G-Comp Builder — Causal G-Computation

G-Comp Builder

DAG import · G-Computation · Counterfactual prediction

Generate ready-made R code for your own G-computation — import the DAG from the Golem Builder, adjust the model, export MCMC draws directly to the Decision Maker.

What is this about?

You have a causal question ("Does X affect Y?") and a model in mind — G-Comp Builder turns it into ready-made, commented R code for g-computation: from the model formula to the counterfactual predictions for ATE, ATT and ATU. The tool does not compute anything itself; it builds the code that you run in your own brms workflow. Unlike the Causal Calculator (guided example with simulated data), it is meant for your own analysis.

The idea behind it

G-computation (causal standardisation) answers the question: what would have happened on average if everyone had been exposed — and what if no one had? This works in three steps: first you fit an outcome model with exposure and confounders. Second, you predict the outcome for each person twice — once with the exposure at the reference value, once at the treatment value (the "two worlds"). Third, you average the difference across persons — depending on the estimand over everyone (ATE), only the exposed (ATT) or only the unexposed (ATU). Because your model is a Bayesian model, this happens for every posterior draw: in the end you get a whole distribution of the causal effect.

How to proceed

① DAG import: the Golem Builder clarifies which variables you must control for. Take exposure, outcome and adjustment set from there — or enter them by hand. A banner shows whether an import is present; a button reloads it.
② Model & formula: the formula is suggested from the DAG and directly editable; optional variables from the DAG can be added with a click. If a mediator was detected, you choose between the total effect (mediator not in the model: X→M→Y plus X→Y) and the direct effect (mediator in the model: only X→Y). Plus random effects, model type (brms, glmmTMB, lm/glm), family and the names of model object and dataframe.
③ Settings: is the exposure binary or continuous? Binary: reference and treatment value, estimands (ATE / ATT / ATU, optionally conditional on a moderator to see effect heterogeneity). Continuous: there is no natural "treated / untreated" — you choose a contrast (min→max, Q1→Q3, ±1 SD or manual) and an artificial ATT/ATU analogue via a cutoff (median or your own value). Note the difference between the AME (effect per unit, a slope) and the contrast from–to.
④ Copy the R code: the code in the output field updates live. Fit your model in R (brm(...)) and run the code — it uses marginaleffects, sets the exposure to the respective value for everyone and averages over the confounders.

What the tool does not settle for you

The code is only as causal as your DAG: it assumes that all relevant confounders are measured, that exposed and unexposed persons occur at every confounder level (overlap) and that the model describes the data decently. The validation box only checks the formal consistency of exposure, outcome and adjustment set — not whether your DAG is right in substance.

And afterwards?

The code exports the MCMC draws of ATE/ATT/ATU as CSV — on the response scale, so directly interpretable. Load them into the Decision Maker to derive a well-founded decision with HDI/ETI/ROPE.

G-Comp Builder — Help

What is G-Computation?

G-Computation (causal standardization) estimates what would happen on average if we set every person to a specific exposure value — as opposed to "which people actually had that value".

Key marginaleffects concept: variables=list(X=val) sets X to val for every observation and averages over all confounder distributions. This is true marginalization — not a "typical person" estimate.

Binary exposure — estimands

ATE (Average Treatment Effect): E[Y(1)−Y(0)] — effect in the full population.
ATT (on the Treated): E[Y(1)−Y(0) | X=1] — effect only among the actually exposed.
ATU (on the Untreated): E[Y(1)−Y(0) | X=0] — effect if the unexposed were treated.

ATE ≠ ATT ≠ ATU when there is effect heterogeneity. Under randomization without interactions they are equal.

Metric exposure — AME & ate_k

AME (Average Marginal Effect): Average partial derivative dE[Y]/dX. Answers: "By how much does Y change per one unit increase in X?"
ate_k (causal contrast): E[Y(x_hi) − Y(x_lo)] — averages the difference between two specific X values over all persons. Answers: "What would the difference be between Min and Max, Q1 and Q3, etc.?"

AME and ate_k are conceptually different: AME is a slope (per unit), ate_k is a total contrast (from–to). Both are causal effects when confounding is controlled.

ate_k — contrast strings in marginaleffects

"minmax" — Min → Max of the exposure
"iqr" — Q1 → Q3 (25th → 75th percentile)
"sd" — Mean−SD → Mean+SD
c(lo, hi) — user-defined interval

ATT/ATU analogue (metric)

For metric exposures there is no natural "treated/untreated". The analogue splits the population artificially at a cutoff (default: median):

att_k: contrast — only for observations with X ≥ cutoff ("highly exposed")
atu_k: contrast — only for observations with X < cutoff ("low exposed")

Useful for checking whether the effect differs for already highly exposed vs. low exposed persons — effect heterogeneity along the exposure axis.

Dose-response curve

Estimates E[Y | do(X=x)] for 100 values from min to max. Each point is a genuine G-Computation estimate: the entire population is "set to X=x" and the marginal mean is computed.

ICE — Individual causal effects

Individual Counterfactual Effects: τ̂ᵢ = Ŷᵢ(x_hi) − Ŷᵢ(x_lo). Shows heterogeneity in the causal effect across persons. Sorted by Ŷᵢ(x_lo) — same row = same person in both plots.

Workflow

Build the causal model in the Golem Builder ↗: draw the DAG, check the adjustment set, choose Total/Direct Effect → click → G-Comp Builder
Formula, adjustment set, effect type, and DAG structure arrive here automatically (via localStorage) — check the formula, fit the model in R
Copy code, compute estimands
Export MCMC draws → Decision Maker

The Golem Builder identifies what needs to be controlled for, but doesn't compute effect estimates itself — that's this tool's job. Without a Golem Builder import, G-Comp Builder also works fully manually: enter exposure/outcome/adjustment set yourself on the left.

Only variables from the adjustment set are required in the formula. Precision covariates (causes of Y but not X) can safely be added — they reduce variance without introducing bias.

1 — DAG Import

2 — Model & Formula

3 — Settings