Mimicking counterfactual outcomes to estimate causal effects
This work is on causal statistical inference where the author investigates continuous time structural nested models (CT-SNM) for the estimation of effects of time varying treatment in the presence of time-dependent confounding by indication. The approach is applicable for the analysis of treatment outcomes in observational studies when treatment may be adapted to patient characteristics which predict the outcome. The author allows this so-called confounding by indication to take place after start of treatments and is such time-dependent, e.g. when treatment is initiated at any time or adapted at several times without fixed protocol in continuous time. The CT-SNM is a model for counterfactual outcomes and describes the relationship between the outcome \(Y^{(t)}\) of a treatment regime as given until time \(t\) and then changed to baseline and \(Y^{(t+h)}\) treatment as given until \(t+h\), with small \(h>0\), \(Y^{(t)}\) and \(Y^{(t+h)}\) generally not observable. The model base is the infinitesimal shift \[ D(y,t;\bar Z_t)=\frac{\partial}{\partial h}\bigg|_{h=0}\left(F^{-1}_{Y^{(t+h)}| \bar Z_t}\circ F_{Y^{(t)}| \bar Z_t}\right)(y), \] expressed as the derivative of the quantile-quantile transform which moves quantiles of the distribution of \(Y^{(t)}\) to quantiles of the distribution of \(Y^{(t+h)}\) given the covariate and treatment history \(\bar Z^t=\{Z(s);0\leq s\leq t\}\) and where \(F\) denotes the cumulative distribution of \(Y^{(t)}\) given the covariate and treatment history and \(F^{-1}\) its generalized inverse. The work is related to [the author, ibid. 36, No. 3, 1464--1507 (2008; Zbl 1360.62511)] and it extends [the author et al., Stat. Neerl. 58, No. 3, 271--295 (2004; Zbl 1059.62102)]. However, the author weakens the condition data that counterfactuals depend deterministically (the so-called local rank preservation) on the observed which is related to an assumption of constant treatment effects. She shows that for the CT-SNMs assumptions about the joint distribution of counterfactuals or the assumption of deterministic treatment effects is not necessary to mimic counterfactual outcomes measured at the end of the study or being survival outcomes. In fact she mimics counterfactuals by constructing mimicking random variables \(X(t)\) that are solutions of a differential equation that have the same distribution as the counterfactuals and that can be used to estimate parameters of the SNM without assuming the treatment effect to be deterministic. Thereby she shows that many previously published methods are robust to violation of the local rank preservation assumption and that a previous postulated conjecture on mimicking counterfactual outcomes holds. Besides a detailed model construction with historical examples, the main results are explicated showing that the mimicking variable \(X(t)\) defined as the solution of a differential equation with a final condition mimics \(Y^{(t)}\) in the sense that it has the same distribution given past treatment and covariate history. The assumption of no unmeasured confounding is formalized to be applicable for the proofs. Mimicking counterfactual survival data is outlines with an example for right censored data. A simulation study based on historical HIV/AIDS data demonstrates the performance of the approach.
- Statistical modeling of causal effects in continuous time
- Statistical models for estimating the effects of intermediate variables in the presence of time-dependent confounders.
- scientific article; zbMATH DE number 1487685
- Counterfactual analyses with graphical models based on local independence
- Structural nested cumulative failure time models to estimate the effects of interventions
- Causal inference for complex longitudinal data: the continuous case.
- Counterfactual analyses with graphical models based on local independence
- The stochastic system approach for estimating dynamic treatments effect
- Some considerations on the back door theorem and conditional randomization
- Statistical models for estimating the effects of intermediate variables in the presence of time-dependent confounders.
- Identifying and estimating net effects of treatments in sequential causal inference
- Structural Nested Models and Standard Software: A Mathematical Foundation through Partial Likelihood
- Semiparametric estimation of structural nested mean models with irregularly spaced longitudinal observations
- Simulation‐based estimators of analytically intractable causal effects
- Impact of discretization of the timeline for longitudinal causal inference methods
This page was built for publication: Mimicking counterfactual outcomes to estimate causal effects
Report a bug (only for logged in users!)Click here to report a bug for this page (MaRDI item Q2012196)
