Weeks 3
SPPH 604 Discussions
Confounding and Introduction
Michael Asamoah-Boaheng
Outline
• Discussion of Week 2 class activity (Expected answers)—Max 15 minutes
• Answering questions on Week 3 materials/Summary of Methods—Max 35 minutes
• Discussing of article and critique (Application of week 3 methods)—Max-50 minutes
• Discussion on the Proposal/SAP submissions, guidelines/Up-coming deadlines—50
minutes
• Overview of Week 4 Materials–10 minutes
2
• Discussion of Week 2 class activity
3
Questions/discussions on pre-reading
materials for Weeks 3
• Could we discuss further the utility and the usefulness of the potential
outcome framework?
4
Counterfactual Framework
• The potential outcome framework/the counterfactual framework/Rubin-Causal-
Model (RCM) is used in the field of causal inference to estimate and analyze causal
effects in observational studies and experiments.
• The framework considers what would have happened to individual under different
treatment or exposure conditions.
• Researchers collect data on actual treatment assignment and observed outcomes for
each individual in the study.
5
• In real life situation, we cannot observe the outcomes for the same person
under two different conditions and can only observe one potential outcome.
• The potential outcome framework estimates causal effect by comparing the
observed outcomes to the counterfactual outcomes.
• It relies on an assumption that the potential outcomes for one unit are not
influenced by treatment assignment by other units. That is the groups should
be comparable/exchangeable/ignorable treatment assignment.
• In other words, it assumes that the treatment assignment is not influenced by
unobserved factors that could also affect the outcome. 6
• So, when this assumption holds, we can estimate or quantify causal effects:
• Notations: given an exposure A, and outcome Y, and with measured
covariates C.
• Let Ya denote the counterfactual outcome or potential outcome that would
have been observed for an individual who had the exposure A.
• We can say that Covariates C is enough to control for confounding if the
counterfactuals Ya are independent of A conditional on C.
7
• Defined as the contrast of counterfactual outcomes, using observed data:
E(Y1-Y0|c) = E(Y|A=1,c)-E(Y|A=0, c)
• Usefulness of the Potential Outcome framework
• To investigate and estimate causal relationships between exposures (such as
treatments or interventions) and outcomes (e.g., health outcomes, policy
effects).
8
• To do comparative analysis. This comparative analysis is essential for
establishing causality and quantifying the effect of an exposure.
• Helps researchers to identify and control for confounding variables by
explicitly considering counterfactual scenarios
9
Principles for confounder selection
• Pre-treatment Criterion: control for any variable that is prior to the
treatment or exposure under study.
• One problem that arises with the “pre-treatment” approach is that in
principle one may end up controlling for a pre-exposure covariate that in
fact introduces bias (for instance when the pre-exposure variable is a
collider)
10
• Common cause” approach: requires relatively minimal knowledge of the
underlying causal structure.
• Adjusts for all pre-exposure covariates that are common causes of
exposure and outcome.
• Limitation : The downside of this criterion is that in certain instances, if
data on some of the covariates that are common causes of the exposure
and the outcome are not available, there might be a different set of
covariates that suffices to control for confounding.
11
• Disjunctive cause criterion: This criterion does not require knowledge of the
full underlying causal diagram relating each of the covariates to all of the other
covariates.
• Adjust for any pre-exposure covariate that is a cause of the exposure, or the
outcome, or both.
• Limitation: The presence of instrumental bias can amplify bias (IV is a cause of
the exposure, but has no relation to the outcome except through the exposure)
12
• Modified disjunctive cause criterion: A practical and theoretically informed
approach to covariate selection would.
• Control for each covariate that is a cause of the exposure, or of the
outcome, or of both; exclude from this set any variable known to be an
instrumental variable; and include as a covariate any proxy for an unmeasured
variable that is a common cause of both the exposure and the outcome.
13
Statistical confounder selection
• Forward and backward selection.
• “Change-in-estimate” approach.
• “High-dimensional propensity score”.
• Targeted maximum likelihood estimation which uses machine learning algorithms.
14
Discussing of article and critique
15
Discussion on the Proposal/SAP
submissions, guidelines/Up-coming
deadlines
16
Overview of Week 4 Materials
• Method: Lecture: Adjustment according to goal
• Scientific Writing: Methods section
• Questions to “Wall of Confusion week 4”: [*At least one question from each group]
• Please let me know ASAP if any of the links are not working.
• Please NOTE that the only link which will NOT be accessible will be the FAQs. It will be
available every Tuesday Morning before class.
17
• Deadlines for submission of Proposal/SAP: Monday September 25, by 4pm. (Pls send it to
my email)à(michael.boaheng@ubc.ca).
• Proposal defense on Thursday September 28, 2023: From 9 am to 12:00 noon. See schedule
for the defense on Canvas.
• Peer review for SAP submission due on October 3, 2023 @ 4pm (Instructions will follow
be made available next week).
• No Lab class next week: Thursday. The time will be used for Proposal defense.
18
THANK YOU
Email: michael.boaheng@ubc.ca
19
