Recruiters as Witnesses: Peer-Informed Multiple Imputation for Longitudinal Respondent-Driven Cohorts

Studies of hidden populations, such as people who inject drugs or sex workers, find participants by chain referral: a few initial participants recruit their peers, those peers recruit theirs, and so on. Nearly all such studies take a single survey per person and score risk once. Hospital research has shown how much that costs. Risk scored at a single time point loses predictive value quickly, and people who look low-risk at baseline are often the ones who deteriorate. The obvious remedy is to follow chain-referred participants over time. Doing so creates a new problem. The people who drop out of follow-up are disproportionately the people whose risk escalated, through overdose, incarceration, or loss of housing. Their missing data is therefore "missing not at random" (MNAR): the chance a value is missing depends on the value itself. Standard multiple imputation cannot correct this. Multiple imputation fills each gap with several plausible random draws and pools the results.

This proposal rests on one structural fact about chain-referral cohorts. Every dropout has a recruiter, and usually some recruits, who were chosen because they share the dropout's social world. Their observed trajectories carry information about the dropout's unobserved state. We claim that conditioning the imputation on recruitment-tree neighbours can turn missingness that is uncorrectable from the person's own data into missingness that is approximately correctable from peer data. We also claim this works only when the imputation is pooled with a variance estimate that respects network dependence. We propose to build this peer-informed imputation, state when it is valid, and test it against strong simpler alternatives on simulated recruitment over real social networks. We state in advance that the simplest version, ordinary chained-equation imputation with recruiter columns added, may turn out to be enough.

Core Claim

Let $Y_{it}$ be participant $i$'s risk state at follow-up wave $t$. Let $R_{it}$ indicate whether it was observed. Let $X_i$ be $i$'s own observed history. Let $Y_{N(i)}$ be the observed data of $i$'s recruiter and recruits. Standard imputation assumes missingness at random (MAR), $R_{it} \perp Y_{it} \mid X_i$. That assumption fails when dropout is driven by the person's own escalating risk. We instead posit peer-conditional MAR:

$$R_{it} \perp Y_{it} \mid X_i,\; Y_{N(i),t}.$$

This holds when the drivers of dropout (a local drug-market shock, a police sweep, a shared housing crisis) are shared along the recruitment ties that homophily creates. Here homophily means recruiting people similar to oneself. The condition cannot be checked from the data alone. We therefore pair it with a delta-adjustment sensitivity analysis that shifts imputed values by a chosen amount $\delta$ to represent residual own-risk dropout. We also give a simulation map of how much homophily is needed for peer conditioning to remove most of the bias.

The second half of the claim concerns uncertainty. Rubin's rules combine a within-imputation variance $W$ and a between-imputation variance $B$ into a total $T = W + (1 + 1/m)B$ over $m$ imputations. In chain-referral data, $W$ computed as if participants were independent is too small. Generalized least squares can overcome the critical threshold in respondent-driven sampling shows that this dependence can break the usual $1/n$ variance decay. Peer-informed imputation adds a further layer of correlation, because imputed values now copy information from neighbours. We therefore replace $W$ with a network-aware within variance $W_{\text{net}}$:

$$T_{\text{net}} = W_{\text{net}} + \left(1 + \tfrac{1}{m}\right) B.$$

$W_{\text{net}}$ comes either from the feasible generalized least squares (fGLS) covariance based on a degree-corrected stochastic blockmodel, as in that paper, or from resampling whole recruitment branches before imputing. The contribution is the combination of peer-conditional identification and dependence-aware pooling. Neither exists in the file.

Position Relative to Prior Work

Two hospital studies motivate the dynamic half. Optimal intensive care outcome prediction over time using machine learning shows that daily cumulative models beat the static APACHE-II score (AUC 0.836 for APACHE-II versus 0.883 on Day 1 and 0.895 on Day 2). Beyond Thresholds uses hidden Markov models, which infer a sequence of unobserved discrete states from noisy measurements, on intraoperative vitals. It raises AUPRC for prolonged ICU stay from 0.244 to 0.339, with the largest AUROC gain (+0.223) in the static lowest-risk tertile. Both work on hospital cohorts with administrative sampling. The first imputes offline with chained equations and notes, without solving it, that dropout through discharge or death biases the analysis. We carry their "hidden vulnerable" question into a setting where the vulnerable are also the ones who vanish.

Two chain-referral studies motivate the application: Needle acquisition patterns among rural PWID in Puerto Rico, where each additional injection partner raises the odds of street-found syringes by 28.2%, and HIV PrEP interest among female sex workers in Guangxi. Both reach insulated populations, and both score risk once, cross-sectionally. We keep their recruitment design and add follow-up waves.

The imputation and network papers supply methods and also constrain the design. SHT-GNN imputes longitudinal biomarkers as link prediction on a subject–variable graph. It is the strongest longitudinal imputer in the file, but it has no subject-to-subject edges, so it cannot use peers. RefiDiff handles MNAR tabular data with a diffusion model but treats rows as independent. Without peer structure, its MNAR handling depends on functional-form assumptions, not on auxiliary information. RENDER reconstructs unobserved ties from recruitment timing (TPR 0.769, FPR 0.106 on Project 90). We reuse it to enlarge each person's neighbour set beyond the coupon tree, which RENDER itself never uses for imputation. MatchThem combines weights with multiply imputed data but assumes ordinary survey dependence. Asymptotic Seed Bias in Respondent-driven Sampling constrains us: short chains that start from non-random seeds do not forget their starting point. Peer information from shallow chains may therefore be biased toward seed characteristics, and our simulations vary chain depth to measure this.

Research Program

We will build peer-informed imputation in three tiers of increasing complexity, so that the evidence decides whether complexity is warranted.

The first tier is chained-equation imputation (MICE). It regresses each incomplete variable on the others in turn. Here the recruiter's and recruits' observed wave-$t$ values and summaries are added as predictor columns, and respondent-driven sampling (RDS) degree weights are included as a covariate. The second tier replaces the regressions with a longitudinal hidden-Markov imputer. A participant's latent risk-state transitions are conditioned on neighbours' inferred states, which pools information both across time and along recruitment ties. The third tier is a graph neural imputer that extends SHT-GNN with subject-to-subject edges from the coupon tree, optionally augmented with ties reconstructed by RENDER.

Each tier is pooled in two ways: naive Rubin pooling, and pooling with $T_{\text{net}}$. For $T_{\text{net}}$ we compare the fGLS covariance with branch-level resampling done before imputation. The downstream analyses are three dynamic-risk estimands. The first is the population-weighted prevalence of the high-risk state at each follow-up wave. The second is the association between baseline network degree and risk escalation, a longitudinal version of the Puerto Rico finding. The third is individual prediction of an escalation event from trajectories versus from the baseline survey alone.

Baselines

The naive reference point is complete-case analysis with Volz–Heckathorn weights. These weights divide each respondent's contribution by their self-reported network size, which offsets the tendency of recruitment to reach well-connected people. Here the weights are applied to whoever remains at each wave. This is what a longitudinal chain-referral study would plausibly do by default. Alongside it we run a static baseline-only risk model fit on the wave-0 survey. That model asks whether follow-up adds anything at all, in the spirit of the APACHE-II comparison in the intensive-care study.

The standard-practice baseline is MICE with degree weights under ordinary MAR, without peer columns, pooled with naive Rubin's rules. This corresponds to weighting after imputation in the style of MatchThem. It is the method an applied epidemiologist would reach for.

The strongest existing imputers are SHT-GNN, run as published without peer edges, and RefiDiff, run in its MNAR mode. SHT-GNN is the most relevant because it is the file's best longitudinal imputer; its reported numbers are a 17.5% RMSE reduction on GLOBEM and an AUC of 82.92% on ADNI. Those results concern different data. For imputation of chain-referral follow-up data there is no reported result found on Lacuna, so the comparison figures will be produced here.

We add inverse-probability-of-retention weighting. This models each person's probability of staying in the study, including from peer covariates, and reweights completers by its inverse. It is the main non-imputation competitor that also uses peer information.

All methods receive the same simulated cohorts, the same auxiliary variables where they can accept them, the same number of imputations ($m = 20$), and an equal hyperparameter-search budget tuned on held-out simulation seeds. A method "beats" the standard baseline if it reduces absolute bias in wave-$t$ prevalence by at least half at moderate homophily while achieving 95% interval coverage within 92–97%. The key ablations remove the peer columns from each tier, which isolates peer information, and swap $T_{\text{net}}$ for naive pooling, which isolates the variance correction.

Why not just…?

Why not just add the recruiter's values as extra columns in ordinary MICE? This is a real threat. It is our first tier, not a strawman. If homophily acts mainly through the direct recruiter and through linear relationships, it may match the hidden-Markov and graph tiers. If it does, we will report the simple method as the recommendation. The more elaborate tiers must earn their place on simulated regimes where information travels over two or more hops or through nonlinear state transitions. What MICE with extra columns cannot fix on its own is pooling. Our coverage experiments test whether naive Rubin intervals under-cover once imputations borrow from neighbours.

Why not just weight the completers by their estimated retention probability? Retention weighting uses the same peer information and avoids imputation models. However, it discards all partial data from dropouts, and it becomes unstable when retention probabilities are small, which is exactly the high-escalation subgroup. It also still requires a network-aware variance estimate. We run it head to head. We expect it to be competitive at mild dropout and to fall behind in both RMSE and interval width as dropout strength rises. If it does not fall behind, that is a useful finding in its own right.

Why not just use a powerful MNAR-capable imputer such as RefiDiff? Without auxiliary information, MNAR corrections rely on the model's assumed shape for the missingness process, which the data cannot check. RefiDiff's MNAR results come from tables of independent rows. Our homophily sweep settles this directly. If peer information is what matters, the gap between RefiDiff and the peer-informed tiers should grow with homophily and close to zero when homophily is zero.

Why not just keep the one-time baseline survey? The hospital evidence (0.836 versus 0.895 AUC; a +0.223 AUROC gain in the lowest-risk tertile) suggests that one-time scores miss people whose risk emerges later. Whether the same holds in community cohorts is an empirical question. We test it as the static reference in the prediction task.

Proposed Experiments

The file contains no public longitudinal chain-referral cohort, so the main evaluation is semi-synthetic, built on the real networks it documents. On the Project 90 and Add Health networks, which fGLS and RENDER also use, we attach to each node a latent three-state risk process whose transitions are coupled to neighbours' states with homophily strength $\rho$. We then run chain-referral recruitment with three coupons per person, varying referral rate, seed selection, and chain depth. We simulate four to six follow-up waves in which the probability of dropout rises with the current latent state at strength $\gamma$. A mixing parameter controls how much of that dropout driver is shared with peers and how much is private.

We report bias, RMSE, 95% coverage, and interval width for the prevalence and degree-association estimands. For prediction of escalation events we report AUROC and AUPRC, including within the baseline lowest-risk tertile.

The decisive checks are these. First, the peer-informed gain should shrink to zero as $\rho \to 0$, which serves as a negative control. Second, gains should fall as the private share of the dropout driver rises, and the delta-adjustment analysis should bracket the truth in that regime. Third, naive Rubin pooling should under-cover at high referral rates while $T_{\text{net}}$ restores nominal coverage. Fourth, adding RENDER-reconstructed ties should help only when the coupon tree misses informative neighbours.

As a real-data check on genuine homophily, we will take the St. Petersburg RDS cohort of 813 injection drug users used by RENDER, provided the recruitment tree and attributes are accessible. We will mask observed attributes with a missingness rule that depends on the masked values, then measure how well each method recovers them. This tests whether real recruitment ties carry enough signal to be worth conditioning on, which is the empirical premise of the whole proposal.