{
  "question": "How sensitive is a focused direct-comparison meta-analysis of walking/jogging versus active controls for depressive symptoms to trial influence, source risk-of-bias ratings, effect-size definition and conventional statistical assumptions?",
  "scope": "Computational robustness audit of the fixed Noetel et al. (2024) public extraction, not an updated systematic review, complete network reproduction, treatment ranking or new primary-trial extraction.",
  "source": {
    "project": "https://osf.io/nzw6u/",
    "file": "cleaned_deprex.RDS",
    "download": "https://osf.io/download/68xnt/",
    "sha256": "50584b02bee4c8f5a181917cffeb7c0be0b3b213baa2bf085db9bd6546667781",
    "csv_conversion": "R base write.csv, row.names=FALSE, na empty string; source RDS preserved. The data and published code implement arm baseline standardization, clarified by correction doi:10.1136/bmj.q1024."
  },
  "prior_knowledge": "The original article, correction, authors calculation code, and existing reviews were read. The dataset was downloaded and converted; column names, categorical labels, missingness, duplicate keys and eligibility metadata were inspected. Twenty-one candidate immediate post-treatment studies have both target categories, and none has overall low risk. No study effect estimate, pooled outcome, influential-study result or sensitivity estimate has been calculated for this audit. Original published pooled walking/jogging network estimates and wider-review summaries were already known.",
  "eligibility": {
    "exercise_label": "Walking / Jogging",
    "active_control_labels": [
      "Educational",
      "Social",
      "Social or educational control",
      "Usual care",
      "Placebo pill",
      "Stretching"
    ],
    "time": "weeks_from_end_of_treatment_to_measurement exactly numeric zero",
    "population": "All original clinically depressed trial populations without an additional adult-only restriction; use the original labels and eligibility, including background usual treatment.",
    "study_identity": "studyID as provided; arm_number identifies distinct arms. Do not count outcomes or timepoints as separate trials.",
    "measure_selection": "Require a depression measure shared by every eligible walking/jogging and active-control arm in the study at time zero. Prefer a measure rated Clinician in every selected arm; then select lexicographically by the exact outcome label. One measure per study. Exclude and log studies with no common measure.",
    "duplicates": "A study/arm_number/outcome/time key must be unique. Halt rather than silently averaging duplicate records.",
    "missing": "Require finite post means and positive SDs and integer post sample sizes greater than one. Do not impute further. Log all exclusions. Missing baseline inputs preclude only the paired change-score sensitivity, not an otherwise eligible endpoint comparison."
  },
  "multiarm": "Combine all eligible arms within each category using participant-weighted means and the exact pooled sample SD including between-arm mean differences. Produce one comparison per study; no control participant is reused in separate comparisons.",
  "primary": {
    "effect": "Post-treatment between-group Hedges g: exercise minus control mean, divided by their pooled post-treatment SD, with exact gamma small-sample correction at df=n_exercise+n_control-2. Lower depression scores are better, so negative g favors walking/jogging.",
    "sampling_variance": "Large-sample SMD variance matching metafor escalc(measure=SMD,vtype=LS): 1/n_exercise + 1/n_control + g^2/(2*(n_exercise+n_control)).",
    "model": "Random effects with restricted maximum likelihood (REML) heterogeneity. Modified Hartung-Knapp interval: multiplier max(1,weighted residual sum/(k-1)), t quantile with k-1 df, two-sided 95% interval.",
    "reported": "k, post-treatment participant counts, pooled g, SE, interval, two-sided nominal p-value, tau squared, Cochran Q and its p-value, I squared, and a conventional t(k-2) prediction interval when k>2. A single primary inferential contrast; no multiplicity adjustment. Sensitivity intervals and nominal p-values are descriptive, not additional confirmatory tests."
  },
  "planned_sensitivity": [
    "REML with normal intervals; unmodified Hartung-Knapp intervals; fixed-effect model; DerSimonian-Laird heterogeneity with normal intervals",
    "Leave each study out in turn, refit REML and the primary interval; report every fit and extrema without choosing a favorable exclusion",
    "Restrict to low risk for both random sequence generation and allocation concealment; separately restrict to low risk for blinded outcome assessor; separately restrict to overall low risk. Use the worst source rating across selected arms. Report insufficient evidence instead of pooling when fewer than two studies remain",
    "Use the sole lexicographically first common measure without preferring clinician ratings, and report any change in selection",
    "On the identical selected arm/measure rows, contrast participant-weighted means of the published arm-level baseline-standardized change g values, with independent-arm variance from the supplied se_smd. This is a different effect definition, not a like-for-like reproduction of the published network contrast",
    "For that arm-change contrast, vary the assumed pre/post correlation to 0, 0.5 and 0.8, recalculating arm variances as 2*(1-r)/n + g_arm^2/(2*n), as in the authors code, then refit REML and the primary interval",
    "Replace the heterogeneous active-control category with usual-care-only arms, rerun deterministic common-measure selection and pooling",
    "Publish selected study contrasts, complete candidate-study inclusion/exclusion reasons and source row identifiers, plus forest and leave-one-out figures"
  ],
  "validation": "Check unique keys and finite positive variances; verify multiarm combination identities and one contribution per study. Cross-check endpoint Hedges g and LS variances and REML estimates against an independent R/metafor implementation. Check a second REML score-root calculation against likelihood minimization. Verify source arm-change calculation against the authors documented Hedges correction implementation and variance formula, reporting any discrepancy rather than repairing source outcomes.",
  "interpretation": "Report only conditional computational estimates. Distinguish robustness of sign, interval exclusion of zero, magnitude and between-study prediction. No treatment ranking, proof of clinical efficacy, causal explanation of subgroup differences, or publication-bias elimination. Carry source extraction, sample attrition, bias and outdated search limitations. Publish all prespecified analyses, including failed or insufficient comparisons; label and log any additional exploratory work."
}