What?
To show multiple adjusted effect estimates shown in a single table from a single regression model.
What are the consequences?
1. Interpretative complexities- create confusion in the interpretation of direct effect estimates to total effect estimates for covariates in the model.
2. Though the effect estimate of the main exposure are not confounded but overall effect estimate is confounded due to including all the covarites in the same model
3. Effect estimate may complicate further by heterogeneity (variation, modification) of the exposure effect estimate are presented
Why we need effect decomposition?:
To identify the interventional factors and understand the pathways to design for intervention to improve health and prevent diseases.
Few term before explaining the concept of effect decomposition:
Primary effect- effect of primary exposure of interest in the initial adjustment model
Secondary effect- effect of covariate (confounder or effect modifier) not of primary interest in the initial adjustment model
Total effect- net of net of all associations of a variable through all causal pathways to the
outcome
Direct effect- an association after blocking or controlling some of causal pathways. In a simple words, direct impact of an exposure on an outcome.
Indirect effect- Part of the exposure effect which is mediated by a given set of potential mediators. In a word, we can say it mediated effect.
Example:
Using logistic regression we will get the following model:
B1 can be interpreted as the conditional total effect of contracting HIV on the log odds of stroke. B2 cannot interpreted as the same way, it is a controlled direct effect of smoking on the log odds when HIV is fixed at a given level. B3 is also a controlled direct effect of age on the log odds when HIV and smoking are fixed at a given level.
To interpret β2 as a direct effect of smoking after blocking its effect on HIV infection, we must adjust for all confounders of both the exposure-outcome relationship and the
mediator-outcome relationship.
Next scenario- after adding a unmeasurable confounding factor U,
And the logistic regression model becomes-
Interpretation of B1 remains same but interpretation of B2 and B3 will change because B2 is now confounded by U and smoking is the collider of the path of Agw-Smoking-U-Stroke, so adjusting with smoking for age-stroke will open a backdoor pathway between U and stroke and biasing B3 as affect estimate.
Avoiding Table 2 fallacy:
Reasonable starting point - construct a causal diagram showing primary exposure, outcome, and potential covariates that encode the causal assumptions used to select covariates for inclusion in the model.
1. Avoid estimates of primary exposure effect measures in different models
2. using different covariate subsets would allow Table 2 to include estimates of total effects of secondary covariates
3. definitions of direct and indirect effects involve combined interventions on both the exposure and mediators; some combinations may resemble nothing anyone would consider in reality, thus violating positivity constraints
Comments
Post a Comment