Survey Methodology

This brief tribute reviews Chris Skinner’s main scientific contributions.

I provide an overview of the evolution of Statistical Disclosure Control (SDC) research over the last decades and how it has evolved to handle the data revolution with more formal definitions of privacy. I emphasize the many contributions by Chris Skinner in the research areas of SDC. I review his seminal research, starting in the 1990’s with his work on the release of UK Census sample microdata. This led to a wide-range of research on measuring the risk of re-identification in survey microdata through probabilistic models. I also focus on other aspects of Chris’ research in SDC. Chris was the recipient of the 2019 Waksberg Award and sadly never got a chance to present hisWaksberg Lecture at the Statistics Canada International Methodology Symposium. This paper follows the outline that Chris had prepared in preparation for that lecture.

My comments consist of three components: (1) A brief account of my professional association with Chris Skinner. (2) Observations on Skinner’s contributions to statistical disclosure control, (3) Some comments on making inferences from masked survey data.

Weight smoothing is a useful technique in improving the efficiency of design-based estimators at the risk of bias due to model misspecification. As an extension of the work of Kim and Skinner (2013), we propose using weight smoothing to construct the conditional likelihood for efficient analytic inference under informative sampling. The Beta prime distribution can be used to build a parameter model for weights in the sample. A score test is developed to test for model misspecification in the weight model. A pretest estimator using the score test can be developed naturally. The pretest estimator is nearly unbiased and can be more efficient than the design-based estimator when the weight model is correctly specified, or the original weights are highly variable. A limited simulation study is presented to investigate the performance of the proposed methods.

The Dutch Health Survey (DHS), conducted by Statistics Netherlands, is designed to produce reliable direct estimates at an annual frequency. Data collection is based on a combination of web interviewing and face-to-face interviewing. Due to lockdown measures during the Covid-19 pandemic there was no or less face-to-face interviewing possible, which resulted in a sudden change in measurement and selection effects in the survey outcomes. Furthermore, the production of annual data about the effect of Covid-19 on health-related themes with a delay of about one year compromises the relevance of the survey. The sample size of the DHS does not allow the production of figures for shorter reference periods. Both issues are solved by developing a bivariate structural time series model (STM) to estimate quarterly figures for eight key health indicators. This model combines two series of direct estimates, a series based on complete response and a series based on web response only and provides model-based predictions for the indicators that are corrected for the loss of face-to-face interviews during the lockdown periods. The model is also used as a form of small area estimation and borrows sample information observed in previous reference periods. In this way timely and relevant statistics describing the effects of the corona crisis on the development of Dutch health are published. In this paper the method based on the bivariate STM is compared with two alternative methods. The first one uses a univariate STM where no correction for the lack of face-to-face observation is applied to the estimates. The second one uses a univariate STM that also contains an intervention variable that models the effect of the loss of face-to-face response during the lockdown.

To improve the precision of inferences and reduce costs there is considerable interest in combining data from several sources such as sample surveys and administrative data. Appropriate methodology is required to ensure satisfactory inferences since the target populations and methods for acquiring data may be quite different. To provide improved inferences we use methodology that has a more general structure than the ones in current practice. We start with the case where the analyst has only summary statistics from each of the sources. In our primary method, uncertain pooling, it is assumed that the analyst can regard one source, survey r, as the single best choice for inference. This method starts with the data from survey r and adds data from those other sources that are shown to form clusters that include survey r. We also consider Dirichlet process mixtures, one of the most popular nonparametric Bayesian methods. We use analytical expressions and the results from numerical studies to show properties of the methodology.

We consider regression analysis in the context of data integration. To combine partial information from external sources, we employ the idea of model calibration which introduces a “working” reduced model based on the observed covariates. The working reduced model is not necessarily correctly specified but can be a useful device to incorporate the partial information from the external data. The actual implementation is based on a novel application of the information projection and model calibration weighting. The proposed method is particularly attractive for combining information from several sources with different missing patterns. The proposed method is applied to a real data example combining survey data from Korean National Health and Nutrition Examination Survey and big data from National Health Insurance Sharing Service in Korea.

Recent work in survey domain estimation allows for estimation of population domain means under a priori assumptions expressed in terms of linear inequality constraints. For example, it might be known that the population means are non-decreasing along ordered domains. Imposing the constraints has been shown to provide estimators with smaller variance and tighter confidence intervals. In this paper we consider a formal test of the null hypothesis that all the constraints are binding, versus the alternative that at least one constraint is non-binding. The test of constant versus increasing domain means is a special case. The power of the test is substantially better than the test with the same null hypothesis and an unconstrained alternative. The new test is used with data from the National Survey of College Graduates, to show that salaries are positively related to the subject’s father’s educational level, across fields of study and over several years of cohorts.

The definition of statistical units is a recurring issue in the domain of sample surveys. Indeed, not all the populations surveyed have a readily available sampling frame. For some populations, the sampled units are distinct from the observation units and producing estimates on the population of interest raises complex questions, which can be addressed by using the weight share method (Deville and Lavallée, 2006). However, the two populations considered in this approach are discrete. In some fields of study, the sampled population is continuous: this is for example the case of forest inventories for which, frequently, the trees surveyed are those located on plots of which the centers are points randomly drawn in a given area. The production of statistical estimates from the sample of trees surveyed poses methodological difficulties, as do the associated variance calculations. The purpose of this paper is to generalize the weight share method to the continuous (sampled population) ‒ discrete (surveyed population) case, from the extension proposed by Cordy (1993) of the Horvitz-Thompson estimator for drawing points carried out in a continuous universe.

Precise and unbiased estimates of response propensities (RPs) play a decisive role in the monitoring, analysis, and adaptation of data collection. In a fixed survey climate, those parameters are stable and their estimates ultimately converge when sufficient historic data is collected. In survey practice, however, response rates gradually vary in time. Understanding time-dependent variation in predicting response rates is key when adapting survey design. This paper illuminates time-dependent variation in response rates through multi-level time-series models. Reliable predictions can be generated by learning from historic time series and updating with new data in a Bayesian framework. As an illustrative case study, we focus on Web response rates in the Dutch Health Survey from 2014 to 2019.

In this paper, with and without-replacement versions of adaptive proportional to size sampling are presented. Unbiased estimators are developed for these methods and their properties are studied. In the two versions, the drawing probabilities are adapted during the sampling process based on the observations already selected. To this end, in the version with-replacement, after each draw and observation of the variable of interest, the vector of the auxiliary variable will be updated using the observed values of the variable of interest to approximate the exact selection probability proportional to size. For the without-replacement version, first, using an initial sample, we model the relationship between the variable of interest and the auxiliary variable. Then, utilizing this relationship, we estimate the unknown (unobserved) population units. Finally, on these estimated population units, we select a new sample proportional to size without-replacement. These approaches can significantly improve the efficiency of designs not only in the case of a positive linear relationship, but also in the case of a non-linear or negative linear relationship between the variables. We investigate the efficiencies of the designs through simulations and real case studies on medicinal flowers, social and economic data.