Hi. Let's say I am doing a causal inference analysis. It is using parametric…
- User
Hi. Let's say I am doing a causal inference analysis. It is using parametric g-computation to look at the effects of shifting an exposure in the whole population like weight or physical activity levels on population prevalence of say, osteoarthritis or knee pain. What are the (formal) assumptions of such an analysis?
- Assistant
Great question! When using parametric g-computation (also called the g-formula) for causal inference—especially to estimate the effect of shifting a continuous exposure (like weight or physical activity) on a population outcome (like osteoarthritis prevalence)—there are several formal assumptions you must make for your estimates to have a valid causal interpretation. These are similar to the assumptions underlying most causal inference methods, but with some nuances for g-computation and continuous exposures.
1. Consistency
- Definition: The observed outcome for each individual under their observed exposure is the same as the potential outcome under that exposure.
- Implication: If you set a person’s weight to what it actually is, their outcome is what you observed.
- Formally: ( Y = Y^{A} ) when ( A ) is the observed exposure.
2. Positivity (Experimental Treatment Assignment)
- Definition: For every combination of covariates, there is a positive probability of receiving every level of the exposure (or at least the levels you want to shift to).
- Implication: There are people at every relevant value of weight or activity, for every combination of confounders.
- Formally: ( P(A = a | L = l) > 0 ) for all ( a ) in the support of ( A ) and all ( l ) in the support of ( L ) (where ( L ) are confounders).
3. Conditional Exchangeability (No Unmeasured Confounding)
- Definition: Given measured covariates, the exposure is independent of the potential outcomes.
- Implication: All confounders of the exposure-outcome relationship are measured and included in your model.
- Formally: ( Y^{a} \perp!!!\perp A | L ) for all ( a ).
4. Correct Model Specification
- Definition: The models you use for the outcome (and sometimes the exposure) are correctly specified.
- Implication: If you use a parametric model (e.g., logistic regression), it must correctly capture the true relationship between exposure, covariates, and outcome.
- Formally: The functional form ( E[Y | A, L] ) is correctly specified.
5. Stable Unit Treatment Value Assumption (SUTVA)
- Definition: No interference between individuals and no hidden versions of treatment.
- Implication: One person’s exposure does not affect another’s outcome, and the exposure is well-defined.
- Formally: ( Y_i^{a} ) does not depend on ( a_j ) for ( j \neq i ).
Summary Table
Assumption Description Formal Statement Consistency Observed outcome = potential outcome under observed exposure ( Y = Y^{A} ) Positivity Nonzero probability of all exposure levels for all covariate strata ( P(A = a Exchangeability No unmeasured confounding given covariates ( Y^{a} \perp!!!\perp A Correct Model Specification Models for outcome (and exposure, if needed) are correct ( E[Y SUTVA No interference, well-defined exposure ( Y_i^{a} ) unaffected by ( a_j )
References
- Robins JM. A new approach to causal inference in mortality studies with a sustained exposure period—application to control of the healthy worker survivor effect. Math Modelling. 1986.
- Hernán MA, Robins JM. Causal Inference: What If. 2020. Free online
- Daniel RM, et al. Methods for dealing with time-dependent confounding. Stat Med. 2013.
If you want to discuss how these assumptions play out in your specific context (e.g., shifting weight in a population), let me know!
- User
thanks. Yeah, well, I was thinking for this analysis I would address these assumptions right up front, rather than tucking them away in a Discussion section. That is I would put in the methods section that they would be addressed for this particular analysis, and then I would do a table or something in the Results to show how they have been addressed. So what I was thinking is, what columns would I put in the table? So one column would be "Assumption", another would be "Description", another would be "What it means in terms of this analysis", the next would be "How this has been addressed in this analysis", and another would be "Has the assumption been met?". And the other thing I was thinking is what names should I use for each of the assumptions? I mean, SUTVA is sometimes called "non-interference", and consistency is sometimes called "well-defined" intervention, and so on.
- Assistant
This is a great approach—being explicit and transparent about assumptions up front is best practice in causal inference, and a table is a clear way to communicate this. Your proposed columns are excellent and will help readers understand both the general and study-specific implications of each assumption.
Column Suggestions
Your columns are well-chosen. Here’s a refined version:
- Assumption (Name)
- Description (General definition)
- What it means in this analysis (Contextualized for your study)
- How addressed in this analysis (What you did to address it)
- Has the assumption been met? (Your assessment: Yes/No/Partially/Unclear, with brief justification)
You could also consider a "References/Notes" column if you want to cite sources or add clarifying comments, but that’s optional.
Naming the Assumptions
You’re right: there are multiple names for some assumptions. Here’s a table of commonly used names and suggestions for what to use (with alternatives in parentheses):
Preferred Name Alternatives Consistency Well-defined intervention Positivity Experimental treatment assignment Exchangeability No unmeasured confounding, ignorability SUTVA Non-interference Correct model specification Model correctness, no model misspecification My suggestion: Use the most widely recognized names, but you can include the alternative in parentheses if you wish.
Example Table Structure
Assumption Description What it means in this analysis How addressed in this analysis Has the assumption been met? Consistency (Well-defined intervention) The observed outcome under the observed exposure equals the potential outcome under that exposure. The outcome (e.g., knee pain) for each person is the same as if we had set their weight/activity to the observed value. The exposure (weight/activity) is clearly defined and measured consistently. Likely met; exposure is well-defined and measured. Positivity Every individual has a nonzero probability of receiving each exposure level, given their covariates. There are people at all relevant levels of weight/activity for all confounder strata. Checked distribution of exposure across covariate strata; excluded rare strata if necessary. Mostly met; some extreme values rare, but main analysis focuses on common ranges. Exchangeability (No unmeasured confounding) Given measured covariates, exposure is independent of potential outcomes. All confounders of weight/activity and outcome are measured and included. Adjusted for age, sex, comorbidities, etc. in models. Partially met; residual confounding possible. SUTVA (Non-interference) No interference between individuals; no hidden versions of exposure. One person’s weight/activity does not affect another’s outcome. Assumed based on study design; no evidence of interference. Likely met. Correct model specification The statistical model for the outcome is correctly specified. The model accurately captures the relationship between exposure, covariates, and outcome. Used flexible modeling (e.g., splines); checked model fit. Reasonably met; some risk of misspecification remains.
Tips
- In the "Has the assumption been met?" column, be honest about uncertainty—“Partially met” or “Unclear” is often appropriate.
- In "How addressed," be specific (e.g., “Adjusted for X, Y, Z; checked overlap of covariates and exposure”).
- In "What it means in this analysis," use plain language and refer to your specific exposure/outcome.
Let me know if you want a more detailed example for your specific exposures/outcomes, or if you want a template you can copy-paste!
Other models, same or similar prompt
Responses were produced in different, uncontrolled contexts (system prompts, settings, prior turns). This is not a benchmark.