Methodology of the Canadian Labour Force Survey
Chapter 3 Dwelling frame creation and maintenance
3.0 Introduction
As described in the previous chapter, the Labour Force Survey (LFS) uses a two-stage sample design in all provinces except for Prince Edward Island. An advantage of this approach is that the sample is concentrated in a limited number of areas; therefore, it is possible to conduct personal interviews. At the first stage, primary sampling units (PSUs) – also called clusters – corresponding to geographic areas are selected. These are relatively small parcels of land, often Census Dissemination Areas (DA). Within the selected PSUs, dwellings are selected at the second stage of sampling.
At both stages of the sampling process, a survey frame, i.e., a list of all the units (clusters or dwellings) that are part of the target population, is required. A good quality frame will have limited coverage errors and facilitate contact with the sampled units. Given that new units are continually being added and removed from the target population, it is important that maintenance and updates are performed on the sampling frame. Details of the frame creation for PSUs and the design aspects for the selection of households were described in Chapter 2.
3.1 Dwelling frame creation
Within selected PSUs, a complete list of dwellings (a frame) is required in order to select the second-stage sample. The list is obtained either through a listing exercise performed in the field or from an existing list, specifically the Address Register (AR). Once the dwelling list is available, it will be used as long as the PSU is in sample. One continuing challenge is to determine which newly sampled PSUs should undergo listing and which can rely on the AR information. Field listing is a more costly option that should be avoided whenever possible. It usually occurs when the information on existing lists is of low quality.
3.1.1 The Address Register
The AR is a database that was initially created for the 1991 Canadian Census of Population, with the purpose of improving census coverage. It was created using several administrative files, such as telephone billing files and building permit files. Immediately after that census, the AR was updated using the list of addresses created during the census enumeration process. Since that first iteration, the AR has continued to be updated quarterly using administrative files and the census listing program, and census information available every five years.
The AR was originally designed to provide and maintain a list of addresses for communities with a population over 50,000. The coverage of the AR was expanded following each subsequent census to include smaller population centres and regions outside population centres. Currently, the AR has national coverage, though it is more accurate in population centres.
In 2015, the AR included over 15 million addresses. The vast majority of these addresses – about 90% – were found to be valid residential dwellings during the 2011 Census. Of the remaining addresses, 7% were obtained through updates from administrative files and field listing in preparation for Census 2016 and 3% were valid dwellings during a previous census.
To appear on the AR, a residential dwelling must possess a valid standard civic address or any sort of descriptive address. For survey purposes, the descriptive addresses are often incomplete and may not provide enough information to locate the dwelling. Where there is a substantial proportion of descriptive addresses, the area may have to be listed in the field.
Two key files are extracted from the AR database for the LFS dwelling frame creation process: the Dwelling Universe File and the Residential Telephone File.
Dwelling Universe File
The Dwelling Universe File (DUF) is an extraction of addresses from the AR. Rules are applied to ensure that the list only contains dwellings that correspond to the target population of the LFS. These rules evolve over time as methods to detect spurious or duplicate addresses improve. Collective dwellings are also a small part of the LFS target population and these dwellings are available through the AR extraction process.
Residential Telephone File
The Residential Telephone File (RTF) is a list of residential telephone numbers valid in Canada. Many of them (88% in 2015) can be associated with a dwelling address found on the DUF. The RTF can therefore be used to add telephone numbers – key contact information – to a large portion of sampled dwellings.
3.1.2 The National Geographic Database
To use the AR in a two-stage design context, each address must first be assigned to a specific PSU. This is achieved by linking the AR to the National Geographic Database (NGD). The NGD contains map layers that include PSU boundaries, street networks, waterways, and other geographical markers. This information can be used to link addresses to street sections. These sections can be at the block level (a block is a polygon with street segment sides contained within a DA) or more precisely at the block-face level (a single street segment). These sections are then associated with a DA or PSU which effectively associates a dwelling with the PSU.
The NGD is managed in partnership with Elections Canada and is constantly changing due to the regular addition of roads and geographic boundary updates such as municipal boundaries. Every three months a new vintage of the NGD is released.
3.1.3 Ordering the list of addresses
The addresses on the dwelling frame must be organized into a list with a specific order that can be maintained over time. This ordering helps to facilitate finding selected dwellings and can help interviewers to recognize any list omissions. The ordering of the addresses is created by a sequencing algorithm which lists the block-faces in an order that covers the entire PSU while minimizing the total distance travelled by the interviewer when verifying the list of addresses. This algorithm uses the geographical information within the PSU coming from the NGD and is most helpful to field staff when all addresses can be block-face geocoded. The algorithm is run for the entire frame of PSUs for each vintage of the NGD. This means that in each selected PSU, the list of addresses is put in a specific order to facilitate and optimize listing.
3.2 Loading and Field Listing
Once the dwellings have been assigned to their PSUs, quality indicators for the list of addresses can be developed. The quality determines if the region will require field listing or if the AR-NGD information will suffice as the list to be used as the sampling frame for dwellings in the PSU.
Ideally, the lists in all PSUs would be verified in the field (field listed), but the budget restricts the number of PSUs that can fall in this category. The quality of the list of addresses for a given PSU depends on the quality of the AR, the quality of the NGD, and the effectiveness of the DUF eligibility rules. The goal of this strategy is to make as much use of the AR as possible while at the same time taking into account the fact that its quality varies for different regions.
The AR quality is known to be highest in population centres. These population centres largely correspond to the “mail-out area”, where the census collection method is to reach households by mail. This area corresponds to about 80% of dwellings. Based on this information, PSUs are classified into one of three groups:
- AR Group 0 are PSUs in the mail-out area. No initial listing in the field is performed and the first sample of units is selected from the AR-based list. While not field listed “as an LFS PSU”, a substantial portion of the mail-out area undergoes field listing under the census listing program. The results of that listing are processed by the AR team and ultimately appear on the DUF.
- AR Group 1 are non-mail-out PSUs with no initial listing. The AR list is assessed to be of good quality, based on a collection of statistics and indicators. The initial sample of dwellings is selected directly from the AR-based list.
- AR Group 2 are non-mail-out PSUs with initial listing. The PSU must be field listed before the first sample selection occurs.
The 2015 allocation assigned 72% of the sampled PSUs to AR Group 0, 19% to AR Group 1 and 9% to AR Group 2. This is a major change from the launch of the 2005 design where fully 61% of PSUs required initial listing. As PSUs rotate in and out of the sample and as the quality of the AR evolves (especially after the 2016 Census) the distribution of the AR Groups will likely change.
3.2.1 Initial loading
For AR Group 0 or AR Group 1 PSUs, the dwelling list used in sample selection is populated from the list of dwelling addresses available on the DUF linked to these PSUs. This process is called initial loading. The LFS sample of dwellings is selected directly from this list.
Unlisted PSUs tend to have a higher proportion of sampled units coded “invalid” or “demolished” at the time of survey collection. Coverage errors will be discussed in Chapter 8.
3.2.2 Initial listing
The PSUs in AR Group 2 must undergo initial listing. The goal of initial listing is to prepare a complete and accurate dwelling list for the first sample selection in a PSU. The initial listing case is pre-filled with the dwellings associated with that PSU according to the DUF. Each dwelling in the list is validated, modified or deactivated by field staff. New dwellings can also be added to the list.
PSU mapping
In order to complete the field listing effectively, the PSU boundaries must be displayed on a map. Proper translation of the map contents in relation to physical features on the ground is paramount in determining which dwellings belong to the PSU. Further, the block numbers and address ranges on the map can help pinpoint specific addresses. Dwelling addresses or descriptions are captured by the field interviewer using the Statistics Canada Listing Application. PSU maps are generated using Generalized Mapping System software in place since 2009. Appendix D contains examples of PSU maps and describes more details about their creation and uses.
Listing collectives
The listing of collectives is not as clear-cut as with privately-occupied dwellings. There are two main criteria for listing collectives. First, inmates of institutions are not part of the population covered by the LFS. Likewise, temporary residents with a usual place of residence elsewhere are not eligible. Generally only the owners’ residence, any staff residences, and dwellings for non-institutionalized residents (e.g., units in a seniors’ residence) would be listed.
3.3 Frame maintenance
Regardless of whether or not the PSU underwent initial listing, each month there is an opportunity to update or correct the dwelling list. Therefore, most frame problems are temporary and can be rectified for subsequent sampling occasions.
3.3.1 List update and list maintenance
Once a PSU has been selected, regular updates to the address list can come either quarterly from each new vintage of the DUF (list update) or monthly from field verification (list maintenance). For AR Group 0 clusters, a combination of list update and list maintenance is used. For AR Group 1 and 2 clusters, list maintenance is the main source.
In list maintenance, dwellings can be added, modified or deactivated (with some reason for the deactivation). Dwellings can be moved in the listing order, affecting a print sequence number, but the permanent within cluster ID number, the listing line, remains fixed. This approach allows the interviewer to have a preferred listing order while effectively preserving the sample history of each dwelling.
Maintenance is normally conducted “on rotation”, meaning during the first month of sampling (e.g., January or July birth months for rotation 1 PSUs). Typically, the interviewer must visit the PSU because at least some of the newly selected dwellings need to be contacted in person.
Interviewer Selected Dwellings
List maintenance can trigger Interviewer Selected Dwellings (ISDs). These are new LFS cases for the CAPI interviewer to complete.
There are two forms of ISDs created during list maintenance. First, during the life of the PSU, the interviewer can add new dwellings on a regular basis as the population grows. Since the dwelling list is open-ended, additional dwellings can be selected in the field. The structures added to the end of the list are sampled using the PSU-level inverse sampling rate (ISR) and next-line-to-be-interviewed provided from the latest sample selection in the PSU. Once a dwelling is selected, the next-line-to-be-interviewed is incremented by the PSU ISR.
The second form of ISDs is known as multiples. During the process of interviewing within a selected dwelling, the interviewer may determine that separate dwellings exist within the structure that were not identified in the list. Typically these are basement or upper units not evident from the street. Since the dwelling list does not contain the extra units as separate lines, these dwellings have no probability of being selected over the lifetime of the PSU. To compensate for the missed dwellings in this and any other similar unresolved cases, all missed units are selected to be in sample along with the original dwelling. They are added to the list as multiples of the originally selected dwelling and the application generates a case for each multiple.
3.3.2 Treatment of growth areas
Since PSU dwelling lists are open-ended, there is potential for extreme growth. Interviewers may not be able to maintain large lists because of the cost associated with this maintenance, and the time required to conduct interviews for the large influx of new sample that such a large list implies. Although this extreme growth is observed in less than 1% of PSUs, options must be available to manage and treat it.
PSU sub-sampling
Based on feedback from the field, the PSUs with large growth may hinder the ability of the interviewer to complete all the assigned interviews. This can be even more difficult during a birth assignment, especially if the fraction of households requiring in-person interviews is high. In such cases of isolated growth, the PSU is sub-sampled to reduce the burden. The LFS uses two forms of sub-sampling.
The first is a simple modification to the sampling rate for the specific PSU. This technique – also called cluster or mechanical sub-sampling – is used for the majority of cases. Often, it is sufficient to decrease the sampling rate by a factor of two in order to reduce the interviewer’s workload by half.
The second form of sub-sampling is the insertion of an additional stage of sample selection. In this technique, sub-clusters are formed as second-stage units (SSUs). By convention, PSUs can be referred to as clusters and parts of PSUs as sub-clusters. In cases of large growth, head office staff delineate four or more sub-clusters of approximately equal size in terms of number of households within the PSU. Two of the SSUs are then selected for survey activity and sub-sampling factors are created.
Sub-sampling modifications affect the sampling probability of the households. Descriptions of the adjustments to account for this are found in the explanation of the weights in Chapter 6.
Stratum update
On rare occasions, the growth in a PSU is so extreme it causes a more than tenfold increase in the number of households. In this scenario, PSU sub-sampling may introduce extreme sampling factors or be insufficient to reduce the interviewer’s workload. In addition, the sub-sampling factors can create high variability amongst the sampling probabilities and may affect the precision of estimates. In such cases, it is better to redesign the stratum. Typically, other PSUs in the stratum will also have exhibited significant growth.
For a stratum level redesign, the original PSUs exhibiting extreme growth are re-delineated into several new PSUs having approximately 230 households each, which is the average size of a PSU. Estimated household counts for all PSUs in the stratum are required, whether they are newly formed or retaining their original boundary. These counts can often be derived based on the latest DUF. With these revised inputs, the stratum update program is run to re-form the random rotation groups and re-establish the PSU level sampling fractions. This program, based on Keyfitz (1951), as modified by Drew, Choudhry, and Gray (1978), retains as many of the selected PSUs as possible at the time of the update.
The newly selected PSUs must be field listed or loaded with information from the AR. The new sample is phased-in over six months.
3.3.3 Monitoring PSU yield
Through time, the PSU yield of households is carefully monitored. PSU with exceptionally small or large household yields may need special attention or treatment. A very low household yield suggests a fundamental change since the design counts were established in June 2013. A large household yield usually indicates areas of growth, but may also indicate dwellings shifted into the wrong PSU on the DUF. Field follow-up or head office investigations are done to justify or correct discrepancies.
3.3.4 Sample size stabilization
Over time there is a slow increase in the size of the population. Left unchecked, this growth would increase the sample size and survey collection costs. In order to keep the sample size under control, stabilization can be used.
Unit targets
The first step of stabilization is to determine where stabilization is necessary. Unit targets – the number of sampled units required in a region to obtain the desired sample of households – are determined. The unit targets take into account that some units on the frame will not necessarily be valid dwellings and that some fraction of valid dwellings are not occupied (i.e., not households). Each stabilization area is a collection of strata that roughly corresponds with an Employment Insurance Economic Region (EIER) or some portion of an EIER. Unit targets should function for all rotations and rarely require updating. The results of recent collection generally indicate where adjustments to the unit targets are warranted – either due to an observed household shortfall or surplus.
Stabilization selection
The unit targets are compared with the number of units obtained from sampling from the most up–to-date dwelling lists at the prescribed rates. Regions requiring stabilization are those where the sample obtained from the most recent dwelling lists contains more units than required according to the unit targets. The number of units to drop is the number of units in the initial sample minus the unit target.
Some areas are defined with no expectation to drop units in that area because of the small sample size and high relative variability. Large growth PSUs with sub-sampling are also exempt from stabilization to avoid further inflation to the sub-sampling factors that are already present. From the remaining units, a systematic subsample of units is selected to drop from the collection process. The selection probabilities of the units not dropped are adjusted to ensure a proper representation of the population.
Other surveys that select units from the LFS frame can do their own stabilization, dropping units from their initial sample. These surveys are discussed in Chapter 9.
Stabilization weight adjustment
The stabilization weight, used to compensate for dwellings dropped from the sample, is calculated after the drop is completed. Not all strata in one stabilization area have the same stratum ISR and the calculation of the weight adjustment takes this into account, ensuring that the sampled units properly represent the population.
The following example illustrates how the stabilization factors do not affect the weighted contribution from the entire stabilization area.
Imagine a stabilization area of three strata, A, B, and C with stratum level ISRs of 400, 500 and 600 and pre-stabilization unit yields of 10, 10 and 10 respectively. Further assume the unit target for this stabilization area is 28, meaning two units should be dropped. For this example, it is assumed that one unit was dropped from stratum A and one unit was dropped from stratum B.
The weighted contribution from this stabilization area should be 15,000 = 10x400+10x500+10x600. With the two units dropped, the weighted contribution becomes 9x400+9x500+10x600=14,100. The stabilization factor is such that the weighted contribution from this stabilization area is preserved. In this example, the factor of 15,000/14,100 is applied to the selected units that remain in sample and the contribution of these units to this stabilization area is exactly 15,000 with this adjustment applied to the weights.
Special considerations
Dwellings selected in the field due to growth in the PSU are identified after the stabilization process and therefore have no chance to be included in the stabilization program. In theory, these dwellings should not have a stabilization weight applied. However, our current systems assign the stabilization factors at the stratum level, and any ISDs are subject to the same stabilization factor as other units in the stratum. The impact is minimal as the number of growth ISDs is small and the stabilization factors are close to or exactly 1. Multiples, multi-unit dwellings misidentified as single residencesNote 1, are given the stabilization weight, in effect appropriating the weight of the main residence.
Note
- Date modified: