Inference from multivariate differential recruitment in respondent-driven sampling data

1. Problem Statement

Respondent-Driven Sampling (RDS) is a chain-referral method used to study hidden or hard-to-reach populations (e.g., drug users, immigrants) where no sampling frame exists. Standard RDS inference relies on the assumption of random recruitment, where participants select peers from their network with equal probability.

However, empirical evidence suggests Differential Recruitment (DR) is common: participants often recruit based on preferences (e.g., age, gender, relationship type).

• Limitation of Existing Methods: Previous work addressed DR but only for a single categorical variable (univariate DR). In reality, recruitment is likely driven by multiple simultaneous factors, including both categorical (e.g., gender, education) and continuous (e.g., age difference, frequency of contact) covariates.
• The Gap: Ignoring these multivariate sources of bias leads to substantial errors in prevalence estimates. There is a lack of a statistical framework that models recruitment probabilities as a function of multiple, simultaneous covariates.

2. Methodology

The authors propose a Multivariate Differential Recruitment (MDR) framework that extends the Markov Chain (MC) modeling of RDS.

A. The MDR Sampling Model

The RDS process is modeled as a Markov Chain where transition probabilities depend on a vector of covariates $x_{ij}$ associated with the recruiter $i$ and the potential recruit $j$.

• Covariates: The vector $x_{ij}$ includes:
  • Node-specific attributes ($r_j$): e.g., age, gender of the recruit.
  • Tie-specific attributes ($w_{ij}$): e.g., absolute age difference, relationship strength.
• Transition Probability: The probability that $i$ recruits $j$ is modeled using a log-linear form:
  $$P^{MDR}_{ij} = \frac{y_{ij} \exp(x_{ij}^\top \beta)}{\sum_{l=1}^N y_{il} \exp(x_{il}^\top \beta)}$$
  Where $y_{ij}$ is the adjacency matrix, and $\beta$ is a vector of unknown coefficients weighting the influence of each covariate.

B. Stationary Distribution and Estimation

• Derivation: The authors prove that under specific conditions, this MC admits a unique stationary distribution $\pi^{MDR}_i$. This distribution serves as the basis for the sampling weights.
  $$\pi^{MDR}_i \propto \sum_{j=1}^N y_{ij} \exp(x_{ij}^\top \beta + r_i^\top \alpha)$$
  (Where $\alpha$ corresponds to the coefficients of node-specific covariates).
• Parameter Estimation: The coefficients $\beta$ are estimated via Maximum Likelihood Estimation (MLE) using the observed recruitment structure (excluding seeds).
• Prevalence Estimators: The authors extend two existing estimators to the MDR framework:
  1. Extended Volz-Heckathorn ($\hat{\mu}^{II}_{MDR}$): A Hájek-type estimator using the MDR-derived weights.
  2. Extended Lu ($\hat{\mu}^{ego}_{MDR}$): An ego-network based estimator that incorporates MDR weights into the calculation of cross-group tie proportions.

C. Variance Estimation

The authors adapt the Neighborhood Bootstrap (NB) method (Yauck et al., 2022) for the MDR context.

• Modification: The original NB does not guarantee a fixed sample size. The authors introduce an algorithm (Algorithm 1) that iteratively adds or removes recruiters and their immediate recruits to ensure every bootstrap replicate has a sample size exactly equal to the original sample size ($n$).
• Confidence Intervals: They construct intervals using the normal approximation (standard normal quantiles) to ensure fair comparison across different bootstrap methods.

3. Key Contributions

1. Multivariate Framework: Introduction of the first RDS model capable of handling multiple simultaneous covariates (both continuous and categorical) influencing recruitment.
2. Theoretical Extension: Derivation of the stationary distribution for an MC with covariate-dependent transitions and the extension of standard prevalence estimators ($\hat{\mu}^{II}$ and $\hat{\mu}^{ego}$) to this setting.
3. Improved Variance Estimation: A modified Neighborhood Bootstrap algorithm that maintains fixed sample sizes, crucial for accurate variance estimation in complex MDR scenarios.
4. Empirical Validation: Application of the method to real-world data involving Venezuelan immigrants in Chile, demonstrating practical utility even with partially observed alter data.

4. Results

Simulation Study

The authors simulated 15 networks across 9 scenarios varying in Homophily (None, Moderate, High) and MDR Strength (None, Moderate, High).

• Bias:
  • In the absence of MDR, all estimators were approximately unbiased.
  • As MDR increased, standard estimators ($\hat{\mu}^{II}_{VH}$) and univariate DR estimators ($\hat{\mu}^{II}_{DR}$) exhibited significant bias.
  • The proposed MDR estimators ($\hat{\mu}^{II}_{MDR}$ and $\hat{\mu}^{ego}_{MDR}$) remained nearly unbiased even under high homophily and high MDR.
• Efficiency (RMSE):
  • $\hat{\mu}^{ego}_{MDR}$ consistently achieved the lowest Root Mean Square Error (RMSE) in scenarios with moderate to high MDR, outperforming all other estimators.
  • Ego-type estimators generally showed lower variability than II-type estimators.
• Coverage:
  • Standard estimators suffered from poor confidence interval coverage (under-coverage) when MDR was present due to uncorrected bias.
  • MDR estimators maintained coverage rates close to or above the nominal 95% level.
  • Note: $\hat{\mu}^{II}_{MDR}$ tended to be conservative (over-coverage) in simple scenarios due to variance overestimation by the bootstrap, while $\hat{\mu}^{ego}_{MDR}$ showed slightly lower coverage in complex scenarios but remained the most robust overall.

Real-World Application

The method was applied to a 2022 RDS study of Venezuelan immigrants in Santiago, Chile, to estimate the proportion of males.

• Original Data: With low-to-moderate MDR, estimates across methods were similar, though $\hat{\mu}^{ego}_{MDR}$ had the lowest standard error.
• Sensitivity Analysis: The authors created a modified dataset with stronger MDR effects (simulating a scenario where males were disproportionately recruited).
  • Under strong MDR, the standard estimators and univariate DR estimators failed to adjust correctly, leading to biased prevalence estimates.
  • The $\hat{\mu}^{ego}_{MDR}$ estimator successfully adjusted for the multivariate bias, yielding the most precise estimate with the lowest standard error.

5. Significance and Conclusion

This paper significantly advances the statistical inference of RDS data by moving beyond the restrictive assumption of random recruitment and the limitations of univariate differential recruitment.

• Practical Impact: It provides researchers with a tool to reduce non-sampling errors in hidden population studies by explicitly modeling how multiple factors (age, gender, relationship dynamics) drive recruitment.
• Robustness: The proposed $\hat{\mu}^{ego}_{MDR}$ estimator is shown to be the most reliable in the presence of complex recruitment biases and network homophily.
• Future Directions: The authors note that the method's performance depends on the availability of covariate data for "alters" (recruits). In practice, this data is often aggregated or missing, requiring imputation. Future work should focus on developing methods to properly account for the uncertainty introduced by reconstructing alter-level data.

In summary, the MDR framework offers a more realistic and statistically rigorous approach to estimating population prevalence in hard-to-reach groups, particularly when recruitment behavior is driven by complex, multivariate social preferences.