The Gold Standard of Experimental Design | Parametric G-Formula, Dynamic Interventions and Monte Carlo Standardisation

1. Specify time zero, order and the target population

Back to contents

Align eligibility, treatment opportunities, covariate measurement, censoring and outcome assessment. Within each interval, state whether covariates are measured before treatment. The simulation will reproduce the order encoded by the analyst, so ambiguous same-visit timing can create impossible histories. Define the population over which risks or means will be standardised.

2. Write each intervention as an executable rule

Back to contents

A static strategy may set treatment to one value at every time. A dynamic strategy may assign treatment from current measured history, such as treating when a biomarker crosses a threshold. State thresholds, grace periods, competing events and what happens when a variable is missing. “Treat as needed” is not a reproducible intervention.

3. Understand the g-formula as longitudinal standardisation

Back to contents

Factor the observed data into models for time-varying covariates, treatment-related processes and outcomes in their temporal order. Under consistency, sequential exchangeability, positivity and suitable model specification, replace the observed treatment mechanism with the chosen intervention and average predicted outcomes over the covariate histories that follow from it.

4. Fit models that respect variable type and history

Back to contents

Use appropriate models for binary, continuous, count, categorical and survival variables, including relevant lags, interactions and nonlinearities. The goal is not merely good one-step prediction; small errors can compound across simulated periods. Avoid conditioning on future information. Preserve deterministic relations and structural zeros rather than asking a generic regression to generate impossible values.

5. Simulate many intervention-specific trajectories

Back to contents

Start from observed baseline covariates, then repeatedly draw each next covariate and outcome from fitted models while setting treatment according to the intervention. A large Monte Carlo sample reduces simulation noise but not model bias. Use reproducible seeds and enough draws that risk contrasts are stable. Keep Monte Carlo error separate from sampling uncertainty, commonly estimated by bootstrap.

6. Worked example: a dynamic threshold rule

Back to contents

A cohort records monthly blood pressure, medication and renal function. Prior medication changes both later pressure and renal function, which influence future treatment and cardiovascular risk. The team compares always treating, never treating and treating when systolic pressure exceeds 140 mmHg. They fit ordered history models, simulate each rule and estimate five-year risk. The dynamic rule looks best, but simulated natural-course renal values drift above the observed range. Revising nonlinear terms improves reproduction and changes the contrast. The intervention estimate becomes credible only after the natural course is repaired.

7. Use the natural course as a diagnostic, not a guarantee

Back to contents

Simulate the observed treatment process or an equivalent natural-course scenario and compare predicted outcome, covariate and event distributions with the data over time. Large divergence signals misspecification, coding errors or unsupported extrapolation. Close agreement is reassuring but cannot prove counterfactual models correct; errors may cancel or appear only under the intervention.

8. Changed-context transfer check: adaptive school support

Back to contents

A school offers tutoring when recent scores fall below a threshold. Earlier tutoring changes later scores, while scores guide future tutoring and final attainment. The transferable method is to encode the dynamic rule and simulate evolving scores. The changed context adds teacher discretion, peer spillover and term schedules. Those mechanisms must be measured or modelled; a clinical implementation cannot simply be renamed.

9. Know the assumptions and common misconceptions

Back to contents

The method does not solve unmeasured confounding, interference, poorly defined treatment or lack of positivity. Dynamic rules may request treatment choices rarely observed for certain histories, forcing extrapolation. Flexible models reduce but do not abolish misspecification. Monte Carlo simulation is computation, not randomisation, and thousands of simulated people do not increase the original information.

10. Audit interventions, models and simulations together

Back to contents

Publish the target trial, temporal diagram, variable definitions, intervention code, model formulas, diagnostics, random seeds, Monte Carlo size, bootstrap procedure and censoring or competing-event handling. Report natural-course checks and support by relevant history. Repeat analyses with alternative functional forms, thresholds and truncation or restriction to supported strategies.

11. Frequently asked questions

Back to contents

How does this differ from weighting? It models the full longitudinal data process and standardises simulated outcomes rather than constructing a weighted pseudo-population. Can it estimate dynamic rules? Yes, when rules and support are explicit. Does natural-course fit validate the result? No, but failure is a warning. Why use Monte Carlo? High-dimensional sums or integrals are approximated by simulated trajectories.

12. The gold standard makes the simulated world inspectable

Back to contents

A credible parametric g-formula analysis defines interventions that could be executed, models variables in their real time order, checks the simulated natural course and exposes where strategies rely on sparse data. The gold standard reports Monte Carlo and sampling uncertainty without mistaking either for identification. Its result is a transparent population-standardised contrast under stated assumptions, not a synthetic trial created from nothing.