The R package NonProbEst for estimation in non-probability surveys

Abstract

Different inference procedures are proposed in the literature to correct selection bias that might be introduced with non-random sampling mechanisms. The R package NonProbEst enables the estimation of parameters using some of these techniques to correct selection bias in non-probability surveys. The mean and the total of the target variable are estimated using Propensity Score Adjustment, calibration, statistical matching, model-based, model-assisted and model-calibratated techniques. Confidence intervals can also obtained for each method. Machine learning algorithms can be used for estimating the propensities or for predicting the unknown values of the target variable for the non-sampled units. Variance of a given estimator is performed by two different Leave-One-Out jackknife procedures. The functionality of the package is illustrated with example data sets.

Why weight? Analytic approaches for large-scale population neuroscience data

“…Doubly robust inference when combining probability and non-probability samples with high dimensional data. Rueda et al, (2020) . The R package NonProbEst for estimation in non-probability surveys.…”

Estimation by kernel weighting of parameters related to employment in the confinement period

“…In order to eliminate the volunteer bias in nonprobability surveys, we must have available auxiliary information that is accurate and closely related to the topic under study. Depending on the type of auxiliary information available, we distinguish different types of bias reduction techniques (Rueda et al, 2020), although in our study we will focus on one, the kernel weighting method, due to its recent appearance and excellent results. This technique is included in the group of those that need a probabilistic reference sample in order to be carried out, from which we only need to know its auxiliary variables.…”

Use of machine learning techniques in non-probabilistic samples

“…Normally the results we obtain from this kind of samples present different types of biases, especially the one known as selection bias, which appears if there is a significant difference between the individuals in our sample and those not sampled. To correct this type of bias there are several techniques, which depend on the type of auxiliary information available (Rueda et al, 2020). If we have a reference probability sample 𝑠 𝑟 , of which we only know the same vector of auxiliary variables as in 𝑠 𝑣 , we can apply the technique known as statistical matching, based on superpopulation models.…”