When we want to understand whether an action causes an outcome, the cleanest approach is a randomised experiment. In many real settings, that is not possible. Businesses cannot randomly “assign” customers to receive a premium experience, hospitals cannot randomly assign every patient to a treatment, and policy teams cannot always randomise interventions across locations. This is where propensity scoring becomes useful.
Propensity scoring is a practical technique used in observational studies to estimate the probability that a subject receives a treatment, given their observed characteristics. Once we can model that probability, we can reduce selection bias and make fairer comparisons between treated and untreated groups. If you are learning applied analytics through a data analytics course in Kolkata, propensity scoring is one of the most valuable tools for moving from simple correlation to stronger causal reasoning.
What Is a Propensity Score?
A propensity score is the estimated probability that an individual receives a treatment (or intervention) based on observed covariates. Formally, it is:
- Propensity score = P(Treatment = 1 | Observed features)
Here, “treatment” can mean many things depending on the problem:
- A marketing offer sent vs not sent
- A customer enrolled in a loyalty programme vs not enrolled
- A patient receiving a new drug vs standard care
- A student attending an extra coaching module vs not attending
The key challenge in observational data is that treated and untreated groups often differ systematically. For example, high-value customers may be more likely to receive premium support, and they may also naturally have higher retention. If we compare outcomes directly, we might wrongly attribute retention differences to premium support rather than the customer’s baseline value. Propensity scores help address this imbalance by creating more comparable groups.
How Do We Estimate Propensity Scores?
In practice, estimating propensity scores means building a model that predicts treatment assignment from pre-treatment variables. The most common method is logistic regression, but machine learning models (random forests, gradient boosting) can also be used if handled carefully.
A typical workflow looks like this:
- Define the treatment clearly
Example: “Received discount coupon in the last 30 days” (yes/no). - Select covariates that influence treatment assignment
Include variables that affect whether a subject gets the treatment and are measured before treatment happens. Examples: past purchases, tenure, demographics, prior engagement. - Fit a model to predict treatment
The model output probability for each subject becomes the propensity score. - Check overlap (common support)
If treated customers have propensity scores near 1 while untreated customers have scores near 0, comparisons become unreliable because the groups do not overlap enough.
A strong analytics team treats propensity score modelling as a design step, not a final answer. The goal is not high predictive accuracy; the goal is balancing the groups so outcome comparisons become fairer.
What Do We Do With Propensity Scores?
After estimating propensity scores, we use them to adjust comparisons between treated and untreated subjects. Common approaches include:
1) Matching
Pair treated subjects with untreated subjects who have similar propensity scores. This creates a matched sample where covariates should be more balanced. Matching is intuitive and easy to communicate, but it can discard data if good matches are not available.
2) Stratification (Subclassification)
Divide the data into bins (for example, quintiles) based on propensity scores. Compare outcomes within each bin and aggregate results. This is simple and works well when there is good overlap.
3) Weighting (IPTW)
Inverse Probability of Treatment Weighting gives more weight to subjects who are underrepresented in their group. The idea is to create a “pseudo-population” where treatment is independent of observed covariates. Weighting can be powerful, but extreme weights can destabilise results, so trimming or stabilising weights is often necessary.
4) Covariate Adjustment Using the Propensity Score
Use the propensity score as a control variable in a regression model of the outcome. This is straightforward, though it relies on model assumptions and may not balance covariates as well as matching or weighting.
In applied learning environments-such as a data analytics course in Kolkata-you often practise more than one method and compare results to understand robustness.
Diagnostics and Common Pitfalls
Propensity scoring is not a magic fix. Its reliability depends on disciplined checks and realistic assumptions.
- Balance checks are mandatory: After matching/weighting, verify whether covariates are balanced (standardised mean differences are commonly used).
- No unmeasured confounding: Propensity scoring only accounts for observed variables. If key drivers are missing (for example, motivation, risk appetite, physician judgement), bias can remain.
- Avoid post-treatment variables: Do not include features that are influenced by the treatment, or you may introduce bias.
- Confirm overlap: If overlap is weak, causal claims become shaky. Sometimes the right conclusion is: “We cannot estimate the effect reliably for this population.”
Conclusion
Propensity scoring helps analysts estimate and correct for selection bias in observational data by modelling the probability of treatment assignment. When used carefully-with proper covariates, strong overlap, balanced diagnostics, and transparent reporting-it strengthens the credibility of impact estimates in marketing, product, healthcare, and policy. For anyone building causal thinking skills through a data analytics course in Kolkata, mastering propensity scores is a practical step toward making decisions based on evidence rather than misleading comparisons.
Comments are closed.