Hybrid Probabilistic-Snowball Sampling

TL;DR

This paper introduces Hybrid Probabilistic-Snowball Sampling Designs (HPSSD), a method combining probabilistic sampling with snowball techniques to reduce bias in social network surveys, supported by simulation results.

Contribution

The paper develops and tests HPSSD, a novel sampling method that integrates random oversampling with snowball sampling to mitigate bias in social network studies.

Findings

HPSSD does not significantly improve reliability in already representative samples.

When homophily is low, HPSSD's unadjusted mean slightly outperforms simple random sampling.

De-biasing HPSSD estimates enhances their accuracy.

Abstract

Snowball sampling is the common name for sampling designs on human populations where respondents are requested to share the questionnaire among their social ties. With some exceptions, estimates from snowball samplings are considered biased. However, the magnitude of the bias is influenced by a combination of elements of the sampling design and features of the target population. Hybrid Probabilistic-Snowball Sampling Designs (HPSSD) aims to reduce the main source of bias in the snowball sample through randomly oversampling the first stage 0 of the snowball. To check the behaviour of HPSSD for applications, we developed an algorithm that, by grafting the edges of a stochastic blockmodel into a graph of cliques, simulates an assortative network of tobacco smokers. Different outcomes of the HPSSD operations are simulated, too. Inference on 8,000 runs of the simulation leads to think that HPSSD does not improve reliability of samples that are already representative. But if homophily in the population is sufficiently low, even the unadjusted sample mean of HPSSD has a slightly better performance than a random, but undersized, sampling. De-biasing the estimates of HPSSD shows improvement in the performance, so an adjusted HPSSD estimator is a desirable development.