Survey Methodology

Sample surveys, as a tool for policy development and evaluation and for scientific, social and economic research, have been employed for over a century. In that time, they have primarily served as tools for collecting data for enumerative purposes. Estimation of these characteristics has been typically based on weighting and repeated sampling, or design-based, inference. However, sample data have also been used for modelling the unobservable processes that gave rise to the finite population data. This type of use has been termed analytic, and often involves integrating the sample data with data from secondary sources.

Alternative approaches to inference in these situations, drawing inspiration from mainstream statistical modelling, have been strongly promoted. The principal focus of these alternatives has been on allowing for informative sampling. Modern survey sampling, though, is more focussed on situations where the sample data are in fact part of a more complex set of data sources all carrying relevant information about the process of interest. When an efficient modelling method such as maximum likelihood is preferred, the issue becomes one of how it should be modified to account for both complex sampling designs and multiple data sources. Here application of the Missing Information Principle provides a clear way forward.

In this paper I review how this principle has been applied to resolve so-called “messy” data analysis issues in sampling. I also discuss a scenario that is a consequence of the rapid growth in auxiliary data sources for survey data analysis. This is where sampled records from one accessible source or register are linked to records from another less accessible source, with values of the response variable of interest drawn from this second source, and where a key output is small area estimates for the response variable for domains defined on the first source.

Jean-Claude Deville, who passed away in October 2021, was one of the most influential researchers in the field of survey statistics over the past 40 years. This article traces some of his contributions that have had a profound impact on both survey theory and practice. This article will cover the topics of balanced sampling using the cube method, calibration, the weight-sharing method, the development of variance expressions of complex estimators using influence function and quota sampling.

In this discussion, I will present some additional aspects of three major areas of survey theory developed or studied by Jean‑Claude Deville: calibration, balanced sampling and the generalized weight-share method.

This article discusses and provides comments on the Ardilly, Haziza, Lavallée and Tillé’s summary presentation of Jean-Claude Deville’s work on survey theory. It sheds light on the context, applications and uses of his findings, and shows how these have become engrained in the role of statisticians, in which Jean-Claude was a trailblazer. It also discusses other aspects of his career and his creative inventions.

Many things have been written about Jean-Claude Deville in tributes from the statistical community (see Tillé, 2022a; Tillé, 2022b; Christine, 2022; Ardilly, 2022; and Matei, 2022) and from the École nationale de la statistique et de l’administration économique (ENSAE) and the Société française de statistique. Pascal Ardilly, David Haziza, Pierre Lavallée and Yves Tillé provide an in-depth look at Jean-Claude Deville’s contributions to survey theory. To pay tribute to him, I would like to discuss Jean-Claude Deville’s contribution to the more day-to-day application of methodology for all the statisticians at the Institut national de la statistique et des études économiques (INSEE) and at the public statistics service. To do this, I will use my work experience, and particularly the four years (1992 to 1996) I spent working with him in the Statistical Methods Unit and the discussions we had thereafter, especially in the 2000s on the rolling census.

Jean-Claude Deville is one of the most prominent researcher in survey sampling theory and practice. His research on balanced sampling, indirect sampling and calibration in particular is internationally recognized and widely used in official statistics. He was also a pioneer in the field of functional data analysis. This discussion gives us the opportunity to recognize the immense work he has accomplished, and to pay tribute to him. In the first part of this article, we recall briefly his contribution to the functional principal analysis. We also detail some recent extension of his work at the intersection of the fields of functional data analysis and survey sampling. In the second part of this paper, we present some extension of Jean-Claude’s work in indirect sampling. These extensions are motivated by concrete applications and illustrate Jean-Claude’s influence on our work as researchers.

In recent decades, many different uses of auxiliary information have enriched survey sampling theory and practice. Jean-Claude Deville contributed significantly to this progress. My comments trace some of the steps on the way to one important theory for the use of auxiliary information: Estimation by calibration.

The article considers sampling designs for populations that can be represented as a $N\times M$ matrix. For instance when investigating tourist activities, the rows could be locations visited by tourists and the columns days in the tourist season. The goal is to sample cells $\mathrm{<mo\; stretchy="false">(</mo>}i\mathrm{,}j\mathrm{<mo\; stretchy="false">)</mo>}$ of the matrix when the number of selections within each row and each column is fixed a priori. The $i^{\text{th}}$ row sample size represents the number of selected cells within row $i;$ the $j^{\text{th}}$ column sample size is the number of selected cells within column $j.$ A matrix sampling design gives an $N\times M$ matrix of sample indicators, with entry 1 at position $\mathrm{<mo\; stretchy="false">(</mo>}i\mathrm{,}j\mathrm{<mo\; stretchy="false">)</mo>}$ if cell $\mathrm{<mo\; stretchy="false">(</mo>}i\mathrm{,}j\mathrm{<mo\; stretchy="false">)</mo>}$ is sampled and 0 otherwise. The first matrix sampling design investigated has one level of sampling, row and column sample sizes are set in advance: the row sample sizes can vary while the column sample sizes are all equal. The fixed margins can be seen as balancing constraints and algorithms available for selecting such samples are reviewed. A new estimator for the variance of the Horvitz-Thompson estimator for the mean of survey variable $y$ is then presented. Several levels of sampling might be necessary to account for all the constraints; this involves multi-level matrix sampling designs that are also investigated.

Sample coordination methods aim to increase (in positive coordination) or decrease (in negative coordination) the size of the overlap between samples. The samples considered can be from different occasions of a repeated survey and/or from different surveys covering a common population. Negative coordination is used to control the response burden in a given period, because some units do not respond to survey questionnaires if they are selected in many samples. Usually, methods for sample coordination do not take into account any measure of the response burden that a unit has already expended in responding to previous surveys. We introduce such a measure into a new method by adapting a spatially balanced sampling scheme, based on a generalization of Poisson sampling, together with a negative coordination method. The goal is to create a double control of the burden for these units: once by using a measure of burden during the sampling process and once by using a negative coordination method. We evaluate the approach using Monte-Carlo simulation and investigate its use for controlling for selection “hot-spots” in business surveys in Statistics Netherlands.

In this paper, we investigate how a big non-probability database can be used to improve estimates of finite population totals from a small probability sample through data integration techniques. In the situation where the study variable is observed in both data sources, Kim and Tam (2021) proposed two design-consistent estimators that can be justified through dual frame survey theory. First, we provide conditions ensuring that these estimators are more efficient than the Horvitz-Thompson estimator when the probability sample is selected using either Poisson sampling or simple random sampling without replacement. Then, we study the class of QR predictors, introduced by Särndal and Wright (1984), to handle the less common case where the non-probability database contains no study variable but auxiliary variables. We also require that the non-probability database is large and can be linked to the probability sample. We provide conditions ensuring that the QR predictor is asymptotically design-unbiased. We derive its asymptotic design variance and provide a consistent design-based variance estimator. We compare the design properties of different predictors, in the class of QR predictors, through a simulation study. This class includes a model-based predictor, a model-assisted estimator and a cosmetic estimator. In our simulation setups, the cosmetic estimator performed slightly better than the model-assisted estimator. These findings are confirmed by an application to La Poste data, which also illustrates that the properties of the cosmetic estimator are preserved irrespective of the observed non-probability sample.

In this article, we use a slightly simplified version of the method by Fickus, Mixon and Poteet (2013) to define a flexible parameterization of the kernels of determinantal sampling designs with fixed first-order inclusion probabilities. For specific values of the multidimensional parameter, we get back to a matrix from the family $P^{\Pi }$ from Loonis and Mary (2019). We speculate that, among the determinantal designs with fixed inclusion probabilities, the minimum variance of the Horvitz and Thompson estimator (1952) of a variable of interest is expressed relative to $P^{\Pi }.$ We provide experimental R programs that facilitate the appropriation of various concepts presented in the article, some of which are described as non-trivial by Fickus et al. (2013). A longer version of this article, including proofs and a more detailed presentation of the determinantal designs, is also available.

Conformal prediction is an assumption-lean approach to generating distribution-free prediction intervals or sets, for nearly arbitrary predictive models, with guaranteed finite-sample coverage. Conformal methods are an active research topic in statistics and machine learning, but only recently have they been extended to non-exchangeable data. In this paper, we invite survey methodologists to begin using and contributing to conformal methods. We introduce how conformal prediction can be applied to data from several common complex sample survey designs, under a framework of design-based inference for a finite population, and we point out gaps where survey methodologists could fruitfully apply their expertise. Our simulations empirically bear out the theoretical guarantees of finite-sample coverage, and our real-data example demonstrates how conformal prediction can be applied to complex sample survey data in practice.

Survey researchers are increasingly turning to multimode data collection to deal with declines in survey response rates and increasing costs. An efficient approach offers the less costly modes (e.g., web) followed with a more expensive mode for a subsample of the units (e.g., households) within each primary sampling unit (PSU). We present two alternatives to this traditional design. One alternative subsamples PSUs rather than units to constrain costs. The second is a hybrid design that includes a clustered (two-stage) sample and an independent, unclustered sample. Using a simulation, we demonstrate the hybrid design has considerable advantages.

Population undercoverage is one of the main hurdles faced by statistical analysis with non-probability survey samples. We discuss two typical scenarios of undercoverage, namely, stochastic undercoverage and deterministic undercoverage. We argue that existing estimation methods under the positivity assumption on the propensity scores (i.e., the participation probabilities) can be directly applied to handle the scenario of stochastic undercoverage. We explore strategies for mitigating biases in estimating the mean of the target population under deterministic undercoverage. In particular, we examine a split population approach based on a convex hull formulation, and construct estimators with reduced biases. A doubly robust estimator can be constructed if a followup subsample of the reference probability survey with measurements on the study variable becomes feasible. Performances of six competing estimators are investigated through a simulation study and issues which require further investigation are briefly discussed.

We present a novel methodology to benchmark county-level estimates of crop area totals to a preset state total subject to inequality constraints and random variances in the Fay-Herriot model. For planted area of the National Agricultural Statistics Service (NASS), an agency of the United States Department of Agriculture (USDA), it is necessary to incorporate the constraint that the estimated totals, derived from survey and other auxiliary data, are no smaller than administrative planted area totals prerecorded by other USDA agencies except NASS. These administrative totals are treated as fixed and known, and this additional coherence requirement adds to the complexity of benchmarking the county-level estimates. A fully Bayesian analysis of the Fay-Herriot model offers an appealing way to incorporate the inequality and benchmarking constraints, and to quantify the resulting uncertainties, but sampling from the posterior densities involves difficult integration, and reasonable approxi­mations must be made. First, we describe a single-shrinkage model, shrinking the means while the variances are assumed known. Second, we extend this model to accommodate double shrinkage, borrowing strength across means and variances. This extended model has two sources of extra variation, but because we are shrinking both means and variances, it is expected that this second model should perform better in terms of goodness of fit (reliability) and possibly precision. The computations are challenging for both models, which are applied to simulated data sets with properties resembling the Illinois corn crop.

We investigate small area prediction of general parameters based on two models for unit-level counts. We construct predictors of parameters, such as quartiles, that may be nonlinear functions of the model response variable. We first develop a procedure to construct empirical best predictors and mean square error estimators of general parameters under a unit-level gamma-Poisson model. We then use a sampling importance resampling algorithm to develop predictors for a generalized linear mixed model (GLMM) with a Poisson response distribution. We compare the two models through simulation and an analysis of data from the Iowa Seat-Belt Use Survey.

Being able to quantify the accuracy (bias, variance) of published output is crucial in official statistics. Output in official statistics is nearly always divided into subpopulations according to some classification variable, such as mean income by categories of educational level. Such output is also referred to as domain statistics. In the current paper, we limit ourselves to binary classification variables. In practice, misclassifications occur and these contribute to the bias and variance of domain statistics. Existing analytical and numerical methods to estimate this effect have two disadvantages. The first disadvantage is that they require that the misclassification probabilities are known beforehand and the second is that the bias and variance estimates are biased themselves. In the current paper we present a new method, a Gaussian mixture model estimated by an Expectation-Maximisation (EM) algorithm combined with a bootstrap, referred to as the EM bootstrap method. This new method does not require that the misclassification probabilities are known beforehand, although it is more efficient when a small audit sample is used that yields a starting value for the misclassification probabilities in the EM algorithm. We compared the performance of the new method with currently available numerical methods: the bootstrap method and the SIMEX method. Previous research has shown that for non-linear parameters the bootstrap outperforms the analytical expressions. For nearly all conditions tested, the bias and variance estimates that are obtained by the EM bootstrap method are closer to their true values than those obtained by the bootstrap and SIMEX methods. We end this paper by discussing the results and possible future extensions of the method.

When a Medicare healthcare provider is suspected of billing abuse, a population of payments $X$ made to that provider over a fixed timeframe is isolated. A certified medical reviewer, in a time-consuming process, can determine the overpayment $Y\mathrm{<mo>=</mo>}X-$ (amount justified by the evidence) associated with each payment. Typically, there are too many payments in the population to examine each with care, so a probability sample is selected. The sample overpayments are then used to calculate a 90% lower confidence bound for the total population overpayment. This bound is the amount demanded for recovery from the provider. Unfortunately, classical methods for calculating this bound sometimes fail to provide the 90% confidence level, especially when using a stratified sample.

In this paper, 166 redacted samples from Medicare integrity investigations are displayed and described, along with 156 associated payment populations. The 7,588 examined $\left(Y\mathrm{,}X\right)$ sample pairs show (1) Medicare audits have high error rates: more than 76% of these payments were considered to have been paid in error; and (2) the patterns in these samples support an “All-or-Nothing” mixture model for $\left(Y\mathrm{,}X\right)$ previously defined in the literature. Model-based Monte Carlo testing procedures for Medicare sampling plans are discussed, as well as stratification methods based on anticipated model moments. In terms of viability (achieving the 90% confidence level) a new stratification method defined here is competitive with the best of the many existing methods tested and seems less sensitive to choice of operating parameters. In terms of overpayment recovery (equivalent to precision) the new method is also comparable to the best of the many existing methods tested. Unfortunately, no stratification algorithm tested was ever viable for more than about half of the 104 test populations.