Causal Machine Learning (CML)
Main idea for our research (public health): How will this disease progress if a patient is
given treatment X? Would this patient still have experienced outcome Z if treatment Y was
received?
Causality → understand how an intervention would affect a specific individual.
STEPS INVOLVED (but not limited to):
● Define a causal graph that models the assumed relationships between variables
(like this example one I made for NHIS data)
● Identify the treatment (the variable whose effect we want to measure) and potential
outcomes (the result we want to assess). This has to be done while controlling for
confounders (variables that affect both the treatment and the outcome)
Example from our current direction: For the NHIS dataset, we can start by estimating
the causal effect of comorbidities (cancer/ diabetes/ lung disease) on the
progression from intermittent to chronic pain (persistent pain). The treatment in this
case is the presence of a comorbidity, while the outcome is the change in pain status
between 2019 and 2020.
● Method: From my search, propensity score matching can be used to balance
treated (those with a comorbidity) and untreated (those without a comorbidity)
groups by their confounding characteristics (age, gender, access to healthcare etc).
● Model: After matching, models such as causal forests or doubly robust estimation
can be built to estimate the Average Treatment Effect (ATE) or the Conditional
Average Treatment Effect (CATE) for subgroups. [More info on this below, from the
Nature paper.]
● This approach involves comparing observed outcomes with counterfactual
scenarios, such as estimating how an individual’s pain status would change if they
had not developed a comorbidity. [I think this is the main difference between
traditional ML and causal ML.]
● Final validation can be done using sensitivity analysis (I have some ideas but I’m not
fully sure how to implement this – can check on it) to ensure the results are reliable.
Causal Machine Learning (CML)
Some more notes…(Feuerriegel, 2024)
Focus of CML → hypothetical scenarios where we predict outcomes under interventions
(also hypothetical).
Traditional Machine Learning → predict outcomes based on correlations
Causal ML → establish cause-and-effect relationships
Based on a causal quantity → changes in patient outcomes due to treatment
Evaluation of causal ML methods → assess the accuracy in predicting patient outcomes
from randomized data.
Key benefit → allows for estimating individualized treatment effects → clinical
decision-making can be personalized to individual patient profiles
Caution → causal inference rests on formal assumptions that cannot be tested
This is termed as the fundamental problem of causal inference: can only observe the
factual patient outcome under the given treatment, but never observe the counterfactual
patient outcome under a different, hypothetical treatment.
Very nice figure that describes the process:
Causal Machine Learning (CML)
Evaluation
Average Treatment Effect (ATE) → measures effects at the level of the study population.
By comparing the average patient outcome for those receiving the treatment versus those
who do not (control group), the ATE helps in understanding how effective a treatment is, on
average, across a specific patient cohort.
Example: when analyzing the comparative effectiveness of a new drug compared to the
standard of care, or when assessing the overall effectiveness or safety of a new drug.
Drawback: ATE cannot offer granular insights into whether patients with specific covariates
may particularly benefit from a treatment, even though such heterogeneity in treatment
effects can be of high interest in clinical practice.
Conditional Average Treatment Effect (CATE) → effect of a treatment for a particular
subgroup of patients defined by the covariates. Understanding the heterogeneity in
treatment effects informs subgroups where treatments are not effective or might even be
harmful, which is relevant for individualizing treatment recommendations to specific
patients.
Note: The estimation of ATE and CATE is often an easier task than predicting potential
outcomes → preferred when it is sufficient for decision-making.
