Table C3

Survey response and weighting

The Canadian IALSS sample has a very complex design, involving stratification, multiple phases, multiple stages, systematic sampling, probability proportional to size sampling, and several overlapping samples. Furthermore, there is a need to compensate for the non-response that occurred at varying levels. Therefore, the estimation of population parameters and the associated standard errors is dependent on the survey weights. Two types of weights were calculated: population weights that are required for the production of population estimates, and jackknife replicate weights that are used to derive the corresponding standard errors.

Population weights

The population weights were derived in four steps: 1) calculation of the design weights, 2) weighting adjustments for non-response, 3) integration of the weights from the different samples, and 4) calibration.

The design weights were defined as the inverse of the probabilities of selection. The overall probability of selection of a sample unit was the product of its probabilities of selection at each phase and stage of selection. The sequential selection of multiple samples in a province was taken into account by factoring in the probability that a unit selected in a given sample was not selected in any of the samples already selected.

The weighting adjustments for non-response were calculated by first categorizing the sample units either as respondents, out-of-scope households, nonrespondent households (those without data from the screener), and non-respondent individuals (screener completed, but no data for the selected respondent). The CHAID algorithm in Knowledge-Seeker software was used successively to form weighting classes (response homogeneous groups) to adjust for non-respondent households and non-responding persons in two separate stages for each province and sample type. Afterward, the design weights of the respondents were adjusted by the factors calculated from each step in order to represent all individuals.

With the overlap in coverage from the various samples, it was necessary to integrate the weights to be able to produce estimates using all units from all samples. The situation is comparable to a multiple frame situation, except that here the samples are dependent. The weights were integrated using Hartley’s method for multiple frames: the entire sample was partitioned according to the sub-populations targeted in the supplementary samples, and the weights were adjusted by coefficients proportional to the realized sample sizes of the various samples within the partition.