Logit regression, more commonly called logistic regression, is a statistical method used to predict outcomes that fall into categories, most often a simple yes or no. Will a patient develop a disease? Will a customer cancel their subscription? Will a lung nodule turn out to be cancerous? Unlike linear regression, which predicts a continuous number (like blood pressure or income), logit regression estimates the probability that something belongs to one group or another.
How It Differs From Linear Regression
The core distinction is the type of outcome you’re trying to predict. Linear regression works with continuous outcomes like days of hospitalization, test scores, or weight. Logit regression works with categorical outcomes: alive or dead, positive or negative, yes or no. This might sound like a small difference, but it changes the entire mathematical machinery underneath.
In linear regression, a coefficient tells you the average change in the outcome for each one-unit change in a predictor. If you’re predicting hospital stay length and the coefficient for age is 0.3, each additional year of age adds an average of 0.3 days. In logit regression, the coefficient instead represents a change in the odds of the outcome occurring. So rather than predicting a number on a scale, you’re predicting how likely something is to happen.
There’s also a practical problem that logit regression solves. If you tried to use linear regression for a yes/no outcome, you could end up with predicted values below 0 or above 1, which makes no sense as a probability. Logit regression avoids this entirely by using a mathematical function that forces every prediction to land between 0 and 1.
The Logit Function and the S-Curve
The “logit” in logit regression refers to a specific mathematical transformation: the natural logarithm of the odds. If the probability of something happening is p, the odds are p divided by (1 minus p). Taking the natural log of those odds gives you the logit. This logit value can range from negative infinity to positive infinity, which makes it compatible with the kind of linear equation used in standard regression.
On the other side of this equation is the sigmoid function (sometimes called the logistic curve), which does the reverse. It takes any number on that infinite scale and squashes it into a value between 0 and 1. Visually, it forms a stretched-out S shape: flat near 0 on the left, rising steeply through the middle, and flattening again near 1 on the right. This is what allows the model to output probabilities. A prediction of 0.92 means the model estimates a 92% chance of the outcome occurring. A prediction of 0.15 means 15%.
These two functions are inverses of each other. The logit converts a probability into a number the model can work with internally, and the sigmoid converts the model’s internal calculation back into a probability you can interpret.
Interpreting the Results: Odds Ratios
The raw output of a logit regression is a set of coefficients expressed as log-odds, which aren’t intuitive for most people. To make them useful, you convert them into odds ratios by raising the mathematical constant e to the power of the coefficient. An odds ratio tells you how much the odds of the outcome change for each one-unit increase in a predictor.
For example, if a model predicting heart disease produces an odds ratio of 1.4 for smoking status, that means smokers have 1.4 times the odds of developing heart disease compared to nonsmokers. An odds ratio of 1.0 means no effect. Values above 1 indicate increased odds, and values below 1 indicate decreased odds. For categorical predictors like sex, the odds ratio compares one group directly against a reference group (for instance, men versus women).
This is one of the reasons logit regression is so popular in health research. Odds ratios are easy to communicate to decision-makers and clinicians, even if the underlying math is complex.
Real-World Applications
Logit regression powers many of the clinical risk calculators that doctors use daily. The TREAT model, for instance, estimates the probability that a lung nodule is cancerous by combining patient demographics like age and sex, clinical data like BMI and history of chronic lung disease, symptoms like unexplained weight loss, and imaging findings. The ACS NSQIP Surgical Risk Calculator uses a similar approach to predict the likelihood of death or serious complications after surgery. The Mayo Clinic model and the Liverpool Lung Project model both use logistic regression to assess lung cancer risk in different patient populations.
Outside of medicine, logit regression shows up in credit scoring (will this borrower default?), marketing (will this customer buy?), and criminal justice (what’s the risk of reoffending?). Its appeal is that it produces a clean probability estimate from a mix of input variables, and its results are relatively transparent compared to more complex machine learning methods.
Assumptions the Model Requires
Logit regression is flexible, but it does require certain conditions to produce reliable results. The observations must be independent of each other, meaning one person’s outcome shouldn’t influence another’s. The relationship between each continuous predictor and the logit of the outcome should be roughly linear. The predictors shouldn’t be too highly correlated with one another (a problem called multicollinearity), and there shouldn’t be extreme outliers that disproportionately pull the model in one direction.
Notably, logit regression does not require the outcome to be normally distributed or the residuals to follow a bell curve, both of which are standard assumptions in linear regression. This makes it well-suited for the messy, categorical outcomes common in real-world data.
Measuring How Well the Model Works
Because logit regression produces probabilities rather than hard predictions, you need to pick a threshold to decide what counts as a “yes.” A common default is 0.5: anything above it gets classified as positive, anything below as negative. But the best threshold depends on the situation. In cancer screening, you might lower the threshold to catch more true positives, even at the cost of more false alarms.
The most widely used tool for evaluating a logit model’s performance is the ROC curve, which plots the true positive rate against the false positive rate across every possible threshold. The area under this curve (AUC) summarizes the model’s overall ability to distinguish between the two groups. An AUC of 0.5 means the model is no better than flipping a coin. An AUC of 1.0 means perfect discrimination. In practice, clinical models with AUC values above 0.8 are generally considered strong.
The AUC has an intuitive interpretation: it’s the probability that the model will rank a randomly chosen positive case higher than a randomly chosen negative case.
Beyond Two Categories: Multinomial Logit
Standard logit regression handles outcomes with exactly two levels. When the outcome has three or more categories that don’t have a natural order, you use multinomial logit regression instead. This extension applies the same underlying logic but estimates separate sets of coefficients for each possible outcome, all relative to a reference category. A classic example is modeling how people choose between three modes of transportation: car, bus, or train. Each option gets its own equation, and the model produces a probability for each one that together add up to 100%.
When the categories do have a natural order (like mild, moderate, and severe), a related approach called ordinal logistic regression is typically more appropriate, since it respects the ranking between groups.