This repository builds upon the original population_airway_microbiome repository. All original files have been preserved in the original structure of the repository and all new work has been created under the directory mediation, which contains a power point presentation with a summary of the main criticisms raised after reviewing the associated original research article, which concern specifically to the mediation analyses reported in the paper. The mediation_IPW.qmd is the main code file that should be opened to reproduce analyses; scripts are sourced into this file, reason why individual scripts are not meant to be ran independently, but should be sourced into the workflow of the qmd file.
I dowloaded supplementary table 3 of the article, in which the proportion of the effect mediated by microbial features is reported for different exposures. Unfortunately, the table does not specify which outcome was used for each estimate reported. Thus, I assumed that these refer to the same outcome.
Directed acyclic graphs (DAG) were reconstructed based on the assumptions reported in the research paper, followed by testing the assumptions in the DAG against the sample dataset provided by the authors of the paper in their github repository, and updated according to the procedure described by Ankan, et al. This was followed by estimating the proportion of the effect mediated in the sample dataset with inverse probability weighting, according to assumptions in the updated DAG, for all exposure-mediator-outcome relationships reported in the original paper (Figure 4e) with the "Airway microbiome health index" as the mediator of interest. Assessment of mediation through inverse probability weighting (IPW) allows to account for complex shared exposure-outcome and mediator-outcome relationships (see paper by Martin Huber).
-
When selecting smoking as the exposure of interest, the total sum of proportion mediation was 3.836, which largely exceeds 1. Said in a different way, the total sum of proportion mediated should not be higher than 100%, but for smoking, mediation by microbial features is allegedly 383.6%.
-
The authors report in the research article that their estimates for proportion mediation are "relatively robust" according to rho values in sensitivity analyses (rho = 0.117, sd = 0.027). Nonetheless, in the paper by Chi et al, a rho value in medsens analyses is said to:
“ρ value is 0.1, indicating that the true ACME would not be significantly different from zero if there exist an unobserved confounder that causes a small correlation between errors for the mediator and the outcome models.”
This would suggest that the estimates are distant from being robust.
Lastly, reporting of the rho values should have been done in the table for every proportion mediated estimate, alongside its confidence interval. Pooling all rho values and reporting as the mean, SD does not allow to interpret robustness of individual estimates.
- Since these estimates are not robust as shown by sensitivity analyses, associations could be explained by residual exposure-outcome confounding, as shown and explained with the accompanying DAGs and the local test results showing that several of the conditional independencies do not hold in the sample dataset:
Figure description: Local tests plot results of the reconstructed DAG based on the assumptions reported in the paper. All confidence intervals that include 0 meet the conditional independecies assumption, whereas those that do not include the null would not meet this assumption. These suggest that adjustment for confounding in the paper could have been suboptimal
The assumptions in this DAG were updated to model with IPW. Here I show that after updating, the conditional independence assumption is consistent with the relationships found in the dataset:
Figure description: Local tests plot results of the updated DAG.
- The models with the continuous mediator AHMI were modelled with inverse probability weighting to estimate the bootstrap 95% confidence interval of proportion mediated while accounting for shared exposure-outcome and mediator-outcome confounding, showing that the estimates provided in the paper could be exagerated. Furthermore, all confidence intervals include 0 which would mean that it is possible that none of the effect of exposures on respiratory health is mediated by the microbiome. These results are shown in the following plot:
Figure description: Proportion of the effect of exposures on respiratory health mediated by the microbiome. The red bars represent the 95% confidence interval of the proportion mediated.
These analyses could be reproduced in the complete dataset. Code would need to be adapted to account for the effect of district.
Studying mediation in the context of this study through causal inference methods could not be appropriate due to violation of the no-interference assumption as it is conceivable that microbial features, including bacteria and fungi are transmitted between individuals.
Another issue is that the authors provide evidence for interaction between the exposures and mediators. Therefore, 4-way decomposition may be needed to reliably estimate mediation in the context of this study.