Analysis

Skip to filters. View results.

What’s new on our website

View results
Sort Help
entries

Results

All (16)

All (16) (0 to 10 of 16 results)

  • 1. Robust variance estimators for generalized regression estimators in cluster samples
    Articles and reports: 12-001-X201900300001
    Description:

    Standard linearization estimators of the variance of the general regression estimator are often too small, leading to confidence intervals that do not cover at the desired rate. Hat matrix adjustments can be used in two-stage sampling that help remedy this problem. We present theory for several new variance estimators and compare them to standard estimators in a series of simulations. The proposed estimators correct negative biases and improve confidence interval coverage rates in a variety of situations that mirror ones that are met in practice.

    Release date: 2019-12-17
  • 2. Model-assisted calibration of non-probability sample survey data using adaptive LASSO Archived
    Articles and reports: 12-001-X201800154963
    Description:

    The probability-sampling-based framework has dominated survey research because it provides precise mathematical tools to assess sampling variability. However increasing costs and declining response rates are expanding the use of non-probability samples, particularly in general population settings, where samples of individuals pulled from web surveys are becoming increasingly cheap and easy to access. But non-probability samples are at risk for selection bias due to differential access, degrees of interest, and other factors. Calibration to known statistical totals in the population provide a means of potentially diminishing the effect of selection bias in non-probability samples. Here we show that model calibration using adaptive LASSO can yield a consistent estimator of a population total as long as a subset of the true predictors is included in the prediction model, thus allowing large numbers of possible covariates to be included without risk of overfitting. We show that the model calibration using adaptive LASSO provides improved estimation with respect to mean square error relative to standard competitors such as generalized regression (GREG) estimators when a large number of covariates are required to determine the true model, with effectively no loss in efficiency over GREG when smaller models will suffice. We also derive closed form variance estimators of population totals, and compare their behavior with bootstrap estimators. We conclude with a real world example using data from the National Health Interview Survey.

    Release date: 2018-06-21
  • 3. A design effect measure for calibration weighting in single-stage samples Archived
    Articles and reports: 12-001-X201500214236
    Description:

    We propose a model-assisted extension of weighting design-effect measures. We develop a summary-level statistic for different variables of interest, in single-stage sampling and under calibration weight adjustments. Our proposed design effect measure captures the joint effects of a non-epsem sampling design, unequal weights produced using calibration adjustments, and the strength of the association between an analysis variable and the auxiliaries used in calibration. We compare our proposed measure to existing design effect measures in simulations using variables like those collected in establishment surveys and telephone surveys of households.

    Release date: 2015-12-17
  • 4. Estimation with Non-probability Surveys and the Question of External Validity Archived
    Articles and reports: 11-522-X201300014288
    Description:

    Probability-based surveys, those including with samples selected through a known randomization mechanism, are considered by many to be the gold standard in contrast to non-probability samples. Probability sampling theory was first developed in the early 1930’s and continues today to justify the estimation of population values from these data. Conversely, studies using non-probability samples have gained attention in recent years but they are not new. Touted as cheaper, faster (even better) than probability designs, these surveys capture participants through various “on the ground” methods (e.g., opt-in web survey). But, which type of survey is better? This paper is the first in a series on the quest for a quality framework under which all surveys, probability- and non-probability-based, may be measured on a more equal footing. First, we highlight a few frameworks currently in use, noting that “better” is almost always relative to a survey’s fit for purpose. Next, we focus on the question of validity, particularly external validity when population estimates are desired. Estimation techniques used to date for non-probability surveys are reviewed, along with a few comparative studies of these estimates against those from a probability-based sample. Finally, the next research steps in the quest are described, followed by a few parting comments.

    Release date: 2014-10-31
  • 5. Condition indexes and variance decompositions for diagnosing collinearity in linear model analysis of survey data Archived
    Articles and reports: 12-001-X201200211757
    Description:

    Collinearities among explanatory variables in linear regression models affect estimates from survey data just as they do in non-survey data. Undesirable effects are unnecessarily inflated standard errors, spuriously low or high t-statistics, and parameter estimates with illogical signs. The available collinearity diagnostics are not generally appropriate for survey data because the variance estimators they incorporate do not properly account for stratification, clustering, and survey weights. In this article, we derive condition indexes and variance decompositions to diagnose collinearity problems in complex survey data. The adapted diagnostics are illustrated with data based on a survey of health characteristics.

    Release date: 2012-12-19
  • 6. Variance inflation factors in the analysis of complex survey data Archived
    Articles and reports: 12-001-X201200111685
    Description:

    Survey data are often used to fit linear regression models. The values of covariates used in modeling are not controlled as they might be in an experiment. Thus, collinearity among the covariates is an inevitable problem in the analysis of survey data. Although many books and articles have described the collinearity problem and proposed strategies to understand, assess and handle its presence, the survey literature has not provided appropriate diagnostic tools to evaluate its impact on regression estimation when the survey complexities are considered. We have developed variance inflation factors (VIFs) that measure the amount that variances of parameter estimators are increased due to having non-orthogonal predictors. The VIFs are appropriate for survey-weighted regression estimators and account for complex design features, e.g., weights, clusters, and strata. Illustrations of these methods are given using a probability sample from a household survey of health and nutrition.

    Release date: 2012-06-27
  • 7. A comparison of variance estimators for poststratification to estimated control totals Archived
    Articles and reports: 12-001-X201000111251
    Description:

    Calibration techniques, such as poststratification, use auxiliary information to improve the efficiency of survey estimates. The control totals, to which sample weights are poststratified (or calibrated), are assumed to be population values. Often, however, the controls are estimated from other surveys. Many researchers apply traditional poststratification variance estimators to situations where the control totals are estimated, thus assuming that any additional sampling variance associated with these controls is negligible. The goal of the research presented here is to evaluate variance estimators for stratified, multi-stage designs under estimated-control (EC) poststratification using design-unbiased controls. We compare the theoretical and empirical properties of linearization and jackknife variance estimators for a poststratified estimator of a population total. Illustrations are given of the effects on variances from different levels of precision in the estimated controls. Our research suggests (i) traditional variance estimators can seriously underestimate the theoretical variance, and (ii) two EC poststratification variance estimators can mitigate the negative bias.

    Release date: 2010-06-29
  • 8. Survey weighted hat matrix and leverages Archived
    Articles and reports: 12-001-X200900110881
    Description:

    Regression diagnostics are geared toward identifying individual points or groups of points that have an important influence on a fitted model. When fitting a model with survey data, the sources of influence are the response variable Y, the predictor variables X, and the survey weights, W. This article discusses the use of the hat matrix and leverages to identify points that may be influential in fitting linear models due to large weights or values of predictors. We also contrast findings that an analyst will obtain if ordinary least squares is used rather than survey weighted least squares to determine which points are influential.

    Release date: 2009-06-22
  • 9. Cell collapsing in poststratification Archived
    Articles and reports: 12-001-X200700210491
    Description:

    Poststratification is a common method of estimation in household surveys. Cells are formed based on characteristics that are known for all sample respondents and for which external control counts are available from a census or another source. The inverses of the poststratification adjustments are usually referred to as coverage ratios. Coverage of some demographic groups may be substantially below 100 percent, and poststratifying serves to correct for biases due to poor coverage. A standard procedure in poststratification is to collapse or combine cells when the sample sizes fall below some minimum or the weight adjustments are above some maximum. Collapsing can either increase or decrease the variance of an estimate but may simultaneously increase its bias. We study the effects on bias and variance of this type of dynamic cell collapsing theoretically and through simulation using a population based on the 2003 National Health Interview Survey. Two alternative estimators are also proposed that restrict the size of weight adjustments when cells are collapsed.

    Release date: 2008-01-03
  • 10. Variance estimation with hot deck imputation: A simulation study of three methods Archived
    Articles and reports: 12-001-X20050029044
    Description:

    Complete data methods for estimating the variances of survey estimates are biased when some data are imputed. This paper uses simulation to compare the performance of the model-assisted, the adjusted jackknife, and the multiple imputation methods for estimating the variance of a total when missing items have been imputed using hot deck imputation. The simulation studies the properties of the variance estimates for imputed estimates of totals for the full population and for domains from a single-stage disproportionate stratified sample design when underlying assumptions, such as unbiasedness of the point estimate and item responses being randomly missing within hot deck cells, do not hold. The variance estimators for full population estimates produce confidence intervals with coverage rates near the nominal level even under modest departures from the assumptions, but this finding does not apply for the domain estimates. Coverage is most sensitive to bias in the point estimates. As the simulation demonstrates, even if an imputation method gives almost unbiased estimates for the full population, estimates for domains may be very biased.

    Release date: 2006-02-17
Stats in brief (0)

Stats in brief (0) (0 results)

No content available at this time.

Articles and reports (16)

Articles and reports (16) (0 to 10 of 16 results)

  • 1. Robust variance estimators for generalized regression estimators in cluster samples
    Articles and reports: 12-001-X201900300001
    Description:

    Standard linearization estimators of the variance of the general regression estimator are often too small, leading to confidence intervals that do not cover at the desired rate. Hat matrix adjustments can be used in two-stage sampling that help remedy this problem. We present theory for several new variance estimators and compare them to standard estimators in a series of simulations. The proposed estimators correct negative biases and improve confidence interval coverage rates in a variety of situations that mirror ones that are met in practice.

    Release date: 2019-12-17
  • 2. Model-assisted calibration of non-probability sample survey data using adaptive LASSO Archived
    Articles and reports: 12-001-X201800154963
    Description:

    The probability-sampling-based framework has dominated survey research because it provides precise mathematical tools to assess sampling variability. However increasing costs and declining response rates are expanding the use of non-probability samples, particularly in general population settings, where samples of individuals pulled from web surveys are becoming increasingly cheap and easy to access. But non-probability samples are at risk for selection bias due to differential access, degrees of interest, and other factors. Calibration to known statistical totals in the population provide a means of potentially diminishing the effect of selection bias in non-probability samples. Here we show that model calibration using adaptive LASSO can yield a consistent estimator of a population total as long as a subset of the true predictors is included in the prediction model, thus allowing large numbers of possible covariates to be included without risk of overfitting. We show that the model calibration using adaptive LASSO provides improved estimation with respect to mean square error relative to standard competitors such as generalized regression (GREG) estimators when a large number of covariates are required to determine the true model, with effectively no loss in efficiency over GREG when smaller models will suffice. We also derive closed form variance estimators of population totals, and compare their behavior with bootstrap estimators. We conclude with a real world example using data from the National Health Interview Survey.

    Release date: 2018-06-21
  • 3. A design effect measure for calibration weighting in single-stage samples Archived
    Articles and reports: 12-001-X201500214236
    Description:

    We propose a model-assisted extension of weighting design-effect measures. We develop a summary-level statistic for different variables of interest, in single-stage sampling and under calibration weight adjustments. Our proposed design effect measure captures the joint effects of a non-epsem sampling design, unequal weights produced using calibration adjustments, and the strength of the association between an analysis variable and the auxiliaries used in calibration. We compare our proposed measure to existing design effect measures in simulations using variables like those collected in establishment surveys and telephone surveys of households.

    Release date: 2015-12-17
  • 4. Estimation with Non-probability Surveys and the Question of External Validity Archived
    Articles and reports: 11-522-X201300014288
    Description:

    Probability-based surveys, those including with samples selected through a known randomization mechanism, are considered by many to be the gold standard in contrast to non-probability samples. Probability sampling theory was first developed in the early 1930’s and continues today to justify the estimation of population values from these data. Conversely, studies using non-probability samples have gained attention in recent years but they are not new. Touted as cheaper, faster (even better) than probability designs, these surveys capture participants through various “on the ground” methods (e.g., opt-in web survey). But, which type of survey is better? This paper is the first in a series on the quest for a quality framework under which all surveys, probability- and non-probability-based, may be measured on a more equal footing. First, we highlight a few frameworks currently in use, noting that “better” is almost always relative to a survey’s fit for purpose. Next, we focus on the question of validity, particularly external validity when population estimates are desired. Estimation techniques used to date for non-probability surveys are reviewed, along with a few comparative studies of these estimates against those from a probability-based sample. Finally, the next research steps in the quest are described, followed by a few parting comments.

    Release date: 2014-10-31
  • 5. Condition indexes and variance decompositions for diagnosing collinearity in linear model analysis of survey data Archived
    Articles and reports: 12-001-X201200211757
    Description:

    Collinearities among explanatory variables in linear regression models affect estimates from survey data just as they do in non-survey data. Undesirable effects are unnecessarily inflated standard errors, spuriously low or high t-statistics, and parameter estimates with illogical signs. The available collinearity diagnostics are not generally appropriate for survey data because the variance estimators they incorporate do not properly account for stratification, clustering, and survey weights. In this article, we derive condition indexes and variance decompositions to diagnose collinearity problems in complex survey data. The adapted diagnostics are illustrated with data based on a survey of health characteristics.

    Release date: 2012-12-19
  • 6. Variance inflation factors in the analysis of complex survey data Archived
    Articles and reports: 12-001-X201200111685
    Description:

    Survey data are often used to fit linear regression models. The values of covariates used in modeling are not controlled as they might be in an experiment. Thus, collinearity among the covariates is an inevitable problem in the analysis of survey data. Although many books and articles have described the collinearity problem and proposed strategies to understand, assess and handle its presence, the survey literature has not provided appropriate diagnostic tools to evaluate its impact on regression estimation when the survey complexities are considered. We have developed variance inflation factors (VIFs) that measure the amount that variances of parameter estimators are increased due to having non-orthogonal predictors. The VIFs are appropriate for survey-weighted regression estimators and account for complex design features, e.g., weights, clusters, and strata. Illustrations of these methods are given using a probability sample from a household survey of health and nutrition.

    Release date: 2012-06-27
  • 7. A comparison of variance estimators for poststratification to estimated control totals Archived
    Articles and reports: 12-001-X201000111251
    Description:

    Calibration techniques, such as poststratification, use auxiliary information to improve the efficiency of survey estimates. The control totals, to which sample weights are poststratified (or calibrated), are assumed to be population values. Often, however, the controls are estimated from other surveys. Many researchers apply traditional poststratification variance estimators to situations where the control totals are estimated, thus assuming that any additional sampling variance associated with these controls is negligible. The goal of the research presented here is to evaluate variance estimators for stratified, multi-stage designs under estimated-control (EC) poststratification using design-unbiased controls. We compare the theoretical and empirical properties of linearization and jackknife variance estimators for a poststratified estimator of a population total. Illustrations are given of the effects on variances from different levels of precision in the estimated controls. Our research suggests (i) traditional variance estimators can seriously underestimate the theoretical variance, and (ii) two EC poststratification variance estimators can mitigate the negative bias.

    Release date: 2010-06-29
  • 8. Survey weighted hat matrix and leverages Archived
    Articles and reports: 12-001-X200900110881
    Description:

    Regression diagnostics are geared toward identifying individual points or groups of points that have an important influence on a fitted model. When fitting a model with survey data, the sources of influence are the response variable Y, the predictor variables X, and the survey weights, W. This article discusses the use of the hat matrix and leverages to identify points that may be influential in fitting linear models due to large weights or values of predictors. We also contrast findings that an analyst will obtain if ordinary least squares is used rather than survey weighted least squares to determine which points are influential.

    Release date: 2009-06-22
  • 9. Cell collapsing in poststratification Archived
    Articles and reports: 12-001-X200700210491
    Description:

    Poststratification is a common method of estimation in household surveys. Cells are formed based on characteristics that are known for all sample respondents and for which external control counts are available from a census or another source. The inverses of the poststratification adjustments are usually referred to as coverage ratios. Coverage of some demographic groups may be substantially below 100 percent, and poststratifying serves to correct for biases due to poor coverage. A standard procedure in poststratification is to collapse or combine cells when the sample sizes fall below some minimum or the weight adjustments are above some maximum. Collapsing can either increase or decrease the variance of an estimate but may simultaneously increase its bias. We study the effects on bias and variance of this type of dynamic cell collapsing theoretically and through simulation using a population based on the 2003 National Health Interview Survey. Two alternative estimators are also proposed that restrict the size of weight adjustments when cells are collapsed.

    Release date: 2008-01-03
  • 10. Variance estimation with hot deck imputation: A simulation study of three methods Archived
    Articles and reports: 12-001-X20050029044
    Description:

    Complete data methods for estimating the variances of survey estimates are biased when some data are imputed. This paper uses simulation to compare the performance of the model-assisted, the adjusted jackknife, and the multiple imputation methods for estimating the variance of a total when missing items have been imputed using hot deck imputation. The simulation studies the properties of the variance estimates for imputed estimates of totals for the full population and for domains from a single-stage disproportionate stratified sample design when underlying assumptions, such as unbiasedness of the point estimate and item responses being randomly missing within hot deck cells, do not hold. The variance estimators for full population estimates produce confidence intervals with coverage rates near the nominal level even under modest departures from the assumptions, but this finding does not apply for the domain estimates. Coverage is most sensitive to bias in the point estimates. As the simulation demonstrates, even if an imputation method gives almost unbiased estimates for the full population, estimates for domains may be very biased.

    Release date: 2006-02-17
Journals and periodicals (0)

Journals and periodicals (0) (0 results)

No content available at this time.

Report a problem or mistake on this page
Date modified:

Search help

How do I use the filters and the search box?

  • You can initiate a search by entering keywords or selecting filters (e.g., under Subject, Geography, etc.) on the left side of the page.

  • Filters can be used together or in any combination. Each time you select a filter, the results on the page are updated.

  • To start a new search, click on the Clear all button above the search box or deselect each filter.

  • Specified keywords and filters appear above the search box. You can deselect any or all the tags to refine or clear your search.

How do I refine my search?

  • You can enter keywords in the search box. It is not necessary to use “+” or “, “ or “AND”.

  • You can delete some or all of the keywords from your search string.

  • Quotes around keywords limit the search to a specific phrase.

    • For example, “labour force survey” will limit results to products containing this phrase.

  • Use “or” between keywords to obtain results that contain one or more search terms.

    • For example, labour or­ force or survey will limit results to products containing any or all of these words.

How does the search work?

  • This search looks for results that have your query term in the title, description, subject, geography, product number or other information about the product.

    • For example, when searching the term ‘Diseases’, all results with the word ‘Diseases’ in the title, description or tagged the subject appear.

  • The text within an article or publication are not searched. For full text searching of articles, use the site search.

Sort help

How are the results ordered?

The Sort by feature provides three options:

  • By date – sorts results by product release date starting with the most recent.
  • By relevance is only needed or beneficial when you use the search box and submit a search query. Results are ordered by a combination of where the search term appears in the content, e.g., if it appears in the title, it likely means the product is more relevant to your search; how up to date or new the content in the product is; how popular the product is; and other factors.
  • By title – sorts results alphabetically.

Sort help

How are the results ordered?

The Sort by feature provides four options:

  • By date – sorts results by product release date starting with the most recent.
  • By most popular – gives priority to results that are most frequently selected by users, followed by the most recent.
  • By relevance is only needed or beneficial when you use the search box and submit a search query. Results are ordered by a combination of where the search term appears in the content, e.g., if it appears in the title, it likely means the product is more relevant to your search; how up to date or new the content in the product is; how popular the product is; and other factors.
  • By title – sorts results alphabetically.