The logistic regression assumptions are the short list of conditions your data must satisfy for the model’s coefficients, p-values, and predicted probabilities to be trustworthy — and they are noticeably gentler than the assumptions linear regression demands. This guide walks through all seven, explains why each one matters, and shows you exactly how to check it before you trust the output.
The seven logistic regression assumptions at a glance
Logistic regression is a generalized linear model, so it inherits a linear backbone but applies it to the log-odds rather than the raw outcome. That single design choice reshapes the whole assumption list. Here are all seven, what each means, and how to verify it:
| Assumption | What it means | How to check |
|---|---|---|
| Binary / categorical outcome | The dependent variable is a class label, not a continuous number | Confirm the target has two (or a fixed set of) categories |
| Independence of observations | Each row is a separate, unrelated case | Check study design; watch for repeated measures or clustering |
| Linearity of the logit | Each continuous predictor is linear in the log-odds, not the probability | Box-Tidwell test; plot logit vs predictor; add a $x\ln x$ term |
| Little multicollinearity | Predictors are not strongly correlated with each other | Variance inflation factor (VIF); correlation matrix |
| No extreme outliers | No single point dominates the fit | Standardized residuals, Cook’s distance, leverage |
| Adequate sample size | Enough events per predictor to estimate stable coefficients | Rule of thumb: about 10 events per predictor |
| No perfect separation | No predictor splits the classes perfectly | Watch for huge coefficients and standard errors that explode |
You can experiment with all of these in practice using our logistic regression calculator. Now let us walk through each assumption in turn.
1. A binary or properly categorical outcome
The first of the logistic regression assumptions is the most basic: the response variable must be categorical. Standard (binary) logistic regression expects exactly two outcomes — coded $0$ and $1$, such as pass/fail, churn/retain, or disease/no disease. If you have more than two unordered categories you need multinomial logistic regression; for ordered categories you need ordinal logistic regression. Feeding a continuous target into a logistic model is a category error — that is a job for linear regression.
How to check: simply inspect your target column. For the classic two-class case, see our guide to binary logistic regression.
2. Independence of observations
Each observation must be independent of the others. The likelihood that the model maximizes assumes that knowing the outcome of one row tells you nothing extra about another. This breaks when you have repeated measurements on the same subject, time-series data, or observations clustered within groups (students within schools, patients within hospitals).
How to check: there is no single statistic — this is mostly a question of study design. If the same unit appears multiple times, or rows are grouped, reach for a mixed-effects (multilevel) or generalized estimating equations model instead. Violating independence makes standard errors too small, so significance gets overstated.
3. Linearity of the logit (not the probability)
This is the assumption beginners most often misread. Logistic regression does not assume a straight-line relationship between predictors and the probability — that relationship is the curved sigmoid. What it assumes is that each continuous predictor is linear in the log-odds:
$$\ln\!\left(\frac{p}{1-p}\right) = \beta_0 + \beta_1 x_1 + \beta_2 x_2 + \dots + \beta_k x_k$$So the linearity assumption lives on the logit scale. If the true relationship between, say, age and the log-odds bends, the model will be misspecified even though the probability curve still looks like a reasonable S.
4. Little or no multicollinearity
Like any regression with multiple inputs, logistic regression assumes the predictors are not highly correlated with one another. When two predictors carry nearly the same information, the model cannot tell their effects apart: coefficients become unstable, their signs can flip, and standard errors balloon.
How to check: compute the variance inflation factor (VIF) for each predictor. A common guideline is that $\text{VIF} > 5$ (or more conservatively $> 10$) signals a multicollinearity problem worth addressing — by dropping a redundant predictor, combining correlated ones, or using regularization. A correlation matrix of the predictors is a quick first screen.
5. No extreme outliers or influential points
A handful of unusual rows can drag the fitted coefficients far from where the bulk of the data would put them. Logistic regression assumes no single observation exerts outsized influence. Outliers in predictor space (high leverage) are especially dangerous when combined with a surprising outcome.
How to check: examine standardized (Pearson or deviance) residuals, leverage values, and Cook’s distance. Points with large residuals or high Cook’s distance deserve a look — verify they are not data-entry errors, and consider how sensitive your conclusions are with and without them.
6. A large enough sample size
Maximum likelihood estimation, which fits logistic regression, only behaves well with enough data — and what matters is not the total row count but the number of cases in the rarer class. The widely cited rule of thumb is about 10 events per predictor (the EPV, or events-per-variable, guideline).
7. No perfect separation
Perfect (or quasi-complete) separation happens when some predictor, or combination of predictors, splits the two classes cleanly — every case above a cutoff is a 1 and every case below it is a 0. When this occurs, the maximum-likelihood estimate of the corresponding coefficient does not exist: the optimizer keeps pushing it toward $\pm\infty$ to make the predicted probabilities approach 0 and 1 exactly.
How to check: watch for the tell-tale symptoms — coefficients and standard errors that blow up to implausibly large values, or software warnings about non-convergence or “fitted probabilities numerically 0 or 1.” Remedies include removing or combining the offending predictor, gathering more data, or using penalized (Firth) logistic regression, which keeps estimates finite.
A short worked sense-check
Imagine predicting loan default from three predictors — income, credit score, and debt-to-income ratio — using $400$ applicants of whom $40$ defaulted. Run the checklist: the outcome is binary (default yes/no) ✓; applicants are independent ✓; you have $40$ events for $3$ predictors, comfortably above the $30$ that the 10-per-predictor rule asks ✓. Then you compute VIF and find income and debt-to-income are heavily correlated (VIF $> 8$) — a multicollinearity flag, so you combine them or drop one. Finally a Box-Tidwell test flags credit score as non-linear in the logit, so you add a spline. In a few minutes you have checked every assumption that matters and fixed the two that failed — that is the entire workflow.
🤖 ML context
Checking the logistic regression assumptions is part of responsible supervised learning. Logistic regression is also the simplest neural network — a single sigmoid neuron — so the same intuitions about linearity and separation carry into deep learning. Build hands-on intuition with the logistic regression calculator, then compare with multivariate logistic regression.
Frequently asked questions
What are the main assumptions of logistic regression?
Does logistic regression assume linearity?
Does logistic regression assume normality of residuals?
How do I check the linearity of the logit assumption?
What is perfect separation in logistic regression?
Key takeaways
The logistic regression assumptions are fewer and friendlier than those of linear regression: a categorical outcome, independent rows, linearity of the logit, low multicollinearity, no dominant outliers, enough events per predictor, and no perfect separation. Crucially, normality and homoscedasticity are not required. Check each with the right tool — VIF, Box-Tidwell, Cook’s distance, an events-per-variable count — and your model’s probabilities and p-values will hold up. Continue with the logistic regression calculator, the binary logistic regression guide, or the formal reference on Wikipedia.