contact-tracing-analytics

Overview

contact-tracing-analytics covers the quantitative side of contact tracing: reconstructing transmission networks from case-contact data, estimating key parameters (secondary attack rates, serial intervals), identifying superspreading, and evaluating whether tracing programs actually work. It bridges field epidemiology and network science.

When to use

• Analyzing case-contact datasets: completeness, timeliness, yield.
• Estimating secondary attack rates by setting and contact type.
• Reconstructing transmission chains and networks.
• Identifying superspreading events and high-risk settings.
• Evaluating tracing performance: speed, coverage, forward vs backward tracing.
• Digital tracing data: app-based proximity, privacy-preserving analytics.

Core concepts

• The tracing cascade. Exposed → identified as contact → reached → tested/quarantined → positive → their contacts traced. Measure yield and delay at each step — programs usually fail at reach and speed, not at epidemiology. Report the full cascade, not just cases found.
• Timeliness is effectiveness. Tracing that reaches contacts after their infectious period prevents nothing. Key metrics: time from index case symptom onset to contact quarantine; proportion of contacts reached within 48 hours. Model the relationship explicitly — each day of delay costs a quantifiable fraction of prevented transmissions.
• Secondary attack rate (SAR). Proportion of contacts who become cases, stratified by setting (household, workplace, social), contact type, and index-case characteristics. SAR by setting is the evidence base for targeted interventions — household SARs are typically an order of magnitude above casual-contact SARs.
• Forward vs backward tracing. Forward tracing finds who the case may have infected (prevents onward transmission); backward tracing finds who infected the case (finds clusters and superspreading events). Backward tracing is disproportionately valuable for overdispersed pathogens — finding one superspreading event beats finding ten isolated transmissions.
• Overdispersion and superspreading. The dispersion parameter k quantifies transmission heterogeneity; small k means most transmission comes from few individuals/events. Tracing data can estimate k from the offspring distribution — and it argues for cluster-focused (backward) strategies when k is small.
• Network reconstruction. Cases as nodes, transmission links as edges (inferred from contact + timing + genomics). Analyze: chain lengths, cluster sizes, setting-specific subnetworks. Genomic data resolves ambiguous links — integrate where available.
• Serial interval. Time between symptom onsets in infector-infectee pairs, estimated from traced pairs with known links. Needed for Rt estimation; watch for truncation bias (recent pairs incomplete) and recall bias in onset dates.
• Digital tracing. Bluetooth proximity apps generate exposure notifications at scale but with noisy distance/duration proxies. Evaluate: adoption rates (network effects dominate — low adoption ≈ no effect), positive predictive value of notifications, and equity (smartphone access). Privacy-preserving designs (decentralized) trade analytic visibility for trust.

Practical workflow

1. Assemble data. Case-contact line lists with dates (onset, test, interview, quarantine), contact types, settings, outcomes. Link records deterministically where possible.
2. Cascade analysis. Compute yield and delay at each step; stratify by demographics and geography to find inequities.
3. Parameter estimation. SAR by setting/type; serial interval from linked pairs (adjust for truncation); offspring distribution and k.
4. Network analysis. Reconstruct chains; identify clusters and superspreading events; characterize high-risk settings.
5. Performance evaluation. Timeliness metrics, coverage (contacts identified per case vs expected), prevented-transmission estimates via modeling.
6. Report. Cascade figures, SAR tables, timeliness distributions, equity stratifications, and concrete operational recommendations (the point is improving the program).

Common pitfalls

• Counting traced cases without measuring delay (slow tracing = theater).
• Forward-only tracing for overdispersed pathogens (missing the cluster-finding value).
• SARs without setting stratification (meaningless averages).
• Serial intervals from unlinked pairs or with truncation bias.
• Ignoring the denominator: contacts never identified.
• Digital-tracing evaluations without adoption-rate context.
• Equity-blind analysis (tracing works worst where it's needed most).
• Overdispersion ignored — forward-only tracing designed as if spread were homogeneous.
• Contacts' contacts never followed, missing second-generation transmission chains.