Logistic regression is one of the most widely used models in statistics and machine learning. But like any model, it comes with conditions, and ignoring them doesn’t make the problem disappear. It just makes it harder to find later.

Think of the assumptions as load-bearing walls. The model can look perfectly healthy from the outside even when one is violated. The problems show up downstream, in predictions that are confidently wrong, or in coefficients that point in the wrong direction entirely.

There are six core assumptions. For each one, here’s what it means in plain English, what a violation looks like in practice, and what you can do about it.

Assumption 1: The Response Variable is Binary

Logistic regression is designed for outcomes with exactly two categories. Pass or fail. Fraud or not fraud. Clicked or didn’t click. The math behind the model, specifically the logit transformation, assumes the outcome can take only one of two values, typically coded as 0 and 1.

What a violation looks like: Suppose you’re predicting which product a customer will buy from three options: Product A, Product B, or Product C. That’s a three-category outcome, not a binary one. Running standard logistic regression here forces a model structure onto a problem it wasn’t built for.

What to do: For three or more unordered categories, use multinomial logistic regression. For ordered categories where the categories have a natural ranking like low, medium, and high, use ordinal logistic regression.

Assumption 2: Observations are Independent

Each row in your dataset should represent a separate, unrelated individual or event. One observation shouldn’t influence or be linked to another. This breaks down whenever your data has a natural grouping structure or repeated measurements on the same subjects.

What a violation looks like: Medical data where the same patient appears multiple times across clinic visits. Or school data where students are nested within classrooms and share a teacher, a curriculum, and a physical environment. Their outcomes aren’t truly independent of one another.

What to do: Use mixed-effects logistic regression, sometimes called multilevel or hierarchical logistic regression. This accounts for the grouping structure and produces more accurate standard errors and predictions.

Assumption 3: No Multicollinearity Among Predictors

The predictor variables shouldn’t be highly correlated with each other. When two predictors move in near-perfect lockstep, the model can’t reliably separate their individual effects on the outcome. Coefficients become unstable, and small changes in the data can cause large swings in estimated effects.

What a violation looks like: Including both “height” and “shoe size” as predictors in the same model. These two variables are so strongly correlated that the model struggles to determine which one is actually driving the outcome.

What to do: Calculate Variance Inflation Factor (VIF) values for each predictor. A VIF above 5 is worth investigating; above 10 generally requires action. When multicollinearity is present, consider removing one of the correlated variables, combining them into a composite, or applying regularization methods like ridge regression.

Assumption 4: Linearity of the Log-Odds

This is the assumption that catches most people off guard. Logistic regression doesn’t assume a linear relationship between the predictors and the outcome probability. It assumes a linear relationship between the predictors and the log-odds of the outcome.

What a violation looks like: Consider age as a predictor of a health event. The underlying relationship may be non-linear: risk accumulates slowly in younger years, then rises steeply in older age. If the true log-odds curve bends but the model forces it straight, predictions will be systematically off at both ends of the age range.

What to do: Use the Box-Tidwell test to check this assumption for each continuous predictor. If it’s violated, a common fix is adding a polynomial term, for example including both age and age-squared as separate predictors. Generalized additive models can also handle curved relationships well.

Assumption 5: No Extreme Outliers

A small number of highly unusual data points can exert disproportionate influence on the model’s estimated coefficients, pulling them away from what the majority of the data supports. Logistic regression is particularly sensitive to extreme values in the predictor space.

What a violation looks like: In a model predicting loan default, one borrower with an income of $10 million, whether a data entry error or a genuine outlier, can distort the income coefficient for the entire dataset, making high income look either more or less protective than it actually is.

What to do: Calculate Cook’s Distance and leverage statistics for each observation to spot points with outsized influence. For each one, ask: is it a data entry error? A legitimate but rare case? A signal the model should capture? The answer determines whether to correct, remove, or keep the observation, and that decision should be documented.

Assumption 6: Large Enough Sample Size

Logistic regression needs enough observations in each outcome category to produce stable, reliable estimates. With too little data, the model overfits and performs poorly on new observations. A widely used rule of thumb is at least 10 events per predictor variable, often abbreviated as EPV.

What a violation looks like: A researcher building a model to predict a rare disease with 1% prevalence in a dataset of 200 patients and five predictor variables. That dataset contains roughly 2 positive cases per predictor, far too few for stable coefficient estimation. The model may produce extreme coefficients or fail to converge entirely.

What to do: If the sample size can’t be increased, consider reducing the number of predictors through subject-matter reasoning before modeling. Penalized regression methods, Firth’s logistic regression in particular, are designed for rare events and small samples and produce less biased estimates under these conditions.

Checking the Assumptions Before You Model

Think of these six assumptions as a pre-flight checklist. A pilot doesn’t run through it because they expect something to go wrong. They run through it because the cost of skipping it is too high if something does.

The same logic applies here. Most violations are detectable before or early in the modeling process, and most have workable solutions. The damage comes from not checking.

The above table provides a quick summary of this article and can be used as a cheat sheet the next time you are running a logistic regression on your data.