Repository files navigation

Causal mediation analysis estimating interventional effects mapped toa target trial

The R packagemedRCT for causal mediation analysis supports theestimation of interventional effects (VanderWeele, Vansteelandt, andRobins 2014), specifically interventional effects that are defined suchthat they map explicitly to a “target trial” (Hernán and Robins 2016),as recently proposed by Moreno-Betancur et al. (2021). In the targettrial, the treatment strategies are specified to reflect hypotheticalinterventions targeting and thus shifting the joint distribution of themediators.medRCT can accommodate any number of potentially correlatedmediators, including mediators that are not of primary interest but thatare intermediate (exposure-induced) mediator-outcome confounders.

Causal mediation analysis generally seeks to investigate the extent towhich the causal effect of an exposure on an outcome is mediated throughintermediate variables. Natural (in)direct effects (Robins and Greenland1992; Pearl 2001) were initially proposed as the estimands of interestin these analyses. Natural effects are defined based on cross-worldcounterfactuals (Robins and Richardson 2011) and their identifiabilityrelies on a cross-world independence assumption. Given their reliance oncross-world counterfactuals, these effects have been criticized for notcapturing the effects of interventions or policy measures that could beconducted in the real world (Naimi, Kaufman, and MacLehose 2014).Further, the independence assumption required can never be guaranteed,even in an experiment (Robins and Richardson 2011; Didelez, Dawid, andGeneletti 2006), and it renders the estimands unidentifiable in thecommon settings of exposure-induced mediator-outcome confounding andmultiple mediators (Avin, Shpitser, and Pearl 2005; VanderWeele,Vansteelandt, and Robins 2014; Vansteelandt and VanderWeele 2012).However, in the context of multiple mediators, certain path-specificnatural effects, also defined in terms of cross-world counterfactuals,can still be identified and may be of substantive interest (VanderWeeleand Vansteelandt 2014).

Interventional effects have been proposed as an alternative to addressthese limitations. Firstly, these effects can be shown to map to ahypothetical randomized trial that evaluates the impact of hypotheticalinterventions shifting the distribution of the mediators(Moreno-Betancur and Carlin 2018). Secondly, interventional effectsremain identifiable in the presence of exposure-induced mediator-outcomeconfounding and multiple interrelated mediators of interest.

ThemedRCT package implements the estimation of interventional effectsthat are defined explicitly as effects in a hypothetical randomizedtrial (the target trial) , as proposed by Moreno-Betancur et al. (2021).This assists with clarifying the research question and ensuring that thestudy findings are meaningful and relevant to policy and practice. Inthe target trial, the treatment strategies are specified to reflecthypothetical interventions targeting and thus shifting the jointmediator distribution. ThemedRCT package implements the estimation ofinterventional effects that correspond to effects of hypotheticalinterventions which:

1. shift the joint distribution of all mediators under exposure to thatunder no exposure,

2. shift the distribution of a specific mediator under exposure, givenconfounders, to match the corresponding distribution under noexposure, independent of and without considering flow-on effects onother mediators,

3. shift the distribution of a specific mediator under exposure, givenconfounders, to match the corresponding distribution under noexposure, while considering flow-on effects on causally descendantmediators.

medRCT estimates these interventional effects using a Monte Carlosimulation-based g-computation approach. It should be noted that thismethod can be computationally intensive and is sensitive to modelmisspecification, as all nuisance parameters are estimated viarestrictive parametric models.

Researchers should consider usingmedRCT when their ultimate goal forconducting mediation analysis is to examine the effects of hypotheticalinterventions targeting multiple, potentially interdependent mediators.

ThemedRCT package is not yet available on CRAN. You can install thelatest stable version fromGitHubusing the following command:

remotes::install_github("T0ngChen/medRCT")

Using a simulated dataset based on a published case study from theLongitudinal Study of Australian Children (Goldfeld et al. 2023), weillustrate how to usemedRCT to estimate the interventional effectsthat emulate a target trial. Specifically, we aim to estimate thedifference in expected outcome (risk of child mental health problems)under exposure (low family socioeconomic position) with versus without ahypothetical intervention that individually shifts the distribution ofeach mediator (parental mental health and preschool attendance) to thelevels in the unexposed (high family socioeconomic position), whileaccounting for baseline confounders, an intermediate (exposure-induced)mediator-outcome confounder (family stressful life events), andcorrelations amongst mediators.

We begin by loading the library and dataset, and defining the confoundervector.

# Load the medRCT packagelibrary(medRCT)# Set a seed for reproducibilityset.seed(2025)# Display the first few rows of the datasethead(LSACdata)#>   child_sex child_atsi mat_cob mat_engl mat_age sep fam_stress parent_mh#> 1         0          1       0        0       1   0          0         0#> 2        NA          0       0        0      NA   0         NA        NA#> 3        NA          0       0        0      NA   0         NA        NA#> 4        NA          0       0        0      NA   0         NA        NA#> 5         1          0       0        0       1   1          0         0#> 6         1          0       0        0       1   0          1         1#>   preschool_att child_mh child_SDQscore#> 1             1        0       8.924660#> 2             0        0       7.349826#> 3             0        1      12.824643#> 4             0        0       6.611369#> 5             0        1      10.329341#> 6             0        1      13.552515# Define confounders for the analysisconfounders<- c("child_sex","child_atsi","mat_cob","mat_engl","mat_age")

Next we run the analyses, estimating interventional effects for ahypothetical intervention that shifts the distribution of each mediatorindividually.Note 1: the dataset has missing data. Incompleterecords are by default deleted before the analysis.Note 2: It isrecommended to perform the analysis with at least 200 Monte Carlosimulations by settingmcsim = 200. For illustration purposes, we usemcsim = 50, which takes approximately 90 seconds to run.

# Estimate interventional effects for a hypothetical intervention# that shifts the distribution of each mediator individuallymed_res<- medRCT(dat=LSACdata,exposure="sep",outcome="child_mh",mediators= c("parent_mh","preschool_att"),intermediate_confs="fam_stress",# intermediate confoundersconfounders=confounders,bootstrap=TRUE,intervention_type="shift_k",mcsim=50                          )#> Conducting complete case analysis, 2499 observations were excluded due to missing data.#> Note: It is recommended to run analysis with no fewer than 200 Monte Carlo simulations.# Summarise the resultssummary(med_res)#>#> Estimated interventional indirect effect:#>#>                      Estimate Std. Error  CI Lower  CI Upper p-value#> IIE_1 (p_trt - p_1)  0.011155   0.004181  0.002814  0.019203  0.0076#> IIE_2 (p_trt - p_2) -0.000763   0.002501 -0.005443  0.004362  0.7604#> TCE (p_trt - p_ctr)  0.128669   0.024554  0.082420  0.178668 1.6e-07#>#> Estimated interventional direct effect:#>#>                     Estimate Std. Error CI Lower CI Upper p-value#> IDE_1 (p_1 - p_ctr)   0.1175     0.0247   0.0712   0.1679 1.9e-06#> IDE_2 (p_2 - p_ctr)   0.1294     0.0244   0.0833   0.1789 1.1e-07#>#> Estimated expected outcome in each trial arm:#>#>       Estimate Std. Error CI Lower CI Upper p-value#> p_1     0.3302     0.0225   0.2872   0.3755  <2e-16#> p_2     0.3421     0.0221   0.2995   0.3862  <2e-16#> p_ctr   0.2127     0.0100   0.1922   0.2315  <2e-16#> p_trt   0.3413     0.0223   0.2987   0.3860  <2e-16#>#> Sample Size: 2608#>#> Simulations: 50#>#> Effect Measure: Risk Difference#> Results are based on all 100 bootstrap samples.

Based on the estimated interventional effect (IIE_1), a hypotheticalintervention improving the mental health of parents of children fromfamilies with low socioeconomic position to the levels of those fromfamilies with high socioeconomic position could potentially prevent 1per 100 cases of child mental health problems. Meanwhile, the effect ofa hypothetical intervention on preschool attendance (IIE_2) isnegligible.

For detailed guidance on using the package to handle more complexscenarios, please refer to thevignette.

For work involving themedRCT R package, please cite the following:

@article{Chen2025medRCT,  author = {Tong Chen and S. Ghazaleh Dashti and Margarita Moreno-Betancur},  title = {{medRCT}: Causal mediation analysis estimating interventional effects mapped to a target trial in {R}},  year = {2025},  doi = {10.21105/joss.08063},  url = {https://doi.org/10.21105/joss.08063},  journal = {Journal of Open Source Software},  volume = {10},  number = {110},  pages = {8063},  publisher = {The Open Journal}}@article{Moreno2021Mediation,   author={Margarita Moreno-Betancur and Paul Moran and Denise Becker and George C Patton and John B Carlin},   title={Mediation effects that emulate a target randomised trial: Simulation-based evaluation of ill-defined interventions on multiple mediators},   journal={Statistical Methods in Medical Research},   volume={30},   number={6},   pages={1395--1412},   year={2021},   URL={https://doi.org/10.1177/0962280221998409},   doi={10.1177/0962280221998409},   publisher={SAGE Publications Ltd}}