women who were married or in a common-law relationship (1.01, 1.01), men and women in rural communities (1.02, 1.01), and women who had moved in the year before the census (1.01).

4.3 Mortality

Of the 3,566,685 Cohort members, 565,615 (15.9%) died during the 16.6-year follow-up. The annual number of deaths rose each calendar year, reaching 40,640 in 2012 (Table 5). About 98% of deaths were ascertained by linkage to the CMDB, which indicates cause of death. The remaining 2% were ascertained from the T1 tax file, which provides only the date of death. In general, the percentages of deaths that were identified from tax files did not vary greatly by socioeconomic and demographic characteristics. The exceptions were visible minority and immigrant Cohort members, among whom 7.2% and 3.5% of deaths, respectively, were ascertained from tax files (data not shown).

Chart 1 shows the proportion of the Cohort who survived during the 16.6-year period (from 1996 to 2012), by age and sex, along with composite life tables (1995–1997 to 2010–2012). The trend in the proportion who survived was similar in each data source. At older ages (80 or older), the proportion surviving was slightly greater in the Cohort than in the life tables, perhaps the result of excluding the institutional population from the Cohort.

Data table for Chart 1

Chart 1
Proportion of respondents surviving the 16.6-year follow-up period (1996 to 2012), by age and sex, compared with Canada life tables
Table summary
This table displays the results of Proportion of respondents surviving the 16.6-year follow-up period (1996 to 2012). The information is grouped by Age at baseline (appearing as row headers), Composite life tables (males), 1996 CanCHEC (males), Composite life tables (females) and 1996 CanCHEC (females), calculated using proportion units of measure (appearing as column headers).

Age at baseline	Composite life tables (males)	1996 CanCHEC (males)	Composite life tables (females)	1996 CanCHEC (females)
	proportion
19	0.985653	0.985	0.993895	0.993
20	0.985320	0.986	0.993603	0.993
21	0.984973	0.986	0.993256	0.993
22	0.984587	0.985	0.992846	0.992
23	0.984130	0.984	0.992370	0.992
24	0.983575	0.985	0.991819	0.992
25	0.982887	0.982	0.991186	0.991
26	0.982055	0.982	0.990468	0.989
27	0.981072	0.983	0.989658	0.990
28	0.979942	0.981	0.988757	0.989
29	0.978659	0.980	0.987758	0.988
30	0.977215	0.978	0.986664	0.986
31	0.975599	0.977	0.985462	0.987
32	0.973800	0.977	0.984157	0.984
33	0.971794	0.974	0.982731	0.984
34	0.969555	0.973	0.981178	0.981
35	0.967052	0.971	0.979482	0.979
36	0.964259	0.967	0.977635	0.978
37	0.961147	0.965	0.975617	0.976
38	0.957686	0.961	0.973412	0.974
39	0.953845	0.956	0.971001	0.972
40	0.949590	0.953	0.968364	0.969
41	0.944882	0.948	0.965475	0.966
42	0.939680	0.946	0.962310	0.961
43	0.933943	0.937	0.958842	0.959
44	0.927620	0.932	0.955037	0.954
45	0.920667	0.927	0.950863	0.950
46	0.913030	0.916	0.946274	0.947
47	0.904657	0.913	0.941240	0.940
48	0.895487	0.905	0.935701	0.938
49	0.885460	0.895	0.929618	0.931
50	0.874515	0.885	0.922924	0.923
51	0.862593	0.870	0.915566	0.914
52	0.849620	0.859	0.907474	0.907
53	0.835527	0.846	0.898577	0.893
54	0.820238	0.824	0.888798	0.889
55	0.803684	0.808	0.878054	0.876
56	0.785785	0.795	0.866253	0.865
57	0.766478	0.778	0.853304	0.849
58	0.745692	0.756	0.839106	0.838
59	0.723373	0.740	0.823553	0.826
60	0.699477	0.705	0.806537	0.809
61	0.673968	0.685	0.787948	0.793
62	0.646828	0.656	0.767676	0.777
63	0.618061	0.618	0.745613	0.750
64	0.587700	0.590	0.721659	0.731
65	0.555805	0.559	0.695719	0.708
66	0.522470	0.519	0.667721	0.674
67	0.487832	0.492	0.637613	0.643
68	0.452071	0.448	0.605377	0.613
69	0.415417	0.411	0.571029	0.577
70	0.378151	0.371	0.534644	0.536
71	0.340601	0.332	0.496353	0.509
72	0.303149	0.296	0.456367	0.462
73	0.266220	0.253	0.414974	0.425
74	0.230278	0.227	0.372564	0.381
75	0.195834	0.194	0.329669	0.343
76	0.163486	0.153	0.287059	0.295
77	0.133817	0.120	0.245631	0.258
78	0.107271	0.105	0.206263	0.226
79	0.084166	0.078	0.169726	0.181
80	0.064642	0.061	0.136671	0.146
81	0.048544	0.049	0.107604	0.121
82	0.035627	0.046	0.082734	0.088
83	0.025550	0.033	0.062060	0.087
84	0.017902	0.021	0.045380	0.063
85	0.012264	0.022	0.032332	0.056
86	0.008221	0.016	0.022441	0.040
87	0.005398	0.010	0.015178	0.032
88	0.003475	0.014	0.010010	0.023
89	0.002196	0.009	0.006445	0.022
90	0.001373	0.012	0.004060	0.016
Note: The 1996 Canadian Census Health and Environment Cohort (CanCHEC) is derived from the 1996 Census of Population and the Derived Record Depository. Sources: Statistics Canada, 1996 Canadian Census Health and Environment Cohort, and authors' calculations based on data from Canada life tables.

Detailed mortality statistics (ASMRs, RRs) were calculated for the population by various socioeconomic and demographic characteristics (Tables 6 and 7). Mortality rates were higher among people who were not in the labour force, compared with those who were employed. A clear gradient by educational attainment was apparent, with the highest ASMRs among people with less than secondary graduation. A gradient was also evident by income adequacy quintile, with ASMRs highest among the lowest quintile. Mortality rates also varied by community size, mobility, Aboriginal identity, and visibility minority status. RRs reflect the relative differences between the ASMRs in each subgroup, with one group (reference group) held at 1.00 and other RRs calculated as a ratio relative to the reference group.

Table 8 shows the number of Cohort members who died, by cause. Most deaths were the result of non-communicable diseases (90% of male deaths, 80% of female deaths), specifically, neoplasms (34% of male deaths, 29% of female deaths) and cardiovascular diseases (33% of male deaths, 28% of female deaths). Within these groups, the most frequently reported causes were ischemic heart disease (20% of male deaths, 14% of female deaths) and cancer of the trachea, bronchus or lung (10% of male deaths, 7% of female deaths).

5 Discussion

The 1996 CanCHEC is a retrospective cohort that followed a sample of nearly one-fifth of Canadians 19 years of age or older to determine mortality over a follow-up period of 16.6 years. With this, the third CanCHEC, it is now possible to examine mortality trends spanning three census cycles and 21 years of data.

The 1996 CanCHEC was constructed by linking 81% of respondents to the census long-form questionnaire to the DRD through probabilistic and deterministic linkages. The false positive error rate was less than 0.5%, indicating that most links were true links. This is comparable to the 2001 CanCHEC, which linked 78.6% of in-scope census respondents to tax files and followed them for mortality (3,537,520 of 4,500,245), and to the original linkage results for the 1991 Census Cohort, where 80.0% of in-scope census respondents were linked to tax files and followed for mortality (2,860,240 of 3,576,485) (Pinault et al. 2016; Peters et al. 2013).

Linkage rates of the 1996 Cohort to the DRD differed by census characteristics. Rates were relatively low for respondents who were in the youngest or oldest age groups, not married, or of low income; who had moved in the past year; and who reported Aboriginal identity. These findings are consistent with those of the 1991 and 2001 Cohorts. The 2001 Cohort had lower linkage rates for respondents who were younger, Aboriginal, and movers in the previous year. The 1991 Cohort had lower linkage rates for respondents who were unmarried, were not working, had low income, or reported Aboriginal ancestry. In weighted analysis, the linkage bias was addressed and representativeness of the Cohort was improved.

The survival curve for the Cohort indicated strong concordance with national life tables. The curves diverged slightly at older ages (particularly at age 80) and among women, likely because of smaller sample sizes in these older age groups, and the exclusion of institutionalized residents from the Cohort. Higher survival rates for older women, compared with life table estimates, was also reported for the 1991 and 2001 Cohorts.

Mortality patterns were broadly consistent with the previous CanCHECs—ASMRs were higher among people who were not in the labour force, in the lowest income quintile, or reported Aboriginal identity (or ancestry in the 1991 Census).

6 Conclusion

This paper describes linkage of respondents to the 1996 Census long-form questionnaire to tax and mortality records to create the 1996 Canadian Census Health and Environment Cohort (CanCHEC), an analytical dataset that can be used to examine mortality trends by demographic and socioeconomic characteristics and for environmental health research. This is the third in a set of similar cohorts, which offers the potential to study trends over three census cycles and 21 years of mortality follow-up. Validation revealed a slight bias which increased the probability of respondents to be included in the 1996 CanCHEC and linked to mortality databases, relative to the Canadian population. However, cohort weights were created to allow researchers to reduce the effect of this bias.

References

Beaumont, J.F., and A.S. Charest. 2012. “Bootstrap variance estimation with survey data when estimating model parameters.” Computational Statistics & Data Analysis 56 (12): 4450–4461.

Beaumont, J.F., and Z. Patak. 2012. “On the generalized bootstrap for sample surveys with special attention to poisson sampling.” International Statistical Review 80 (1): 127–148.

Bushnik, T., S. Yang, M.S. Kramer, J.S. Kaufman, A.J. Sheppard and R. Wilkins. 2016. “The 2006 Canadian Birth-Census Cohort.” Health Reports 27 (1): 11–19. Statistics Canada Catalogue no. 82-003-X.

Carriere, K.C., and L.L. Roos. 1997. “A method of comparison for standardized rates of low-incidence events.” Medical Care 35 (1): 57–69.

Christidis, T., and M. Tjepkema. 2017. 1996 Canadian Census Health and Environment Cohort (CanCHEC): Technical Report. Ottawa: Statistics Canada. Unpublished.

Crouse, D.L., P.A. Peters, A. van Donkelaar, M.S. Goldberg, P.J. Villeneuve, O. Brion, S. Khan, D.O. Atari, M. Jerrett, C.A. Pope, III, M. Brauer, J.R. Brook, R.V. Martin, D. Stieb, and T. Burnett. 2012. “Risk of nonaccidental and cardiovascular mortality in relation to long-term exposure to low concentrations of fine particulate matter: A Canadian national-level cohort study.” Environmental Health Perspectives 120 (5): 708–714.

Fellegi, I.P., and A.B. Sunter. 1969. “A theory for record linkage.” Journal of the American Statistical Association 64 (328): 1183–1210.

Jeon, S.H., and R.V. Pohl. 2016. Health and Work in the Family: Evidence from Spouses’ Cancer Diagnoses. Analytical Studies Branch Research Paper Series, no. 381. Statistics Canada Catalogue no. 11F0019M. Ottawa: Statistics Canada.

Izrael, D., M.P. Battaglia, and M.R. Frankel. Extreme survey weight adjustment as a component of sample balancing (a.k.a. raking). Paper presented at the SAS Global Forum, Washington, D.C., 2009.

Kalton, G., and I. Flores-Cervantes. 2003. “Weighting methods.” Journal of Official Statistics 19 (2): 81–97.

Mathers C.D., A.D. López, and C.J. Murray. 2006. “The burden of disease and mortality by conditions: Data, methods and results for 2001.” In Global Burden of Disease and Risk Factors, ed. A.D. Lopez, C.D. Mathers, M. Ezzati, D.T. Jamison, and C.J.L. Murray, Chapter 3, Annex table 3A.2. New York: Oxford University Press. Available at: https://www.ncbi.nlm.nih.gov/books/NBK11808/ (accessed November 9, 2017).

Peters, P.A., M. Tjepkema, R. Wilkins, P. Finès, D.L. Crouse, P.C.W. Chan, and R.T. Burnett. 2013. “Data resource profile: 1991 Canadian census cohort.” International Journal of Epidemiology 42(5): 1319–1326.

Pinault, L., P. Finès, M. Tjepkema, F. Labrecque-Synnott, and A. Saidi. 2016. The 2001 Canadian Census–Tax–Mortality Cohort: A 10-Year Follow-up. Analytical Studies: Methods and References, no. 003. Statistics Canada Catalogue no. 11-633-X. Ottawa: Statistics Canada.

Research Triangle Institute. 2013. SUDAAN (software), Release 11.0.1. Research Triangle Park, North Carolina: Research Triangle Institute.

Rotermann, M., C. Sanmartin, R. Trudeau, and H. St-Jean. 2015. “Linking 2006 Census and hospital data in Canada.” Health Reports 26 (10): 10–20. Statistics Canada Catalogue no. 82-003-X.

St-Jean, H. 2016. Social Data Linkage Environment (SDLE) Methodology Report: External linkage between the Canadian Mortality Database (CMDB 1970 to 2011 and CMDB 2012) and the SDLE Derived Record Depository (version 4). Ottawa: Statistics Canada. Unpublished.

Statistics Canada. n.d.a. “2016 submissions,” “Approved microdata linkages,” Microdata linkage at Statistics Canada. Last updated June 14, 2017. Available at:
http://www.statcan.gc.ca/eng/record/2016 (accessed November 2, 2017).

Statistics Canada. n.d.b. Social Data Linkage Environment (SDLE). Last updated July 7, 2017. Available at: http://www.statcan.gc.ca/eng/sdle/index (accessed October 19, 2017).

Statistics Canada. 1999a. Coverage 1996 Census Technical Reports. Statistics Canada Catalogue No. 92-370-XIE. Ottawa: Statistics Canada.

Statistics Canada 1999b. 1996 Census Dictionary Final Edition Reference. Statistics Canada Catalogue No. 92-351-UIE. Ottawa: Statistics Canada.

Tjepkema, M., R. Wilkins, and A. Long. 2013. “Socio-economic inequalities in cause-specific mortality: A 16-year follow-up study.” Canadian Journal of Public Health 104 (7): 472–478.

Weichenthal, S., D.L. Crouse, L., Pinault, K. Godri-Pollitt, E.Lavigne, G. Evans, A. van Donkelaar, R.V. Martin, and R.T. Burnett. 2016. “Oxidative burden of fine particulate air pollution and risk of cause-specific mortality in the Canadian Census Health and Environment Cohort (CanCHEC).” Environmental Research 146: 92–99.

Wilkins, R., M. Tjepkema, C. Mustard, and R. Choinière. 2008. “The Canadian census mortality follow-up study, 1991 through 2001.” Health Reports 19 (3): 25–43. Statistics Canada Catalogue no. 82-003-X.

Withrow, D.R., J.D. Pole, E.D. Nishri, M. Tjepkema, and L.D. Marrett. 2017. “Cancer survival disparities between First Nations and non-Aboriginal adults in Canada: Follow-up of the 1991 Census Mortality Cohort.” Cancer Epidemiology, Biomarkers & Prevention 26 (1): 145–151. DOI: 10.1158/1055-9965.EPI-16-0706.