Coverage Technical Report, Census of Population, 2021
6. Dwelling Classification Survey

6.1 Introduction

The Dwelling Classification Survey (DCS) measures dwelling classification errors in the census. The DCS also estimates the number of usual residents in occupied non-response dwellings. Using these results, the occupancy status of non-response dwellings is adjusted, and people are imputed into the occupied dwellings during the whole household imputation (WHI) procedure, based on DCS control total files. In this chapter, the term “dwelling” will always refer to private dwellings.^{Note 1}

One of the potential sources of error in a census is the misclassification of a dwelling. When a household does not return a questionnaire, the enumerator has to determine whether the dwelling is occupied. Two types of dwelling classification errors can occur in this situation. First, an occupied dwelling may be misclassified as unoccupied. If these dwellings were ultimately considered to be unoccupied, this would result in census population undercoverage because the dwelling occupants would not be counted. Second, an unoccupied dwelling may be misclassified as occupied. If these dwellings were ultimately considered to be occupied, this would result in census population overcoverage because people would be counted when, in fact, no one lives at that dwelling. Estimates from the DCS are used to adjust census data for these two coverage errors.

Additionally, a third type of dwelling classification error measured by the DCS is the error incurred when marginal dwellings or dwellings under construction are misclassified as dwellings. This will result in dwelling overcoverage and could result in population overcoverage if the dwelling is classified as occupied. Census data are not adjusted for these dwellings, so census estimates of the housing stock include some degree of overcoverage. This is discussed further in Section 6.3.1.2 and Section 6.3.2.2.

6.2 Methodology

6.2.1 Stratification and sample selection

The DCS target population included all private dwellings classified as either unoccupied or non-response dwellings, excluding dwellings in First Nations communities, Métis Settlements, Inuit regions and other remote areas, as well as private dwellings attached to a collective dwelling.^{Note 2} Those areas were excluded because of cost and operational considerations.

The DCS sample size was set at 1,903 collection units (CUs). The sampling frame consisted of all non-remote and non-reserve CUs (i.e., CUs with a collection methodology of mail-out, mail-out with drop-off [MODO], list/leave and seasonal).^{Note 3} Consequently, Nunavut had no in-scope CUs, and the DCS was not conducted there. The sample design was as follows. All in‑scope CUs in Yukon (42 CUs) and the Northwest Territories (21 CUs) formed two strata, one for each territory. All their in‑scope CUs were selected for the DCS sample with certainty. All the CUs in Prince Edward Island formed a third stratum, from which a simple random sample of 51 CUs was selected.

The remaining CUs in each province were grouped into urban and rural strata. A CU was initially considered urban if it had been part of a census metropolitan area (CMA) or a census agglomeration (CA) that had 40,000 or more dwellings. Further, all the CUs within a crew leader district (CLD) were considered urban if more than 50% of the CUs in the CLD were urban. All the remaining CUs formed the rural strata. Urban CUs were stratified by CMA and CA. A simple random sample of at least five CUs was selected within each stratum. From past census data, it was determined that five CUs was an appropriate workload for an interviewer. There were 1,037 urban CUs in the sample. To control field costs, CUs in close proximity to each other were chosen for the rural sample. This was done via a two-stage stratified random sampling design. In the first stage, CLDs were selected within each province using simple random sampling. In the second stage, five CUs were randomly chosen from each of the selected CLDs. There were 866 rural CUs in the sample.

Dwelling subsampling within a sampled CU occurred when the number of unoccupied and non-response dwellings exceeded a maximum dwelling parameter, which was 35 in mail-out and list/leave CUs, 40 in MODO CUs, and 135 in seasonal CUs. Subsampling of in-scope dwellings occurred in 584 CUs. Otherwise, all unoccupied dwellings and non-response dwellings^{Note 4} in the sampled CUs formed the DCS dwelling sample. A total of 34,459 unoccupied dwellings and 11,346 non-response dwellings were sampled in 2021. Table 6.2.1 shows the distribution of the sample by province and territory.

6.2.2 Field interviews

A DCS questionnaire was used to verify the true occupancy status on Census Day of sampled dwellings in the sampled CUs. Occupancy status was verified from mid-June to mid-August 2021. DCS operations ideally started once a predetermined response rate threshold for the Census of Population was obtained in the sampled CU, but sometimes they had to start earlier because of operational constraints.

To determine occupancy status and to collect other information, enumerators were instructed to contact current occupants, neighbours, landlords or any other person with knowledge about the dwelling. Up to three contact attempts were made for each dwelling. If the dwelling was found to have been occupied on Census Day, the number of occupants on Census Day was also obtained, along with occupants’ sex and age, when possible.

6.2.3 Processing and estimation

All completed questionnaires were sent to head office in Ottawa for processing.

The first processing step was data capture. Once this was completed, the questionnaires were subjected to an extensive set of consistency edits. The questionnaires that failed the edits were examined manually to resolve inconsistencies.

At this point in the processing, the unoccupied dwellings and the non-response dwellings in the sample were separated, and the dwellings’ classifications were confirmed against the census database. The questionnaires completed for each sampled CU were matched to the final census lists of unoccupied dwellings and non-response dwellings (i.e., no census questionnaire was received for the dwelling). If a match could not be found, meaning that the status of the dwelling had changed after the DCS sample was selected, the sampled dwelling was discarded and no further processing was required. Dwellings listed as unoccupied or as non-response on the census lists in the sampled CUs for which no DCS questionnaire was received were considered as total non-response to the DCS and proceeded to the next processing step.

Total non-response to the DCS was addressed by a weighting adjustment, and item imputation was used for item non-response. The procedure was the same for unoccupied dwellings and non-response dwellings. When there was no information for a dwelling, the design weights of the respondents were adjusted to account for the design weight of the non-respondents. The adjustment was made separately by geographic post-stratum. Refer to Hong (2023) for more details about the post-strata used. Design weights were then adjusted so that the sum of the adjusted weights for each geographic post-stratum equalled the total number of unoccupied or non-response dwellings. Item non-response for occupancy status and the number of usual residents was addressed by imputation. Occupancy status was imputed first and then used in the imputation of the other variables.

Finally, the household size, collected on the DCS questionnaire when a dwelling was found occupied, was used to produce an estimate, by household size, of the occupancy rate of private dwellings classified by the census as unoccupied or as non‑response.

6.2.4 Census whole household imputation

The WHI procedure strives to represent non-response private dwellings in the census and, as such, imputes for total non‑response to the census. The strategy varies based on whether the DCS was conducted in the area. The WHI procedure results in all private dwellings being classified as either occupied or unoccupied in the census database (i.e., there are no longer any non‑response dwellings).

In geographic areas where the DCS was conducted (i.e., CUs with mail-out, MODO, list/leave and seasonal collection types, excluding private dwellings attached to a collective dwelling), the DCS estimates were used as input to the WHI algorithm to impute occupied private dwellings based on post‑strata and household size distributions. First, within a DCS geographic post‑stratum, all the dwellings without a response were identified (this was done separately for the unoccupied and non‑response dwelling universes). Second, for the non‑response dwelling universe only, any non‑response dwelling for which field collection had obtained the number of usual residents was deemed to be occupied and was assigned the recorded household size. Finally, dwellings without a response were randomly selected and imputed as occupied. The selection was made so that the final number of non‑response and unoccupied dwellings converted to occupied dwellings in the post‑strata equalled the DCS estimate of occupied dwellings in the non‑response and unoccupied dwelling universes. In addition, new for the 2021 cycle, control total files were used such that the DCS estimate of the occupancy rate in each post‑stratum was applied in each census subdivision. Once a private unoccupied or non‑response dwelling was imputed as occupied, a procedure, subject to the constraints of DCS estimates by post‑stratum and household dwelling size, was used to impute the household dwelling size. Then, the geographically nearest dwelling of the predetermined household size was chosen as a donor. The answers to the census short‑form questionnaire from this donor household were then assigned to the unoccupied or non‑response dwelling that was imputed as occupied.

In geographic areas where the DCS was not conducted (i.e., CUs with remote and reserve collection types, as well as private dwellings attached to a collective dwelling), the vast majority of private dwellings classified as non‑response were imputed as occupied, with the household size coming either from field collection or from a donor household. The geographically nearest dwelling of the predetermined household size was chosen as a donor. If the dwelling did not have a household size already assigned, the household size was chosen to be the donor’s household size. The information from this donor household was then assigned to the unoccupied or non-response dwelling that was imputed as occupied. For dwellings in remote and reserve collection areas, the complete record was assigned. For private dwellings attached to a collective dwelling, only the census short-form questionnaire answers were assigned.

Several adaptations were implemented in the 2021 Census collection plan to mitigate the impact of the COVID-19 pandemic and other potential risks. A statistical contingency plan was developed based on the secure, responsible and appropriate use of administrative data to support the 2021 Census in the event of disruptions to census collection. Linked administrative data from federal and provincial data sources were used in specific areas to improve the imputation of non-response households at the imputation stage after the 2021 Census collection ended.^{Note 5} Households constructed from administrative data were added to the donor pool in those specific areas.

More information on WHI can be found in the report by Hong (2023).

6.3 Estimates

Census data were adjusted using DCS estimates, via the WHI procedure, for non-response dwellings and for occupied dwellings that were misclassified as unoccupied. The estimates are provided in sections 6.3.1.1 and 6.3.2.1. Census data were not adjusted for marginal dwellings or dwellings under construction that were misclassified as dwellings. Sections 6.3.1.2 and 6.3.2.2 present the estimates of the number of marginal dwellings and dwellings under construction that were wrongly classified as dwellings and were therefore erroneously included in the housing stock.

6.3.1 Unoccupied dwellings

6.3.1.1 Occupied dwellings misclassified as unoccupied

Table 6.3.1.1.1 provides the estimated number of occupied dwellings that were misclassified as unoccupied and the corresponding error rate for unoccupied dwellings by urban and rural areas^{Note 6} and by province and territory. For comparison, Table 6.3.1.1.2 provides the estimates for the same variables for the 2016 Census.

Table 6.3.1.1.1 shows that 17.3% of all dwellings classified as unoccupied in 2021 were truly occupied. This is a slight increase from 15.0% in 2016 and is a statistically significant change. Dwelling misclassification was more prevalent in urban areas (23.8%) than in rural areas (10.9%). Both urban and rural areas show an increase from 2016. The misclassification rate increased or remained nearly unchanged for all provinces and territories except for the Northwest Territories, where it decreased.

Among the provinces and territories, New Brunswick had the highest misclassification rate (20.2%), followed by Ontario and British Columbia (18.9%), and Quebec (17.7%). The rates for the other provinces and territories ranged from 8.1% for Prince Edward Island to 16.6% for Alberta.

Table 6.3.1.1.3 provides the estimated number of occupied dwellings that were misclassified as unoccupied, the estimated number of people living in those dwellings and the standard error. These estimates were used to decide how many dwellings to impute as occupied in WHI. Table 6.3.1.1.4 provides the estimates for the same variables for the 2016 Census. Because of errors in the initial dwelling classification, approximately 218,042 households were not enumerated in the 2021 Census. This is the number of households added to the census during WHI.

6.3.1.2 Housing stock overcoverage in dwellings misclassified as unoccupied

Table 6.3.1.2 shows the estimated number of dwellings classified as unoccupied that were not in the housing stock and the corresponding error rate for unoccupied dwellings for various geographic areas. No adjustments were made to the census database to account for dwellings not in the housing stock that were misclassified as unoccupied.

The enumeration of unoccupied dwellings that fall outside the housing universe results in dwelling overcoverage. Dwellings are outside the housing universe if they are used for commercial purposes, are not habitable year round or are double counted in the census. Double counting can occur when the dwelling appears to have two addresses associated with it or when two questionnaires are mistakenly returned for a dwelling that no longer contains a separate apartment.

The DCS estimates of the number of dwellings not in the housing stock that were misclassified as unoccupied dwellings were not used to adjust the census database because of the degree of subjectivity associated with classifying a dwelling as suitable for year-round occupancy. A dwelling must have a source of heat or power and provide complete shelter from the elements to be considered suitable for year-round occupancy. It is sometimes difficult to tell whether a dwelling is habitable, such as when a dwelling is a cottage, is under construction and almost complete, or has deteriorated. Note that the conditions used to identify dwellings that fall outside the housing universe were updated since the 2016 cycle to better reflect the definition provided in this paragraph. With the updated conditions, dwellings outside the housing stock would have accounted for 27.5% of all dwellings classified as unoccupied in 2016.

This percentage was similar in 2021, where dwellings outside the housing stock accounted for 26.3% of all dwellings classified as unoccupied. Among the provinces and territories, the incidence of dwellings outside the housing stock being classified as unoccupied ranged from 15.1% in Newfoundland and Labrador to 38.0% in New Brunswick. The problem was more prevalent in rural areas (27.7%) than in urban areas (24.8%).

Dwellings actually outside the housing stock represented 2.1% of all private dwellings in the 2021 Census. This was the same as the 2016 error rate (2.1%). Among the provinces and territories, the error ranged from 0.8% in the Northwest Territories to 4.1% in Prince Edward Island.

6.3.2 Non-response dwellings

6.3.2.1 Occupancy rate among non-response dwellings

Table 6.3.2.1.1 provides the estimated number and rate of occupied non-response dwellings in the census by urban and rural area^{Note 7} and by province and territory. Table 6.3.2.1.2 provides this information for the 2016 Census. Table 6.3.2.1.1 shows that 61.5% of all dwellings classified as non-response were occupied. This is a slight decrease from the 63.1% occupancy rate in 2016, though the difference is not statistically significant. The 2021 Census was relatively consistent in classifying non-response dwellings in urban (61.2%) and rural (62.2%) areas. At the provincial and territorial level in 2021, Yukon had the highest occupancy rate for non-response dwellings, at 69.5%, while Newfoundland and Labrador had the lowest rate, at 41.6%.

Table 6.3.2.1.3 shows the number of occupied non-response dwellings in the 2021 Census and provides the number of people added in those dwellings through the DCS. Table 6.3.2.1.4 shows the data from the 2016 Census. In 2021, 421,202 people were added to the census in 210,370 dwellings. The comparable 2016 numbers were 357,666 people in 179,823 dwellings.

6.3.2.2 Housing stock overcoverage in dwellings misclassified as non-response

Table 6.3.2.2 shows the 2021 Census dwelling classification error from dwellings misclassified as non-response when they should not have been included in the housing stock. Section 6.3.1.2 provides the definition of dwellings outside the housing universe and comments on the difficulty in determining whether a dwelling should be included in the housing stock. Note that the conditions used to identify dwellings that fall outside the housing universe were updated since the 2016 cycle to better reflect the definition provided in this paragraph. With the updated conditions, non-response dwellings that were outside the housing stock accounted for 7.0% of non-response dwellings in 2016.

At the national level, dwellings outside the housing stock accounted for 7.2% of all non-response dwellings in 2021. The error rate was higher in urban areas (7.6%) than in rural areas (6.2%). For provinces and territories, the incidence of dwellings outside the housing stock being classified as non-response ranged from 4.5% in Newfoundland and Labrador to 15.4% in Saskatchewan. At the national level, non-response dwellings outside the housing stock accounted for 0.2% of all private dwellings. This was similar to the 2016 error rate of 0.1%. Among the provinces and territories, the error ranged from 0.1% to 0.4% when looking at the error by province or territory.