Propensity Pre-Balancing on the Target Variable

Some data sets are like a party where one group eats all the pizza. The other group gets one sad olive. This is common when we model a target variable. One class may be huge. The other may be tiny. Propensity pre-balancing is a way to make the party more fair before the real modeling begins.

What does this mean in plain words?

Let us slow down. The phrase sounds scary. It is not.

A target variable is the thing you want to predict or study. It might be:

• Will a customer churn?
• Will a patient recover?
• Will a loan default?
• Will a user click?
• Will a machine break?

The target often has groups. For example, churn and no churn. Or default and no default.

Now imagine your churn group is tiny. Only 5 out of 100 customers churn. Your model may get lazy. It may learn to say “no churn” all day. It will look accurate. But it will be useless.

That is the classic target imbalance problem.

But there is another problem too. The groups may be different in many ways. Maybe churners are younger. Maybe they have fewer support tickets. Maybe they live in one region. Maybe they bought a cheaper plan. These differences can confuse your learning process.

Propensity pre-balancing tries to fix this before the main task. It asks a simple question:

Based on the features, how likely is each row to belong to a target group?

That “how likely” number is the propensity score.

The tiny robot bouncer idea

Think of a tiny robot bouncer at a data club. Each row wants to enter. The robot looks at the row’s features. It says, “Hmm. You look like someone who might churn.” Or, “You look like someone who probably will not churn.”

The robot gives each row a score from 0 to 1.

• 0.05 means “very unlikely to be in the target group.”
• 0.50 means “could go either way.”
• 0.95 means “very likely to be in the target group.”

Then we use those scores to balance the data. We can match rows. We can weight rows. We can sample rows. The goal is simple.

Make the target groups more comparable.

Not identical. That would be magic. Just more fair.

Why do this before modeling?

Because raw data can be messy. Really messy. Like spaghetti in a backpack.

If one target group is very different from the other, your model may learn shortcuts. It may learn patterns that are not meaningful. It may say:

“People on the cheap plan churn.”

But maybe that is not the full story. Maybe cheap plan users are also newer users. Maybe they had less onboarding. Maybe they live in a region with poor service. The target difference is tangled with many other features.

Pre-balancing helps untangle the knot.

It can help in several ways:

• It can reduce bias.
• It can make comparisons cleaner.
• It can improve learning for rare target groups.
• It can stop big groups from dominating small groups.
• It can make model behavior easier to inspect.

It is not a magic wand. It is more like a broom. It tidies the floor before you dance.

A simple example

Suppose you run a streaming app. You want to predict churn. Your data has 10,000 users.

• 9,500 users did not churn.
• 500 users did churn.

That is not balanced. The churn group is small.

You have features like:

• Age
• Plan type
• Watch hours
• Support tickets
• Days since signup
• Number of devices

First, you build a simple model. Its job is not the final prediction. Its job is only to estimate the propensity of being in the churn group.

So it predicts:

“Given these features, how likely is this user to be a churner?”

Now every user has a propensity score.

Next, you balance. You may match churners with non-churners who have similar scores. For example, a churner with a score of 0.72 gets matched with a non-churner with a score of 0.71. They look similar, based on features. But their target outcomes differ.

That is useful. Now the comparison is more apples to apples.

Three common ways to pre-balance

There are many methods. But three are especially common.

1. Matching

Matching pairs rows from different target groups that have similar propensity scores.

It is like finding data twins.

One churner. One non-churner. Similar background. Different outcome.

This can create a cleaner training set. It can also shrink the data. You may drop rows that do not have a good match.

2. Weighting

Weighting keeps more rows. But it changes how much each row matters.

A rare or important row may get a bigger weight. A common row may get a smaller weight.

It is like turning up the volume on quiet voices. The small target group finally gets heard.

3. Sampling

Sampling means you pick rows in a smart way. You may undersample the giant group. You may oversample the small group. You may use propensity scores to choose which rows are most useful.

This is like curating a playlist. You do not need every song. You need the right mix.

Where the “target variable” part gets tricky

This is important. Please put on your tiny safety helmet.

When you use the target variable to guide balancing, you must avoid target leakage.

Leakage happens when information from the answer sneaks into the training process in a way that would not exist in real life. It is like giving students the answer key before the test. Scores go up. Reality goes down.

Propensity pre-balancing uses the target group to build the balancing plan. That can be okay. But you must do it only inside the training workflow.

Do not balance the full data set before splitting into train and test. That can leak information from the test set into training.

A safer flow is:

1. Split the data into training and test sets.
2. Use only the training set to estimate propensity scores.
3. Balance only the training set.
4. Train your final model on the balanced training set.
5. Evaluate on the untouched test set.

The test set should stay boring. Untouched. Pure. Like a sealed snack.

Propensity score does not mean probability of truth

A propensity score looks like a probability. But do not worship it.

It is estimated by a model. That model may be wrong. It may miss important features. It may be too simple. It may be too complex. It may behave like a raccoon with a calculator.

So you should check the balance after using the scores.

Ask:

• Are feature distributions more similar now?
• Did the tiny target group get better representation?
• Did we lose too much data?
• Are extreme weights causing chaos?
• Does performance improve on untouched data?

The goal is not a pretty score. The goal is a fairer learning setup.

How to know if balancing worked

You need diagnostics. This sounds fancy. It mostly means “look before you leap.”

Compare the target groups before and after balancing. Look at key features. For each feature, ask if the groups are still wildly different.

For example, before balancing:

• Churners average 18 watch hours.
• Non-churners average 80 watch hours.

After balancing:

• Churners average 22 watch hours.
• Non-churners average 25 watch hours.

That is much closer. Nice.

You can also look at charts. Histograms are helpful. Box plots are helpful. Love them like little data sandwiches.

When should you use this?

Propensity pre-balancing is useful when target groups are not just unequal in size, but also uneven in feature mix.

It can help when:

• The rare target class matters a lot.
• You want cleaner group comparisons.
• Your model is learning lazy shortcuts.
• You need better fairness across subgroups.
• You are studying outcomes, not just chasing accuracy.

It is common in causal thinking. It is also useful in predictive modeling. But the goal may differ.

In causal analysis, you care about fair comparisons. You want groups to look alike, except for the thing being studied.

In machine learning, you may care about improving learning. You want the model to see enough useful cases from each target group.

Both goals are valid. Just be clear about which game you are playing.

When should you avoid it?

Do not use it blindly.

Avoid it when:

• You have very little data.
• Your propensity model is poor.
• The groups have almost no overlap.
• You cannot explain the weighting or sampling choices.
• You are accidentally using future information.

Overlap matters a lot. If churners and non-churners are totally different, matching may fail. Imagine trying to match penguins with bicycles. It is not a match. It is a cartoon.

Bad overlap means the data cannot support fair comparison in some areas. In that case, you may need better data. Or a narrower question.

A simple recipe

Here is a friendly process.

1. Define the target. Be precise. Know what “positive” means.
2. Choose features. Use only features available before prediction time.
3. Split the data. Keep test data untouched.
4. Fit a propensity model. Use training data only.
5. Create scores. Each row gets a target propensity score.
6. Balance the training set. Use matching, weighting, or sampling.
7. Check balance. Compare features before and after.
8. Train the final model. Use the balanced training data.
9. Evaluate honestly. Use the original test distribution.

That last step matters. Real life will not be balanced just because your training set is. Your model must face the wild world.

A quick note on fairness

Pre-balancing can help with fairness. But it can also hurt fairness if done carelessly.

For example, if some demographic group is underrepresented in the target class, balancing may remove even more of them. Or weighting may amplify noisy patterns.

So check performance across important groups. Do not only check one big metric. Accuracy can smile while fairness cries in the corner.

Use group metrics. Use error rates. Use calibration checks. Keep humans in the loop.

The big idea

Propensity pre-balancing on the target variable is a way to make data less lopsided before learning from it. It gives each row a score. That score says how likely the row is to belong to a target group, based on its features. Then you use the score to match, weight, or sample rows.

The result is a training set that is often more balanced and less biased. It can make models smarter. It can make comparisons cleaner. It can make rare target groups less invisible.

But it needs care. Split first. Balance second. Test honestly. Check your work.

Think of it like setting the table before dinner. You are not cooking the whole meal yet. You are making sure everyone has a plate. And in data science, that is already a very good start.