Survey design
Results
All (6)
All (6) ((6 results))
- Articles and reports: 12-001-X19980013904Description:
Many economic and agricultural surveys are multi-purpose. It would be convenient if one could stratify the target population of such a survey in a number of different purposes and then combine the samples for enumeration. We explore four different sampling methods that select similar samples across all stratifications thereby reducing the overall sample size. Data from an agriculture survey is used to evaluate the effectiveness of these alternative sampling strategies. We then show how a calibration (i.e., reweighted) estimator can increase statistical efficiency by capturing what is known about the original stratum sizes in the estimation. Raking, which has been suggested in the literature for this purpose, is simply one method of calibration.
Release date: 1998-07-31 - Articles and reports: 12-001-X19980013905Description:
Two-phase sampling designs offer a variety of possibilities for use of auxiliary information. We begin by reviewing the different forms that auxiliary information may take in two-phase surveys. We then set up the procedure by which this information is transformed into calibrated weights, which we use to construct efficient estimators of a population total. The calibration is done in two steps: (i) at the population level; (ii) at the level of the first-phase sample. We go on to show that the resulting calibration estimators are also derivable via regression fitting in two steps. We examine these estimators for a special case of interest, namely, when auxiliary information is available for population subgroups called calibration groups. Postrata are the simplest example of such groups. Estimation for domains of interest and variance estimation are also discussed. These results are illustrated by applying them to two-phase designs at Statistics Canada. The general theory for using auxiliary information in two-phase sampling is being incorporated into Statistics Canada's Generalized Estimation System.
Release date: 1998-07-31 - Articles and reports: 12-001-X19980013910Description:
Let A be a population domain of interest and assume that the elements of A cannot be identified on the sampling frame and the number of elements in A is not known. Further assume that a sample of fixed size (say n) is selected from the entire frame and the resulting domain sample size (say n_A) is random. The problem addressed is the construction of a confidence interval for a domain parameter such as the domain aggregate T_A = \sum_{i \in A} x_i. The usual approach to this problem is to redefine x_i, by setting x_i = 0 if i \notin A. Thus, the construction of a confidence interval for the domain total is recast as the construction of a confidence interval for a population total which can be addressed (at least asymptotically in n) by normal theory. As an alternative, we condition on n_A and construct confidence intervals which have approximately nominal coverage under certain assumptions regarding the domain population. We evaluate the new approach empirically using artificial populations and data from the Bureau of Labor Statistics (BLS) Occupational Compensation Survey.
Release date: 1998-07-31 - Articles and reports: 12-001-X19970023615Description:
This paper demonstrates the utility of a multi-stage survey design that obtains a total count of health facilities and of the potential client population in an area. The design has been used for a state-level survey conducted in mid-1995 in Uttar Pradesh, India. The design involves a multi-stage, areal cluster sample, wherein the primary sampling unit is either an urban block or rural village. All health service delivery points, either self-standing facilities or distribution agents, in or formally assigned to the primary sampling unit are mapped, listed, and selected. A systematic sample of households is selected, and all resident females meeting predetermined eligibility criteria are interviewed. Sample weights for facilities and individuals are applied. For facilities, the weights are adjusted for survey response levels. The survey estimate of the total number of government facilities compares well against the total published counts. Similarly the female client population estimated in the survey compares well with the total enumerated in the 1991 census.
Release date: 1998-03-12 - Articles and reports: 12-001-X19970023616Description:
A standard method for correcting for unequal sampling probabilities and nonresponse in sample surveys is poststratification: that is, dividing the population into several categories, estimating the distribution of responses in each category, and then counting each category in proportion to its size in the population. We consider poststratification as a general framework that includes many weighting schemes used in survey analysis (see Little 1993). We construct a hierarchical logistic regression model for the mean of a binary response variable conditional on poststratification cells. The hierarchical model allows us to fit many more cells than is possible using classical methods, and thus to include much more population-level information, while at the same time including all the information used in standard survey sampling inferences. We are thus combining the modeling approach often used in small-area estimation with the population information used in poststratification. We apply the method to a set of U.S. pre-election polls, poststratified by state as well as the usual demographic variables. We evaluate the models graphically by comparing to state-level election outcomes.
Release date: 1998-03-12 - Articles and reports: 12-001-X19970023618Description:
Statistical agencies often constitute their business panels by Poisson sampling, or by stratified sampling of fixed size and uniform probabilities in each stratum. This stampling corresponds to algorithms which use permanent numbers following a uniform distribution. Since the characteristics of the units change over time, it is necessary to periodically conduct resamplings while endeavouring to conserve the maximum number of units. The solution by Poisson sampling is the simplest and provides the maximum theoretical coverage, but with the disadvantage of a random sample size. On the other hand, in the case of stratified sampling of fixed size, the changes in strata cause difficulties precisely because of these fixed size constraints. An initial difficulty is that the finer the stratification, the more the coverage is decreased. Indeed, this is likely to occur if births constitute separate strata. We show how this effect can be corrected by rendering the numbers equidistant before resampling. The disadvantage, a fairly minor one, is that in each stratum the sampling is no longer a simple random sampling, which makes the estimation of the variance less rigorous. Another difficulty is reconciling the resampling with an eventual rotation of the units in the sample. We present a type of algorithm which extends after resampling the rotation before resampling. It is based on transformations of the random numbers used for the sampling, so as to return to resampling without rotation. These transformations are particularly simple when they involve equidistant numbers, but can also be carried out with the numbers following a uniform distribution.
Release date: 1998-03-12
Data (0)
Data (0) (0 results)
No content available at this time.
Analysis (6)
Analysis (6) ((6 results))
- Articles and reports: 12-001-X19980013904Description:
Many economic and agricultural surveys are multi-purpose. It would be convenient if one could stratify the target population of such a survey in a number of different purposes and then combine the samples for enumeration. We explore four different sampling methods that select similar samples across all stratifications thereby reducing the overall sample size. Data from an agriculture survey is used to evaluate the effectiveness of these alternative sampling strategies. We then show how a calibration (i.e., reweighted) estimator can increase statistical efficiency by capturing what is known about the original stratum sizes in the estimation. Raking, which has been suggested in the literature for this purpose, is simply one method of calibration.
Release date: 1998-07-31 - Articles and reports: 12-001-X19980013905Description:
Two-phase sampling designs offer a variety of possibilities for use of auxiliary information. We begin by reviewing the different forms that auxiliary information may take in two-phase surveys. We then set up the procedure by which this information is transformed into calibrated weights, which we use to construct efficient estimators of a population total. The calibration is done in two steps: (i) at the population level; (ii) at the level of the first-phase sample. We go on to show that the resulting calibration estimators are also derivable via regression fitting in two steps. We examine these estimators for a special case of interest, namely, when auxiliary information is available for population subgroups called calibration groups. Postrata are the simplest example of such groups. Estimation for domains of interest and variance estimation are also discussed. These results are illustrated by applying them to two-phase designs at Statistics Canada. The general theory for using auxiliary information in two-phase sampling is being incorporated into Statistics Canada's Generalized Estimation System.
Release date: 1998-07-31 - Articles and reports: 12-001-X19980013910Description:
Let A be a population domain of interest and assume that the elements of A cannot be identified on the sampling frame and the number of elements in A is not known. Further assume that a sample of fixed size (say n) is selected from the entire frame and the resulting domain sample size (say n_A) is random. The problem addressed is the construction of a confidence interval for a domain parameter such as the domain aggregate T_A = \sum_{i \in A} x_i. The usual approach to this problem is to redefine x_i, by setting x_i = 0 if i \notin A. Thus, the construction of a confidence interval for the domain total is recast as the construction of a confidence interval for a population total which can be addressed (at least asymptotically in n) by normal theory. As an alternative, we condition on n_A and construct confidence intervals which have approximately nominal coverage under certain assumptions regarding the domain population. We evaluate the new approach empirically using artificial populations and data from the Bureau of Labor Statistics (BLS) Occupational Compensation Survey.
Release date: 1998-07-31 - Articles and reports: 12-001-X19970023615Description:
This paper demonstrates the utility of a multi-stage survey design that obtains a total count of health facilities and of the potential client population in an area. The design has been used for a state-level survey conducted in mid-1995 in Uttar Pradesh, India. The design involves a multi-stage, areal cluster sample, wherein the primary sampling unit is either an urban block or rural village. All health service delivery points, either self-standing facilities or distribution agents, in or formally assigned to the primary sampling unit are mapped, listed, and selected. A systematic sample of households is selected, and all resident females meeting predetermined eligibility criteria are interviewed. Sample weights for facilities and individuals are applied. For facilities, the weights are adjusted for survey response levels. The survey estimate of the total number of government facilities compares well against the total published counts. Similarly the female client population estimated in the survey compares well with the total enumerated in the 1991 census.
Release date: 1998-03-12 - Articles and reports: 12-001-X19970023616Description:
A standard method for correcting for unequal sampling probabilities and nonresponse in sample surveys is poststratification: that is, dividing the population into several categories, estimating the distribution of responses in each category, and then counting each category in proportion to its size in the population. We consider poststratification as a general framework that includes many weighting schemes used in survey analysis (see Little 1993). We construct a hierarchical logistic regression model for the mean of a binary response variable conditional on poststratification cells. The hierarchical model allows us to fit many more cells than is possible using classical methods, and thus to include much more population-level information, while at the same time including all the information used in standard survey sampling inferences. We are thus combining the modeling approach often used in small-area estimation with the population information used in poststratification. We apply the method to a set of U.S. pre-election polls, poststratified by state as well as the usual demographic variables. We evaluate the models graphically by comparing to state-level election outcomes.
Release date: 1998-03-12 - Articles and reports: 12-001-X19970023618Description:
Statistical agencies often constitute their business panels by Poisson sampling, or by stratified sampling of fixed size and uniform probabilities in each stratum. This stampling corresponds to algorithms which use permanent numbers following a uniform distribution. Since the characteristics of the units change over time, it is necessary to periodically conduct resamplings while endeavouring to conserve the maximum number of units. The solution by Poisson sampling is the simplest and provides the maximum theoretical coverage, but with the disadvantage of a random sample size. On the other hand, in the case of stratified sampling of fixed size, the changes in strata cause difficulties precisely because of these fixed size constraints. An initial difficulty is that the finer the stratification, the more the coverage is decreased. Indeed, this is likely to occur if births constitute separate strata. We show how this effect can be corrected by rendering the numbers equidistant before resampling. The disadvantage, a fairly minor one, is that in each stratum the sampling is no longer a simple random sampling, which makes the estimation of the variance less rigorous. Another difficulty is reconciling the resampling with an eventual rotation of the units in the sample. We present a type of algorithm which extends after resampling the rotation before resampling. It is based on transformations of the random numbers used for the sampling, so as to return to resampling without rotation. These transformations are particularly simple when they involve equidistant numbers, but can also be carried out with the numbers following a uniform distribution.
Release date: 1998-03-12
Reference (0)
Reference (0) (0 results)
No content available at this time.
- Date modified: