The two most widely used models in business analytics share a surname and a lot of machinery, but they answer different kinds of question. The choice between them is decided almost entirely by what you are trying to predict.
| Linear regression | Logistic regression | |
|---|---|---|
| Outcome | A number on a continuous scale | A yes/no event |
| Prediction | The expected value | The probability of “yes” |
| Business examples | Next month’s spend per customer, delivery time, units sold, salary for a role | Will this customer churn, will this loan default, will this lead convert, is this claim fraudulent |
| Coefficient reading | Change in the outcome per unit change in the input | Change in the log-odds; exponentiated, a multiplier on the odds |
| Fitting method | Least squares | Maximum likelihood |
Why not just use linear regression for yes/no?
You can code churn as 1 and stay as 0 and fit a straight line. This is called a linear probability model, and economists sometimes use it because its coefficients are easy to read. For prediction, though, it has an obvious defect: a straight line keeps going. For a customer with very low engagement it might predict a churn “probability” of 1.3, and for a very engaged one, minus 0.2. Neither makes sense.
Logistic regression solves this by passing the linear combination of inputs through an S-shaped curve, the logistic or sigmoid function, which squeezes any number into the range between 0 and 1. Very negative scores map close to 0, very positive scores close to 1, and a score of 0 maps to exactly 0.5.
Odds, log-odds and what a coefficient means
The linear part of a logistic model predicts the log of the odds. Odds are the probability of an event divided by the probability of it not happening. A 10% churn probability means odds of 0.1 divided by 0.9, roughly 0.11, or “1 to 9”. A 50% probability means odds of 1.
Suppose a churn model for a broadband provider gives the input “number of complaints in the last 90 days” a coefficient of 0.4. That means each extra complaint adds 0.4 to the log-odds of churning. Exponentiate it and you get about 1.49: each complaint multiplies the odds of churn by roughly 1.5. This exponentiated value is the odds ratio, and it is the most common way to report logistic coefficients.
The important subtlety is that a constant odds ratio does not mean a constant change in probability. Watch what one extra complaint does at different starting points:
- A customer at 10% churn probability moves to about 14.2%, an increase of about 4 percentage points.
- A customer at 50% moves to about 59.9%, an increase of about 10 points.
- A customer at 90% moves to about 93.1%, an increase of about 3 points.
The same factor has its largest effect on customers who are on the fence. That is one reason retention teams often target middle-risk customers rather than the highest-risk ones: the highest-risk group may be almost certain to leave regardless.
Linear regression is simpler to read. If a model of monthly spend gives “number of app sessions” a coefficient of ₹35, then each extra session is associated with ₹35 more spend, the same at any level of spend.
From probability to decision: the threshold
A logistic model outputs probabilities. A business process often needs a yes or no: call this customer or not, approve this loan or not. That conversion requires a threshold, and 0.5 is a lazy default rather than a principled choice.
The right threshold depends on the costs of the two kinds of mistakes. If a retention call costs ₹50 and saving a customer is worth ₹3,000 of margin, it is worth calling customers with even a modest churn probability, provided the call actually changes behaviour for some of them. For loan approval, where a default can cost lakhs, the threshold for approving will be strict. Plot precision and recall, or expected profit, across thresholds and choose with the business owner, not alone.
This also means the evaluation metrics differ. Linear models are judged on error in the outcome’s own units, such as mean absolute error in rupees, and on R-squared. Logistic models are judged on how well they rank cases (area under the ROC curve), how accurate their probabilities are (calibration), and how they perform at the chosen threshold (precision, recall, confusion matrix). Plain accuracy is misleading when the event is rare. If 3% of transactions are fraudulent, a model that says “not fraud” every time is 97% accurate and completely useless.
Shared assumptions, shared pitfalls
Both models assume the relationship with each input is linear on their own scale (the outcome for linear regression, the log-odds for logistic). If the effect of tenure on churn is high in the first three months, low in the middle and rising again at contract renewal, a single coefficient will average that away. Binning the input or adding terms for curvature helps.
Both suffer from overlapping inputs, omitted variables and the gap between association and causation. A churn model that finds “customers who called support churn more” has not shown that support calls cause churn; it may be that unhappy customers call support.
And both are worth trying before anything fancier. A gradient-boosted model may predict churn a little better, but a logistic regression whose coefficients the retention team can read and challenge is often the model that actually gets used.
Beyond the two
Some outcomes fit neither model cleanly. Counts, such as number of support tickets per month, are better handled by Poisson or negative binomial regression. Outcomes with more than two categories, such as which of four plans a customer picks, need multinomial logistic regression. Time until an event, such as days until churn, calls for survival analysis. Recognising the type of outcome is the first modelling decision, and it is the one that most determines whether everything after it makes sense.
Written by the RP Strativa Team. Coefficients, costs and the broadband example are illustrative.
Leave a Reply