coreSR.AIcoreSR
Usage Guide

DT-MBNMA

Dose-time model-based network meta-analysis

1

Why use DT-MBNMA

  1. 1Standard NMA treats each treatment as a separate node and usually uses a single end-of-trial timepoint — so different doses of the same drug become disconnected nodes, and the time-course of the effect is discarded.
  2. 2DT-MBNMA instead treats the drug AGENT as the node and models DOSE as a continuous covariate and TIME as a longitudinal trajectory, fitting parametric dose-response and time-course curves by Bayesian MCMC (JAGS).
  3. 3This borrows strength across doses and timepoints: you can combine early phase-2 dose-finding trials (many doses, short follow-up) with phase-3 trials (few doses, long follow-up) in one coherent model.
  4. 4Use it when your evidence base varies by dose and/or has repeated measurements over time, and you want to predict the relative effect at any dose or timepoint — including doses/times no single trial reported.
Tip: DT-MBNMA reduces to the familiar cases: a single timepoint → dose-response MBNMA (MBNMAdose); discrete doses → time-course MBNMA (MBNMAtime). It also fits the joint dose×time response surface directly (a custom JAGS model), so you can predict the effect at any dose AND any timepoint at once.
2

How your data should be structured

  1. 1Use LONG format: one row per study × arm × timepoint. A single arm measured at 4 timepoints contributes 4 rows.
  2. 2Required columns: study (trial id), agent (the drug / node, e.g. "semaglutide"; use "placebo" for the reference), dose (a number; 0 for placebo), time (a number, e.g. weeks), n (arm sample size).
  3. 3For a continuous outcome (e.g. % weight change) add mean and se (standard error of the mean). For a binary outcome add events instead of mean/se.
  4. 4Pick one agent as the reference (the placebo / dose-0 anchor). Every other agent is estimated relative to it.
  5. 5For a pure dose-response (single-timepoint) analysis, give one timepoint per arm. For a time-course analysis, give the repeated timepoints.
Tip: Column names are flexible — the upload step auto-detects common variants (study/studyid/trial, agent/drug/treatment, dose, time/week, n/total, mean/y, se, events/r). You confirm the mapping before transforming.
3

Example dataset (continuous outcome, dose-response)

  1. 1A minimal worked example: a 3-agent network (placebo, drugA, drugB) measured at week 52, with several doses per drug. Download it, upload it on the Data tab, map the columns, then run a "dose" analysis with an Emax dose-response function.
Tip: For a TIME-course example, repeat each arm at several times (e.g. weeks 4, 12, 28, 52) with the response at each — keep dose fixed per arm — then run a "time" analysis with an Emax or exponential time-course function.
One row per study arm. placebo is dose 0 (the reference); drugA and drugB each appear at multiple doses. Outcome = mean change (lower is better) with its standard error.
studyagentdosetimenmeanse
S1placebo05280-10.5
S1drugA15280-3.50.5
S1drugA35280-60.5
S2placebo05275-1.20.5
S2drugB25275-40.55
S2drugB45275-6.50.55
S3drugA35260-6.20.6
S3drugB45260-6.80.6
S4placebo05290-0.80.45
S4drugA15290-3.20.45
Download example CSV
4

Set up the agents (network nodes)

  1. 1After uploading + transforming, the Agents tab lists every agent it found.
  2. 2Mark exactly one agent as the Reference (placebo / dose-0 anchor).
  3. 3Choose a dose-response function per agent: emax (saturating; the usual choice), loglinear, linear, or nonparam (treat each dose as its own level).
  4. 4Save the agents before running.
5

Run the analysis

  1. 1The Analyze tab shows a "Recommended for this data" card — a data-driven starting point for the mode, response function(s) and effects model — that you can apply with one click.
  2. 2On the Analyze tab choose the mode: Dose-response (MBNMAdose), Time-course (MBNMAtime), or Joint dose × time (the response surface).
  3. 3Pick the response function(s) — joint mode uses both a dose-response and a time-course shape — and the effects model (random or common). To choose rigorously, click "Compare functions (DIC)": it fits the candidate shapes and ranks them by DIC (lower = better fit), and you can apply the winner.
  4. 4For Time-course mode you can also model within-arm correlation across repeated timepoints: AR(1) (decays with the time gap) or compound symmetry (a constant ρ). The joint surface treats repeated measures as independent.
  5. 5Set MCMC controls (chains, iterations, burn-in, thinning, seed) — these are just a starting point: if it has not converged, the sampler auto-escalates (burn-in + iterations double and it re-runs, up to 3 times).
  6. 6Click Run. Progress streams live; results land on the Outputs tab.
Tip: Watch the convergence chip (max R-hat ≤ 1.1 with ESS ≥ 400 on the structural parameters). Emax-type parameters can be weakly identified when the data don’t bracket the half-max — if it still says not converged after auto-escalation, raise the base iterations or start from a common-effect fit.
6

Read the outputs and GRADE

  1. 1The Outputs tab shows the dose-response or time-course curves (predicted relative effect vs dose / time with 95% credible bands) — or, in joint mode, the dose×time response surface (one line per timepoint across dose) plus its marginal slices — alongside the structural-parameter table (e.g. Emax, ED50 / ET50 per agent), between-study heterogeneity SDs, model fit (DIC), and convergence diagnostics.
  2. 2The GRADE tab lets you record certainty of evidence per agent-vs-reference contrast.
  3. 3Large between-study SDs on the structural parameters flag possible model misspecification or inconsistency — a built-in diagnostic of the model-based approach.
coreSR.AICoreSR

AI-powered evidence synthesis