Keyword search

The U.S. Bureau of the Census implemented major changes to the design of the Survey of Income and Program Participation (SIPP) with the panel begun in 1996. The revised survey design emphasized longitudinal applications and the Census Bureau attempted to understand and resolve the seam bias common to longitudinal surveys. In addition to the substantive and administrative redesign of the survey, the Census Bureau is improving the data processing procedures which yield microdata files for the public to analyse. The wave-by-wave data products are being edited and imputed with a longitudinal element rather than cross-sectionally, carrying forward information from a prior wave that is missing in the current wave. The longitudinal data products will be enhanced, both by the redesigned survey and new processing procedures. Simple methods of imputing data over time are being replaced with more sophisticated methods that do not attenuate seam bias. The longitudinal sample is expanding to include more observations which were nonrespondents in one or more waves. Longitudinal weights will be applied to the file to support person-based longitudinal analysis for calendar years or longer periods of time (up to four years).

This article extends and further develops the method proposed by Pfeffermann, Skinner and Humphreys (1998) for the estimation of gross flows in the presence of classification errors. The main feature of that method is the use of auxiliary information at the individual level which circumvents the need for validation data for estimating the misclassification rates. The new developments in this article are the establishment of conditions for model identification, a study of the properties of a model goodness of fit statistic and modifications to the sample likelihood to account for missing data and informative sampling. The new developments are illustrated by a small Monte-Carlo simulation study.

The study of social mobility, between labour market statuses or between income levels, for example, is often based on the analysis of mobility matrices. When comparing these transition matrices, with a view to evaluating behavioural changes, one often forgets that the data derive from a sample survey and are therefore affected by sampling variances. Similarly, it is assumed that the responses collected correspond to the ' true value.'

A longitudinal study on a cohort of pupils in the secondary school has been conducted in an Italian region since 1986 in order to study the transition from school to working life. The information have been collected at every sweep by a mail questionnaire and, at the final sweep, by a face-to-face interview, where retrospective questions referring back to the whole observation period have been asked. The gross flows between different discrete states - still in the school system, in the labour force without a job, in the labour force with a job - may then be estimated both from prospective and retrospective data, and the recall effect may be evaluated. Moreover, the conditions observed by the two different techniques may be regarded as two indicators of the 'true' unobservable condition, thus leading to the specification and estimation of a latent class model. In this framework, a Markov chain hypothesis may be introduced and evaluated in order to estimate the transition probabilities between the states, once they are corrected or the classification errors. Since the information collected by mail show a given amount of missing data in terms of unit nonresponse, the 'missing' category is also introduced in the model specification.

The log-linear modelling of categorical longitudinal survey data on income is studied. An emphasis is on inference about change. Special attention is paid to modelling of longitudinal data from two waves. A small illustration is based on data from the Canadian Survey of Labour and Income Dynamics.

The purpose of the present study is to utilize panel data from the Current Population Survey (CPS) to examine the effects of unit nonresponse. Because most nonrespondents to the CPS are respondents during at least one month-in-sample, data from other months can be used to compare the characteristics of complete respondents and panel nonrespondents and to evaluate nonresponse adjustment procedures. In the current paper we present analyses utilizing CPS panel data to illustrate the effects of unit nonresponse. After adjusting for nonresponse, additional comparisons are also made to evaluate the effects of nonresponse adjustment. The implications of the findings and suggestions for further research are discussed.

The disseminated results of annual business surveys inevitably contain statistics that are changing. Since the economic sphere is increasingly dynamic, a simple difference of aggregates between n-l and n is no longer sufficient to provide an overall description of what has happened. The change calculation module in the new generation of annual business surveys divides overall change into various components (births, deaths, inter-industry migration) and calculates change on the basis of a constant field, assigning special importance to restructurings. The main difficulties lie in establishing subsamples, reweighting, calibrating according to calculable changes, and taking account of restructuring.

We address the problem of estimation for the income dynamics statistics calculated from complex longitudinal surveys. In addition, we compare two design-based estimators of longitudinal proportions and transition rates in terms of variability under large attrition rates. One estimator is based on the cross-sectional samples for the estimation of the income class boundaries at each time period and on the longitudinal sample for the estimation of the longitudinal counts; the other estimator is entirely based on the longitudinal sample, both for the estimation of the class boundaries and the longitudinal counts. We develop Taylor linearization-type variance estimators for both the longitudinal and the mixed estimator under the assumption of no change in the population, and for the mixed estimator when there is change.

In longitudinal surveys, sample subjects are observed over several time points. This feature typically leads to dependent observations on the same subject, in addition to the customary correlations across subjects induced by the sample design. Much research in the literature has focussed on modeling the marginal mean of a response as a function of covariates. Liang and Zeger (1986) used generalized estimating equations (GEE), requiring only correct specification of the marginal mean, and obtained standard errors of regression parameter estimates and associated Wald tests, assuming a "working" correlation structure for the repeated measurements on a sample subject. Rotnitzky and Jewell (1990) developed quasi-score tests and Rao-Scott adjustments to "working" quasi-score tests under marginal models. These methods are asymptotically robust to misspecification of the within-subject correlation structure, but assume independence of sample subjects which is not satisfied for complex longitudinal survey data based on stratified multi-stage sampling. We proposed asymptotically valid Wald and quasi-score tests for longitudinal survey data, using the Taylor Linearization and jackknife methods. Alternative tests, based on Rao-Scott adjustments to naive tests that ignore survey design features and on Bonferroni-t, are also developed. These tests are particularly useful when the effective degrees of freedom, usually taken as the total number of sample primary units (clusters) minus the number of strata, is small.

Two-phase sampling designs have been conducted in waves to estimate the incidence of a rare disease such as dementia. Estimation of disease incidence from longitudinal dementia study has to appropriately adjust for data missing by death as well as the sampling design used at each study wave. In this paper we adopt a selection model approach to model the missing data by death and use a likelihood approach to derive incidence estimates. A modified EM algorithm is used to deal with data missing by sampling selection. The non-paramedic jackknife variance estimator is used to derive variance estimates for the model parameters and the incidence estimates. The proposed approaches are applied to data from the Indianapolis-Ibadan Dementia Study.

This document brings together data from a number of Statistics Canada surveys and provides a visual perspective on the following subject areas: crime, police administration, adult and youth court activity, the correctional population, costs of the criminal justice system, violence against women, Canadians' experiences with crime, and their perceptions and fears of crime.

This compendium of vital statistics includes summary data on births, deaths, marriages and divorces. The introduction covers the data sources, data quality, and methods pertaining to each event, and includes a glossary defining the terms used. The first chapter is a brief overview of vital statistics for 1996. Subsequent chapters treat marriage, divorce, birth, fetal and infant mortality, total mortality, causes of death, vital statistics by census division, and international comparisons. Most charts and tables show Canada data for 1986 though 1996, while the charts and tables for causes of death show Canada data for 1979 through1996. Data for the provinces and territories are usually shown for 1995 and 1996. Appendices include population denominator data, age-standardized mortality rate (ASMR) calculation methods, and leading causes of death methodology.

Operating ratios for top carriers improved by one point in the first and second quarter of 1999 over the same period one year earlier. Average revenue per carrier fell 1% in the first quarter and rose only 3% in the second quarter.

On an industry wide basis, Canadian bus companies continue to grow and prosper. Gross revenues in 1998 (excluding subsidies) were 8.6 percent higher than in 1995.

With the exception of the third quarter (0.94), for-hire motor carriers of freight posted seasonally adjusted operating ratios of 0.93 in three out of the four quarters of 1998.

Data on registration of motor vehicles by type including passenger automobiles, trucks, motorcycles, buses, trailers and others are presented in this publication. A historical table of total registrations is provided. Motor vehicle registrations are shown by census divisions and municipalities where available. Data definitions, analysis, the methodology employed, an explanation of data quality and a bibliography are included.

This on-line publication provides detailed capital expenditures by type of asset on both construction and machinery and equipment made by private and public organizations in Canada. For each province and territory and for the 19 divisions of the Canadian economy, it details capital expenditures according to four types of residential construction, 95 types of non-residential construction, and 56 categories of machinery and equipment. Included are data on capital expenditures for major renovation and alteration of construction assets as well as for major retrofit and refurbishing of machinery and equipment assets. Trade and general construction contractors, suppliers of construction materials, and suppliers of machinery and equipment will find these data useful for market analysis.

This monthly publication presents data by weight on the acquisition, stock, production and shipment of raw and refined sugar for the month and the year-to-date from refiners of raw cane or beet sugars as well as those who further process such sugars. Geographic detail is at the national level. The December issue includes a list of reporting firms.

Consulting Engineering Services Price Index (CEPI) is an annual index that measures changes in the prices for services provided by consulting engineers. These services encompass advisory and design work as well as construction or project management. They are provided for many types of projects (fields of specialization), and to both Canadian and foreign clients. Price indexes are published for 10 fields of specialization as well as for national, regional, and foreign markets.

This occasional on-line bulletin series presents price indexes published by the Goods and Services Producing Industries Section, Prices Division. Each issue will contain different service price indexes.

This paper introduces Poisson Mixture sampling, a family of sampling designs so named because each member of the family is a mixture of two Poisson sampling designs, Poisson nps sampling and Bernoulli sampling. These two designs are at opposite ends of a continuous spectrum, indexed by a continuous parameter. Poisson Mixture sampling is conceived for use with the highly skewed populations often arising in business surveys. It gives the statistician a range of different options for the extent of the sample coordination and the control of response burden. Some Poisson Mixture sampling designs give considerably more precise estimates than the usual Poisson nps sampling. This result is noteworthy, because Poisson nps is in itself highly efficient, assuming it is based on a strong measure of size.

We develop an approach to estimating variances for X-11 seasonal adjustments that recognizes the effects of sampling error and errors from forecast extension. In our approach, seasonal adjustment error in the central values of a sufficiently long series results only from the effect of the X-11 filtering on the sampling errors. Towards either end of the series, we also recognize the contribution to seasonal adjustment error from forecast and backcast errors. We extend the approach to produce variances of errors in X-11 trend estimates, and to recognize error in estimation of regression coefficients used to model, e.g., calendar effects. In empirical results, the contribution of sampling error often dominated the seasonal adjustment variances. Trend estimate variances, however, showed large increases at the ends of series due to the effects of fore/backcast error. Nonstationarities in the sampling errors produced striking patterns in the seasonal adjustment and trend estimate variances.

Most statistical offices select the sample of commodities of which prices are collected for their Consumer Price Indexes with non-probability techniques. In the Netherlands, and in many other countries as well, those judgemental sampling methods come close to some kind of cut-off selection, in which a large part of the population (usually the items with the lowest expenditures) is deliberately left unobserved. This method obviously yields biased price index numbers. The question arises whether probability sampling would lead to better results in terms of the mean square error. We have considered simple random sampling, stratified sampling and systematic sampling proportional to expenditure. Monte Carlo simulations using scanner data on coffee, baby's napkins and toilet paper were carried out to assess the performance of the four sampling designs. Surprisingly perhaps, cut-off selection is shown to be a successful strategy for item sampling in the consumer price index.

We consider the use of calibration estimators when outliers occur. An extension is obtained for the class of Deville and Särndal (1992) calibration estimators based on Wright (1983) QR estimators. It is also obtained by minimizing a general metric subject to constraints on the calibration variables and weights. As an application, this class of estimators helps us consider robust calibration estimators by choosing parameters carefully. This makes it possible, e.g., for cosmetic reasons, to limit robuts weights to a predetermined interval. The use of robust estimators with a high breakdown point is also considered. In the specific case of the mean square metric, the estimator proposed by the author is a generalization of a Lee (1991) proposition. The new methodology is illustrated by means of a short simulation study.

This paper investigates a repeated sampling approach to take into account auxiliary information in order to improve the precision of estimators. The objective is to build an estimator with a small conditional bias by weighting the observed values by the inverses of the conditional inclusion probabilities. A general approximation is proposed in cases when the auxiliary static is a vector of Horvitz-Thompson estimators. This approximation is quite close to the optimal estimator discussed by Fuller and Isaki (1981), Montanari (1987, 1997), Deville (1992) and Rao (1994, 1997). Next, the optimal estimator is applied to a stratified sampling design and it is shown that the optimal estimator can be viewed as a generalised regression estimator for which the stratification indicator variables are also used at the estimation stage. Finally, the application field of this estimator is discussed in the general context of the use of auxiliary information.

Robust small area estimation is studied under a simple random effects model consisting of a basic (or fixed effects) model and a linking model that treats the fixed effects as realizations of a random variable. Under this model a model-assisted estimator of a small area mean is obtained. This estimator depends on the survey weights and remains design-consistent. A model-based estimator of its mean squared error (MSE) is also obtained. Simulation results suggest that the proposed estimator and Kott's (1989) model-assisted estimator are equally efficient, and that the proposed MSE estimator is often much more stable than Kott's MSE estimator, even under moderate deviations of the linking model. The method is also extended to nested error regression models.

In this paper a general multilevel model framework is used to provide estimates for small areas using survey data. This class of models allows for variation between areas because of: (i) differences in the distributions of unit level variables between areas, (ii) differences in the distribution of area level variables between areas (iii) area specific components of variance which make provision for additional local variation which cannot be explained by unit-level or area-level covariates. Small area estimators are derived for this multilevel model formulation and an approximation to the mean square error (MSE) of each small area estimates for this general class of mixed models is provided together with an estimator of this MSE. Both the approximations to the MSE and the estimator of MSE take into account three sources of variation: (i) the prediction MSE assuming that both the fixed and components of variance terms in the multilevel model are knows, (ii) the additional component due to the fact that the fixed coefficients must be estimated, and (iii) the further component due to the fact that the components of variance in the model must be estimated. The proposed methods are estimated using a large data set as a basis for numerical investigation. The results confirm that the extra components of variance contained in multilevel models as well as small area covariates can improve small area estimates and that the MSE approximation and estimator are satisfactory.

The Gallup Organization has been conducting household surveys to study state-wide prevalences of alcohol and drug (e.g., cocaine, marijuana, etc.) use. Traditional design-based survey estimates of use and dependence for counties and select demographic groups have unacceptably large standard errors because sample sizes in sub-state groups are two small. Synthetic estimation incorporates demographic information and social indicators in estimates of prevalence through an implicit regression model. Synthetic estimates tend to have smaller variances than design-based estimates, but can be very homogeneous across counties when auxiliary variables are homogeneous. Composite estimates for small areas are weighted averages of design-based survey estimates and synthetic estimates. A second problem generally not encountered at the state level but present for sub-state areas and groups concerns estimating standard errors of estimated prevalences that are close to zero. This difficulty affects not only telephone household survey estimates, but also composite estimates. A hierarchical model is proposed to address this problem. Empirical Bayes composite estimators, which incorporate survey weights, of prevalences and jackknife estimators of their mean squared errors are presented and illustrated.

Two design-based estimators of gross flows and transition rates are considered. One makes use of the cross-sectional samples for the estimation of the income class boundaries at each time period and the longitudinal sample for the estimation of counts of units in the longitudinal population (longitudinal counts); this is the mixed estimator. The other one is entirely based on the longitudinal sample, both for the estimation of the class boundaries and the longitudinal counts; this is the longitudinal estimator. We compare the two estimators in the presence of large attrition rates, by means of a simulation. We find that under a less than perfect model of compensation for attrition, the mixed estimator is usually more sensitive to model bias than the longitudinal estimator. Furthermore, we find that for the mixed estimator, the magnitude of this bias overshadows the small gain in precision when compared to the longitudinal estimator. The results are illustrated with data from the Survey of Labour and Income Dynamics and the Longitudinal Administrative Database of Statistics Canada.

The British Labour Froce Survey (LFS) uses a rotating sample design, with each sample household retained for five consecutive quarters. Linking together the information on the same persons across quarters produces a potentially very rich source of longitudinal data. There are however serious risks of distortion in the results from such longitudinal linking, mainly arising from sample attrition, and from response errors, which can produce spurious flows between economic activity states. This paper describes the initial results of investigations by the Office for National Statistics (ONS) into the nature and extent of the problems.

At the end of 1993, Eurostat lauched a 'community' panel of households. The first wave, carried out in 1994 in the 12 countries of the European Union, included some 7,300 households in France, and at least 14,000 adults 17 years or over. Each individual was then followed up and interviewed each year, even if they had moved. The individuals leaving the sample present a particular profile. In the first part, we present a sketch of how our sample evolves and an analysis of the main characteristics of the non-respondents. We then propose 2 models to correct for non-response per homogeneous category. We then describe the longitudinal weight distribution obtained from the two models, and the cross-sectional weights using the weight share method. Finally, we compare some indicators calculated using both weighting methods.

The U.S. Bureau of the Census implemented major changes to the design of the Survey of Income and Program Participation (SIPP) with the panel begun in 1996. The revised survey design emphasized longitudinal applications and the Census Bureau attempted to understand and resolve the seam bias common to longitudinal surveys. In addition to the substantive and administrative redesign of the survey, the Census Bureau is improving the data processing procedures which yield microdata files for the public to analyse. The wave-by-wave data products are being edited and imputed with a longitudinal element rather than cross-sectionally, carrying forward information from a prior wave that is missing in the current wave. The longitudinal data products will be enhanced, both by the redesigned survey and new processing procedures. Simple methods of imputing data over time are being replaced with more sophisticated methods that do not attenuate seam bias. The longitudinal sample is expanding to include more observations which were nonrespondents in one or more waves. Longitudinal weights will be applied to the file to support person-based longitudinal analysis for calendar years or longer periods of time (up to four years).

This article extends and further develops the method proposed by Pfeffermann, Skinner and Humphreys (1998) for the estimation of gross flows in the presence of classification errors. The main feature of that method is the use of auxiliary information at the individual level which circumvents the need for validation data for estimating the misclassification rates. The new developments in this article are the establishment of conditions for model identification, a study of the properties of a model goodness of fit statistic and modifications to the sample likelihood to account for missing data and informative sampling. The new developments are illustrated by a small Monte-Carlo simulation study.

The study of social mobility, between labour market statuses or between income levels, for example, is often based on the analysis of mobility matrices. When comparing these transition matrices, with a view to evaluating behavioural changes, one often forgets that the data derive from a sample survey and are therefore affected by sampling variances. Similarly, it is assumed that the responses collected correspond to the ' true value.'

A longitudinal study on a cohort of pupils in the secondary school has been conducted in an Italian region since 1986 in order to study the transition from school to working life. The information have been collected at every sweep by a mail questionnaire and, at the final sweep, by a face-to-face interview, where retrospective questions referring back to the whole observation period have been asked. The gross flows between different discrete states - still in the school system, in the labour force without a job, in the labour force with a job - may then be estimated both from prospective and retrospective data, and the recall effect may be evaluated. Moreover, the conditions observed by the two different techniques may be regarded as two indicators of the 'true' unobservable condition, thus leading to the specification and estimation of a latent class model. In this framework, a Markov chain hypothesis may be introduced and evaluated in order to estimate the transition probabilities between the states, once they are corrected or the classification errors. Since the information collected by mail show a given amount of missing data in terms of unit nonresponse, the 'missing' category is also introduced in the model specification.

The log-linear modelling of categorical longitudinal survey data on income is studied. An emphasis is on inference about change. Special attention is paid to modelling of longitudinal data from two waves. A small illustration is based on data from the Canadian Survey of Labour and Income Dynamics.

The purpose of the present study is to utilize panel data from the Current Population Survey (CPS) to examine the effects of unit nonresponse. Because most nonrespondents to the CPS are respondents during at least one month-in-sample, data from other months can be used to compare the characteristics of complete respondents and panel nonrespondents and to evaluate nonresponse adjustment procedures. In the current paper we present analyses utilizing CPS panel data to illustrate the effects of unit nonresponse. After adjusting for nonresponse, additional comparisons are also made to evaluate the effects of nonresponse adjustment. The implications of the findings and suggestions for further research are discussed.

The disseminated results of annual business surveys inevitably contain statistics that are changing. Since the economic sphere is increasingly dynamic, a simple difference of aggregates between n-l and n is no longer sufficient to provide an overall description of what has happened. The change calculation module in the new generation of annual business surveys divides overall change into various components (births, deaths, inter-industry migration) and calculates change on the basis of a constant field, assigning special importance to restructurings. The main difficulties lie in establishing subsamples, reweighting, calibrating according to calculable changes, and taking account of restructuring.

We address the problem of estimation for the income dynamics statistics calculated from complex longitudinal surveys. In addition, we compare two design-based estimators of longitudinal proportions and transition rates in terms of variability under large attrition rates. One estimator is based on the cross-sectional samples for the estimation of the income class boundaries at each time period and on the longitudinal sample for the estimation of the longitudinal counts; the other estimator is entirely based on the longitudinal sample, both for the estimation of the class boundaries and the longitudinal counts. We develop Taylor linearization-type variance estimators for both the longitudinal and the mixed estimator under the assumption of no change in the population, and for the mixed estimator when there is change.

In longitudinal surveys, sample subjects are observed over several time points. This feature typically leads to dependent observations on the same subject, in addition to the customary correlations across subjects induced by the sample design. Much research in the literature has focussed on modeling the marginal mean of a response as a function of covariates. Liang and Zeger (1986) used generalized estimating equations (GEE), requiring only correct specification of the marginal mean, and obtained standard errors of regression parameter estimates and associated Wald tests, assuming a "working" correlation structure for the repeated measurements on a sample subject. Rotnitzky and Jewell (1990) developed quasi-score tests and Rao-Scott adjustments to "working" quasi-score tests under marginal models. These methods are asymptotically robust to misspecification of the within-subject correlation structure, but assume independence of sample subjects which is not satisfied for complex longitudinal survey data based on stratified multi-stage sampling. We proposed asymptotically valid Wald and quasi-score tests for longitudinal survey data, using the Taylor Linearization and jackknife methods. Alternative tests, based on Rao-Scott adjustments to naive tests that ignore survey design features and on Bonferroni-t, are also developed. These tests are particularly useful when the effective degrees of freedom, usually taken as the total number of sample primary units (clusters) minus the number of strata, is small.