The paper

Methods

Registration. The plan, codebooks, decided trial designs, and analysis code were registered before any number was extracted (the registration; files under plan/). Before registering, the author had read the abstracts of the main reports of MORDOR, AVENIR, CHAT, LAKANA, MIRAMA, and PRET's mortality analysis, so their headline estimates were known; what the registration fixed is which trials and comparisons count, which estimate each contributes, the models, the sensitivity analyses, and how each result would be read. Changes after registration are listed in deviations.json.

Question and strategies. The review includes randomized trials in which azithromycin was given to children at population scale regardless of illness, against placebo, no treatment, delayed treatment, or another azithromycin schedule, with all-cause mortality among children measured. Each randomized comparison has one strategy, coded from the trial's design: all_children_mda (communities randomized; azithromycin to every child 1 to 59 months old), infant_only_mda (communities randomized; azithromycin to infants 1 to 11 months old only), individual_infant (individuals randomized; azithromycin to every neonate or infant at a visit), individual_child, community_wide_mda (trachoma programmes treating everyone), and schedule_comparison (no untreated arm). Each azithromycin arm against the trial's untreated arm is one comparison; arms sharing one control are marked as sharing it; a trial randomized separately within countries gives one comparison per country.

Search and screening. PubMed and Europe PMC (titles and abstracts) were searched on 2026-10-17 for azithromycin, mass or community administration, mortality, and randomization (277 distinct records), and ClinicalTrials.gov for interventional studies of azithromycin with mortality or survival among their outcomes (172 registrations); PubMed was also searched for each eligible registration's number, which found reports whose abstracts the terms missed (data/search.json, data/screening.json). Two coders, separate instances of the model that wrote this study, screened every record and registration alone from plan/screening_codebook.md; they agreed on 251 of 277 records (κ = 0.74) and all 172 registrations, and a third instance decided the 26 records they differed on (all trial substudies one coder kept in case they reported deaths; it kept one).

Designs. Reports were grouped into 17 trials (data/groups.json). Two further coders coded each trial's design alone from its registration and the Methods of its open reports and protocols (plan/design_codebook.md); they agreed on eligibility, unit of randomization, control, and masking for all 16 trials coded, and on the comparisons' strategies for 14; a third reader decided the other two. Fourteen trials are eligible (data/trials.json). Three trials' registrations were first submitted after enrollment began (MORDOR II, PRET, and the trial adding azithromycin to seasonal malaria chemoprevention); no report carries a retraction or expression of concern.

Extraction. After registration, two coders copied every printed all-cause mortality estimate for each comparison, by age band, period, country, and subgroup, from the Results, tables, and figure captions of the trials' open reports, the abstracts of reports without open full text, and results posted on ClinicalTrials.gov (plan/extraction_codebook.md). Entries were paired by label and then by identical numbers: 146 pairs agreed exactly, and a third reader decided the remaining 24 items against the sources (data/coding/). The decided entries are in data/extraction.json.

Effect measure. All intervals in this paper are 95% confidence intervals. Each estimate is the log of a mortality ratio, azithromycin against control, with its standard error from the printed 95% interval; a printed percent reduction $p$ becomes the ratio $1 - p/100$. Incidence rate ratios, hazard ratios, and risk ratios are treated as one measure, as deaths are rare. An estimate printed only as deaths and person-time per arm gets its Poisson standard error times the square root of the trial's design effect. For each comparison and age band the code takes the whole-follow-up estimate for all children in the band from the primary analysis, preferring the main report, and combines narrower bands by inverse variance where the band itself isn't printed. Two comparisons of one trial that share a control and enter one analysis (AVENIR's arms in P3) have their variances recomputed with the control split between them.

Analyses. All pooled estimates are random-effects models by restricted maximum likelihood with Knapp–Hartung intervals (Knapp and Hartung (2003); Viechtbauer (2010)). P1 pools the all_children_mda comparisons against placebo on mortality at 1 to 59 months; P2 pools the twice-yearly infant_only_mda comparisons on infant mortality (1 to 11 months); P3 is the coefficient of infant_only in a meta-regression of infant mortality over the P1 comparisons' infant estimates and the P2 comparisons. Holm's correction covers the three. Pre-registered sensitivity analyses: a common-effect model; a Bayesian random-effects model with a half-normal prior of scale 0.5 on τ (Röver et al. (2021)); leaving out each comparison; double-masked placebo comparisons only; for P2, LAKANA's four-times-yearly arm, both LAKANA arms, without MIRAMA, and with the individual_infant comparisons. Secondary analyses: a meta-regression on the log of the control arm's mortality rate; the slope of the log rate ratio on years since the first distribution; infants against children 12 to 59 months within all-children trials; the individual_infant comparisons; dosing frequency; and the registrations due to report (primary completion at least 12 months before the search) that hadn't. The registration fixed how each result would be read: P1 robust if its Holm-corrected p, its common-effect p, and every leave-one-out p are below 0.05; P2, if not significant, read by the largest reduction its interval allows and whether that interval excludes P1's reduction; P3 robust only if AVENIR's printed within-trial contrast also favours treating all children. Methods follow the Cochrane Handbook where it applies (Higgins et al. (2019)).

After registration. Three things were added and are reported as exploratory (code/summarize.R, deviations.json). MORDOR II and AVENIR's contrast of its two treated arms print each arm's mortality rate with its interval but no ratio and no counts, so the registered rules leave them out; their ratios were estimated from the two rates, with the standard error of the log from each rate's interval, the arms taken as independent. The two individually randomized infant trials were also combined by a common-effect model, since Knapp–Hartung intervals from two studies carry little information. A bug in the registered code, which let a band of 1 to 59 months stand for 12 to 59 months, affected only the secondary age analysis and was fixed.

Sources of each trial's numbers. MORDOR: Keenan et al. (2018), its third year in Niger Keenan et al. (2019), and Tanzania's age-specific estimates Bloch et al. (2021). AVENIR: O'Brien et al. (2024) and Arzika et al. (2025). CHAT: Oldenburg et al. (2024). LAKANA: Haidara et al. (2025). MIRAMA: Yago-Wienne et al. (2026). The well-infant visit trial: Sié et al. (2024). NAITRE: Oldenburg et al. (2022) and its 12-month report Sié et al. (2022). PRET: O'Brien et al. (2018). TANA: Porco et al. (2009). MORDOR II's numbers are its results posted on ClinicalTrials.gov (NCT03338244).

Running it. code/run builds every result from data/ in the container env/Dockerfile builds (R 4.6.1, metafor 5.2-1, jsonlite 2.0.0), in seconds.

Results

Treating all children under five. Across 5 comparisons from 3 trials, mortality among children under five was 14.1% lower with azithromycin (rate ratio 0.859, CI 0.797 to 0.926; p = 0.005, p = 0.015 after Holm correction; τ² 0, I² 0%). The common-effect estimate was 0.859 (0.813 to 0.908), the Bayesian estimate 0.865 (0.784 to 0.973), and leaving out each comparison in turn gave estimates from 0.847081068925482 to 0.88111303973325, the largest p 0.0440892455174892: robust by the registered rule. Table 1 gives each comparison. Adding MORDOR II, estimated from its printed rates (0.929, 0.839 to 1.029), gave 0.877 (0.819 to 0.938; p = 0.004), an exploratory estimate.

Table 1. Each comparison's mortality rate ratio and confidence interval as the analyses used it.

The AVENIR rows for infants carry the wider intervals P3 used, with the shared placebo arm split between them.

Treating infants alone. In the 3 trials that gave azithromycin twice a year to infants only, infant mortality was not detectably lower (rate ratio 0.971, CI 0.867 to 1.086; p = 0.370; I² 0%). The interval allows at most a 13.3% reduction, which excludes the 14.1% reduction of treating all children. The common-effect estimate was 0.971 (0.872 to 1.08) and the Bayesian 0.977 (0.753 to 1.298). With LAKANA's four-times-yearly arm in place of its twice-yearly arm the estimate was 0.948 (0.856 to 1.05); with both LAKANA arms, 0.961 (0.897 to 1.029); without MIRAMA, 0.963 (0.656 to 1.413); and with the individually randomized infant trials added, 0.969 (0.881 to 1.066).

Infants under each strategy. In the trials that treated all children, infant mortality was lower with azithromycin (pooled rate ratio 0.823, CI 0.707 to 0.958). The registered meta-regression comparing the infant rate ratios of the two strategies gave a ratio of 1.187 (0.974 to 1.447; p = 0.074, p = 0.148 after Holm correction), so the difference is not established. The registered robustness check, AVENIR's printed contrast of its child and infant arms, could not be run: its open sources print each arm's infant mortality rate but not their ratio. From those rates, the contrast is 0.83 (0.717 to 0.959; p = 0.0118), and combining the two ratios AVENIR prints for infants with and without an older sibling gives 0.827 (0.719 to 0.952); both are exploratory.

Where mortality was higher. Across 8 mass-distribution comparisons, the log rate ratio changed by -0.078 per unit of the log control mortality rate (CI -0.144 to -0.013; p = 0.0271), allowing for strategy: larger reductions where more children died.

Over time. In the two trials that printed estimates by year, the log rate ratio changed by -0.008 per year of treatment (CI -0.207 to 0.191).

By age within all-children trials. Only two comparisons print both an infant and an older-child estimate after the bug fix; their pooled ratio was 1.036 (0.083 to 12.953), uninformative. In AVENIR, from its printed rates, children aged one to four years had a rate ratio of 0.892 (0.807 to 0.985), and the ratio of its infant to its older-child rate ratio was 0.864 (0.726 to 1.027), exploratory.

Individual dosing of infants. The registered random-effects estimate from the 2 individually randomized trials was 0.956 (CI 0.147 to 6.225), too wide to inform; the common-effect estimate added after registration was 0.963 (0.76 to 1.222), and 0.996 (0.777 to 1.277) with NAITRE's six-month primary outcome.

Frequency and other designs. In LAKANA, four doses a year against two gave an infant mortality ratio of 0.93 (0.756 to 1.144). In PRET, twice-yearly treatment of children against annual treatment of everyone gave 0.82 (0.675 to 0.997) for the widest band PRET prints, and 0.81 (0.66 to 1) for the band its report names as its mortality outcome. TANA's community-wide trachoma treatment against delayed treatment gave an odds ratio of 0.51 (0.289 to 0.898) for children aged one to nine years. The trial adding azithromycin to seasonal malaria chemoprevention printed deaths only within a composite with hospital admission.

Unreported trials. Of 13 eligible registrations, 12 were due to report and 3 had not: SANTE (Mali) and ICARIA (Sierra Leone), which give azithromycin to infants at routine visits, and TIRET (Ethiopia), a trachoma trial.

Discussion

What the trials show together. Each trial of treating all children under five found lower mortality or a non-significant reduction in the same direction, and pooled by strategy they agree closely, with no detectable heterogeneity. Each trial of treating infants alone found no significant effect, and pooled they bound the effect away from what treating all children achieved. The 2020 guideline restricted mass azithromycin to infants on the grounds that MORDOR's effect was largest in infants and that treating fewer children would select less resistance (World Health Organization (2020)). Three trials designed to test that restriction have since reported, AVENIR in Niger, LAKANA in Mali, and MIRAMA in Burkina Faso, and none found a significant reduction in infant mortality with infant-only treatment; AVENIR's authors concluded that treating all children under five was more effective, and MIRAMA's that policy should consider treating all children under five (O'Brien et al. (2024); Haidara et al. (2025); Yago-Wienne et al. (2026)). This review adds the pooled estimate across the three, with its upper limit on the benefit.

Earlier syntheses. A 2019 pooled analysis of three trials in four countries found a 14.4% reduction in child mortality, which this review's estimate for treating all children matches (Oldenburg et al. (2019)). A 2025 meta-analysis pooled five trials by the age at which children died, and found the effect on infant mortality uncertain (risk ratio 0.90, 95% CI 0.78 to 1.04) while mortality at 12 to 59 months fell (Gurunthalingam et al. (2025)); it was searched before LAKANA and MIRAMA reported. This review differs in grouping the trials by whom they treated and in including those two trials: infants had lower mortality where everyone under five was treated, and not where only infants were. A 2025 review of registered azithromycin trials described their designs without pooling outcomes (Kong and So (2025)).

Why treating infants alone might not work. One explanation is that azithromycin protects infants partly by treating infections, such as malaria, carried by the older children around them, a path infant-only distribution would miss. AVENIR's secondary analysis found the advantage of treating all children over treating infants larger among infants with an older sibling at home (0.78 against 0.91), though the difference between those subgroups wasn't significant (Arzika et al. (2025)), and in CHAT the only cause of death that fell significantly was malaria (Sié et al. (2026)). Other explanations are open: the control arms of LAKANA and MIRAMA recorded lower infant mortality than AVENIR's (11.9 and 2.5 deaths per 1,000 person-years against 23.9, MIRAMA's counted from birth histories), and this review's mortality gradient suggests smaller effects where fewer children die, though it rests on few comparisons; LAKANA also counted deaths only in the three months after each dose. AVENIR, though, tested both strategies in one setting, where the infant arm's reduction was small too. What would tell these apart is within-trial evidence: AVENIR's contrast of its two treated arms, which its open sources don't print as a ratio, and the ongoing AVENIR II trial (Arzika et al. (2026)). Giving azithromycin to individual infants at clinic visits (Sié et al. (2024); Oldenburg et al. (2022)) also showed no effect, which fits an effect that depends on treating whole communities, though both trials recorded few deaths, which limits what they can show.

What follows. For a guideline weighing child survival against resistance, the evidence now differs by strategy: treating all children under five lowers mortality by 14.1% in the settings tested, while treating only infants has not been shown to lower it, so restricting distribution to infants, meant to keep the benefit while limiting resistance, has not kept the benefit in the trials that tested it. Treating all children uses more azithromycin and selects more resistance: in AVENIR, macrolide resistance genes in children's guts rose most in the arm that treated all children (Doan et al. (2026)). In a trial run beside AVENIR, treating all children cost less per dose delivered than treating infants alone, and caregivers found it more acceptable (Arzika et al. (2025)). Weighing resistance against survival is beyond this review; it shows that the infant-only option, as tested, offers little survival benefit to set against the resistance it spares. The larger reductions where control mortality was higher fit targeting by mortality, as the guideline proposed. The effect didn't wane over two to three years in the two trials that report it, and MORDOR II, which re-randomized Niger's communities after three years of treatment, gave a consistent estimate.

What would settle the rest. The results of SANTE and ICARIA, two trials of azithromycin given to infants at routine visits that were due to report and haven't (Driscoll et al. (2023); Owusu-Kyei et al. (2024)), would sharpen the estimate for individual infant dosing. A printed ratio for AVENIR's child-arm against infant-arm contrast, and AVENIR II, would settle whether treating older children is what protects infants. Trials outside the Sahel and East Africa would show how far the all-children effect travels.

Limitations

• Few trials per strategy. The all-children estimate rests on three trials (five comparisons) and the infant-only estimate on three; the infant-only interval excludes the all-children reduction only narrowly, and doesn't exclude a benefit of a few percent.
• Between-trial comparison. P3 compares strategies across trials that differ in setting, mortality, and how deaths were counted; only AVENIR compares the strategies within one trial, and its ratio isn't printed in its open sources. The exploratory estimates from printed rates assume each arm's interval reflects its uncertainty and that the arms are independent.
• Estimates left out by the registered rules. MORDOR II and AVENIR's within-trial contrast print only per-arm rates; MORDOR's infant estimates are printed for Tanzania but not for Niger or Malawi, so MORDOR contributes only Tanzania's infants to P3; the trial adding azithromycin to seasonal malaria chemoprevention prints deaths only within a composite.
• Different outcomes. Deaths were counted by census in most trials, from birth histories in MIRAMA, and over the three months after each dose in LAKANA; the analyses treat rate, hazard, and risk ratios as one measure.
• Sources. AVENIR's main report and the well-infant trial's report aren't open; their numbers come from abstracts, secondary reports, and registry results. Search covered PubMed, Europe PMC, and ClinicalTrials.gov only.
• One model. Screening, coding, and extraction were done by instances of one model family; errors it shares across instances wouldn't show as disagreements.
• Resistance and costs were not analysed.
• A bug in the registered code affected one secondary analysis and was fixed after registration.

Provenance

A Claude-family model (claude-opus-5-5) chose the question, designed the review, searched PubMed, Europe PMC and ClinicalTrials.gov, wrote the plan, codebooks, and analysis code, and registered the plan. Separate instances of the same model screened the records and registrations, coded the trials' designs, and extracted the numbers, two independently and a third deciding wherever they differed. The model then ran the analysis, wrote the claims and this paper, and applied the hazard screen. Full texts came only from open sources; no full text or abstract is in the bundle, only identifiers, titles, codings, and the numbers copied from the trials' reports and registry results. A person lent the agent and asked it to research a non-mathematical science question that matters; no person chose the question or wrote, coded, extracted, analysed, or reviewed any part of the work.