Design Experiments From Reported Results

 Quick Analysis Plan Copied

 Long Analysis Plan

Trial Architecture

Twenty-four farms are randomized to 4 sequences of 6 farms each. Every farm contributes one baseline (control) month and three post-crossover (intervention) months. Weekly AGI surveillance covers all farm workers/residents (target n≈10/farm), yielding 24 farms × 4 months × 4 weeks × 10 workers ≈ 3,840 person-weeks. The stepped-wedge design is chosen because every farm eventually receives the intervention (ethical advantage in an occupational setting) and because within-cluster time-trend adjustment is more efficient than parallel allocation when cluster count is modest ACS QUIK stepped-wedge design paper.

Intervention: (a) point-of-use chlorination of cooking water with residual monitoring (0.2–0.5 mg/L free chlorine twice weekly by field staff using DPD colorimetry); (b) dedicated handwashing stations with soap at kitchen and latrine entrances. Fidelity metrics: residual chlorine ≥0.2 mg/L in ≥80% of spot-checks, soap availability ≥90% of visits.

Sample Size & Power Framework

Weekly AGI (≥3 loose stools/24h or ≥2 vomiting episodes) is binary per person-week. Assuming control incidence p_c=0.15 and target p_t=0.075 (50% relative reduction), the within-person variance s²₁ = [0.15×0.85 + 0.075×0.925]/2 = 0.0984. With farm-level ICC ρ=0.05 and 160 person-weeks per farm, s²_tot = 0.0984/(1−0.05) = 0.1036. The multilevel variance-inflation framework for stepped-wedge designs with >2 clustering levels (person-week → farm → sequence) provides VIF_SW factors that account for both the cluster structure and the timing of crossover Multilevel stepped-wedge sample-size methodology. Preliminary VIF_SW≈0.3 (for T=4 periods, ICC=0.05) suggests 24 farms × 160 person-weeks = 3,840 person-weeks yields ~80% power at α=0.05 for a 7.5-percentage-point absolute reduction.

Note: values above are design-stage estimates assuming constant ICC over time, no secular trend beyond fixed period effects, and cross-sectional sampling within farm-periods. Simulation-based power verification is recommended before finalizing cluster count, especially with fewer than 30 clusters Simulation recommendation for small stepped-wedge designs.

Pathogen Bioinformatics Pipeline

Stool is collected from all consenting workers at baseline (Month 1) and at each farm's Month 4 (3 months post-intervention for intervention arms; matched timepoint for Seq 1 control). Samples are processed via a two-tier approach:

Tier 1 — Multiplex qPCR panel (all samples): Targets: Norovirus GI/GII, Rotavirus A, Adenovirus 40/41, Astrovirus, Sapovirus, E. coli (stx1/stx2, eae, lt, st), Shigella/EIEC (ipaH), Salmonella (invA), Campylobacter jejuni/coli (hipO/glyA), Cryptosporidium (18S), Giardia lamblia (18S), Entamoeba histolytica (18S). Ct ≤35 = positive; Ct 35–40 = indeterminate (counted as negative in primary, positive in sensitivity).

Tier 2 — Shotgun metagenomics (10% random subsample + all qPCR-positives for pathogen-level genotyping): DNA extraction (QIAamp Fast DNA Stool Mini Kit), Illumina NovaSeq 2×150bp, host-read removal with Bowtie2+GRCh38, taxonomic classification with Kraken2+Bracken, virulence-gene confirmation with ABRicate+VFDB, assembly-based genotyping of norovirus (VP1) and E. coli (serotype/WGS) with SPAdes+shovill.

Primary computational analysis (Python3/Snakemake):

• AGI incidence model: Mixed-effects logistic regression, glmer(agi ~ intervention + factor(period) + (1|farm), family=binomial), ITT, fixed period effects, random farm intercept; odds ratio with 95% CI.
• Pathogen-level effect: For each pathogen, mixed-effects logistic (positive/negative at each timepoint), farm as random effect, period as fixed, intervention as fixed; FDR-corrected (Benjamini-Hochberg, q<0.10) across 12 pathogen tests.
• Pathogen load change: For qPCR-positive samples, log₁₀ genome copies/g as outcome; linear mixed model; tests whether intervention reduces pathogen burden conditional on infection.
• Co-detection network: Jaccard similarity of pathogen co-occurrence across samples; igraph network comparing baseline vs follow-up structure; permutation test for network density change.
• Mediation (exploratory): Tests whether pathogen-specific prevalence reduction mediates the AGI incidence effect; bootstrap CIs (n=5,000).

Analysis Plan & Model Specification

Sensitivity analyses: (a) Per-protocol (farms with ≥80% chlorine compliance); (b) Exclude Seq 1 control-period to test Hawthorne surveillance effects; (c) Lag model (1-week lag between intervention and effect); (d) Negative-control outcome (non-infectious injury) to detect residual confounding.

Assumptions, Blindspots & What Would Change Conclusions

Assumptions requiring verification: (1) Control AGI incidence of 15%/person-week is a design assumption; if baseline is lower (e.g., 8%), the same 24 farms yield <50% power for a 50% relative reduction. (2) ICC=0.05 is borrowed from healthcare cluster trials; farm-level ICC for AGI may differ (shared meals, shared water source, animal contact). (3) Constant correlations over time are assumed by the analytical VIF framework; seasonality in enteric pathogens violates this and is modeled via fixed period effects but cannot be fully separated from step timing with only 4 periods. (4) The VIF_SW≈0.3 estimate assumes cross-sectional sampling; if the same workers are followed longitudinally (cohort design), individual-level correlation adds another layer Design assumptions and simulation validation. (5) Stool sampling at only two timepoints cannot capture transient pathogen carriage between samples; qPCR detects DNA from non-viable organisms (especially bacteria), potentially inflating prevalence. (6) No blinding of workers or field staff is feasible; outcome reporting (AGI symptoms) is self-reported and subject to reporting bias that may differ by intervention status (social desirability bias favoring the intervention).

What would disprove the design rationale: If baseline surveillance reveals farm-level AGI incidence <5%/person-week, the cluster count must increase to ≥48 farms for adequate power. If farm-level ICC exceeds 0.15, the design effect overwhelms the stepped-wedge efficiency advantage and a different randomization unit (e.g., household within farm) becomes necessary. If qPCR panel results show ≥80% pathogen positivity at baseline (polyparasitism), pathogen-specific effects will be confounded by co-infection and require multivariate pathogen models.