Causa Nostra: The Potentially Legitimate Business of Drawing Causal Inferences From Observational Data

• Published in: CPT: pharmacometrics and systems pharmacology Vol. 8; no. 5; pp. 253 - 255
• Main Author: Rogers, James A.
• Format: Journal Article
• Language: English
• Published: United States John Wiley & Sons, Inc 01-05-2019; John Wiley and Sons Inc; Wiley
• Subjects: Alcohol; Estimates; Population; Simulation; Surgery
• ISSN: 2163-8306; 2163-8306
• DOI: 10.1002/psp4.12395

(In this simple example, resampling is equivalent to parametric Bernoulli sampling, but the intention here is to hew as closely as possible to a typical pharmacometric workflow.) 2.2.For every patient in the virtual population, fix treatment at one level, e.g., “open surgery.” 2.3.Use the estimated‐response model to simulate a response value for each member of the virtual population as a function of that patient's covariate values (in this example, kidney stone size) and treatment assignment. 2.4.Compute the average probability of success in the simulated population. 2.5.Repeat all of the above steps, replacing the treatment in step 2.2 with the comparative intervention (in this example, “percutaneous nephrolithotomy”). In G‐computation, step 2.1 of the previous procedure would not involve resampling to create a large “virtual population” but would instead involve the creation of a “virtual sample,” having exactly the same sample size and exactly the same covariate distribution as in the original data set In G‐computation, step 2.3 of the previous procedure would not involve the introduction of stochastic elements and would replace the simulated responses with conditional predictions of the response These two differences are motivated by the same consideration, namely, that pharmacometric response models are typically nonlinear mixed‐effects models for which predicted values are not directly available as a function of model parameters. Causal Concepts and Causal Diagrams As the previous example is intended to illustrate, it would hardly be novel to propose that the law of total probability be applied in the context of pharmacometric simulation to properly average over covariate distributions. [...]to do so would be to merely advance various probabilistic and statistical methods, which misses the point of formal causal reasoning entirely.