Dealing with nonresponse in survey sampling: a latent modeling approach

TL;DR

This paper introduces a latent modeling approach to address both unit and item nonresponse in surveys, leveraging latent trait models to estimate response probabilities and reduce bias, especially for sensitive questions.

Contribution

It proposes a novel method combining latent trait models with logistic regression to handle non-ignorable nonresponse without requiring auxiliary data.

Findings

Method effectively reduces nonresponse bias in simulations.

Latent variable approach performs well with sensitive survey items.

Theoretical properties of estimators are established.

Abstract

Nonresponse is present in almost all surveys and can severely bias estimates. It is usually distinguished between unit and item nonresponse: in the former, we completely fail to have information from a unit selected in the sample, while in the latter, we observe only part of the information on the selected unit. Unit nonresponse is usually dealt with by reweighting: each unit selected in the sample has associated a sampling weight and an unknown response probability; the initial sampling weight is multiplied by the inverse of estimated response probability. Item nonresponse is usually dealt with by imputation. By noting that for a particular survey variable, we just have observed and unobserved values, in this work we exploit the connection between unit and item nonresponse. In particular, we assume that the factors that drive unit response are the same as those that drive item response on selected variables of interest. Response probabilities are then estimated by using a logistic regression with a latent covariate that measures such will to respond and that can explain part of the unknown behavior of a unit to participate in the survey. The latent covariate is estimated using latent trait models. Such approach is particularly relevant for sensitive items and, therefore, can handle non-ignorable nonresponse. Auxiliary information known for both respondents and nonrespondents can be included either in the latent variable model or in the logistic model. The approach can be also used when auxiliary information is not available, and we focus here on this case. The theoretical properties of the proposed estimators are sketched and simulations studies are conducted to illustrate their finite size sample performance.