Most of the standard advice in data work amounts to “just train a model on historical data,” and it quietly wrecks decision quality. Models built on association are good at spotting patterns and predicting what comes next. They are much worse at a question like whether cutting the price of a product will cause more sales. That question belongs to Causal ML, which deals with how the system behaves rather than what tends to happen in it.
I have watched plenty of “data-driven” features fail in production, and mixing up correlation with causation is usually somewhere in the wreckage. Causal methods add the “what if” part: estimating the outcome of an action nobody has taken yet, with confounding variables separated out from the effect you care about.
The potential outcomes framework
A causal study is not aimed at loss minimization. What you want is the Average Treatment Effect, and the potential outcomes framework is how you write it down. Every unit has two states: Y(1), the outcome if the treatment was applied, and Y(0), the outcome if it was not. For any single individual you only ever observe one of them. That missing half is the whole difficulty of causal inference.
If the data did not come from a randomized controlled trial (RCT), the groups are not exchangeable. Picture a doctor who prescribes heart medication only to the sickest patients. Feed that history to a standard model and it will report that the drug causes heart attacks. The selection bias has to come out of the data before any coefficient means anything.
Drawing the logic with DAGs
Directed acyclic graphs (DAGs) are the blueprint for a causal analysis. Variables are nodes and arrows are one-way causal flows. A confounder gives itself away by pointing at both the treatment and the outcome. Leave those backdoor paths unaccounted for and your model’s coefficients are garbage.
If you want to see how these statistical foundations play out in bigger systems, I covered some of that in my senior dev insights on applied statistics. Once the DAG has told you which confounders matter, the next step is adjusting for them mathematically, usually with linear regression or matching.
De-biasing with OLS regression
Multiple linear regression is the first tool most people reach for here. Put the confounders in as independent variables and the model estimates the relationship you want while holding the rest constant. The example below uses Python and the statsmodels library to control for an “Initial Health” confounder.
import statsmodels.api as sm
import pandas as pd
# bbioon_estimate_causal_effect
# Assume df has 'severity', 'drug', and 'initial_health'
def bbioon_estimate_causal_effect(df):
X = df[['drug', 'initial_health']]
X = sm.add_constant(X)
y = df['severity']
model = sm.OLS(y, X).fit()
return model.summary()
# The coefficient for 'drug' is now the adjusted treatment effect.The number to read is the coefficient on the treatment variable. Raw data can point to a negative impact while the adjusted coefficient comes out positive. For more depth on these methods, Matheus Facure’s open-source material is worth the time.
Matching and propensity scoring
When confounders and outcomes do not relate linearly, regression stops working. Matching handles that case. Instead of fitting a formula, you search the control group for a twin of every treated individual. A propensity score, meaning the probability of receiving treatment, collapses high-dimensional data into a single metric you can search on.
Pairing individuals with similar propensity scores builds what people call a synthetic RCT. The groups become exchangeable, so you can compare outcomes directly. It is a data-restructuring step that forces balance onto your covariates.
Difference-in-Differences (DiD)
Some of the factors that matter were never recorded. If those factors are time-invariant, difference-in-differences gives you a way through. DiD compares two groups across two periods, before and after the treatment. You measure the change in the control group, assume the treated group would have moved by the same amount, and read whatever delta is left as the causal effect.
That cancels out the baseline differences between the groups. Time-varying confounders are the hazard. If a hospital upgrades its equipment at the same time it starts a new drug trial, the DiD estimator splits the credit between them and the estimate is polluted. Check your features for collinearity before you ship the analysis.
If this kind of analysis is eating your dev hours, I take on that work. I have been building on WordPress since the 4.x days and writing complex backend integrations since before AI was a buzzword.
Measuring instead of guessing
Real data is messy and a clean causal signal is rare. Knowing when to reach for OLS, matching, or DiD is what moves you from guessing to measuring real impact. Leaning on a prediction model while ignoring selection bias is technical debt you will pay for later. Causal ML only works when your variables genuinely support the adjustment you are making, so a good share of the effort goes into checking that they do.
