A chart review, a registry pull, or a database extract does not let you choose n — it is whatever the data already gave you. The planning question that matters is different from a prospective trial’s: given the events you have, how many covariates can a logistic or Cox model support before it overfits? This tool applies the events-per-variable (EPV) heuristic to answer it.
The chart pull, registry cohort, or database extract you already have.
The rarer of the two outcome levels — EPV is built on events, not patients.
Everything you are considering putting in the model — not just the ones you expect to matter.
30 events in the rarer outcome category — your covariate count exceeds the budget.
10 adjusters exceed the EPV budget of 3 (≈30 events / 10 per variable) — overfitting risk: reduce to ≤3, use a penalized model (LASSO/ridge), or increase the number of events.
Events-per-variable heuristic (Peduzzi et al., 1996): max covariates = floor(events ÷ EPV), EPV ≥ 10 — counted from events in the rarer outcome category, not total n.
The calculator above is deterministic and free, and it only checks the arithmetic. If you want a statistician to look at the actual model — which covariates belong, whether a penalized approach fits your data better than trimming the list — tell us, and we will let you know if we build it.
Total sample size is the wrong denominator for a covariate budget. A cohort of 2,000 patients with a 1.5% outcome rate has only about 30 events to spend — the same budget as a cohort of 300 patients with a 10% outcome rate. EPV counts the rarer outcome category because that is the quantity that actually constrains how many parameters a logistic or Cox model can estimate reliably; total n can be large while the model is still starved of events.
The heuristic itself is simple — max covariates = floor(events ÷ 10) — but it only answers one part of the modeling question. Before the covariate count is even the right thing to worry about, you need the right primary analysis for your comparison structure. If you have not settled that yet, pick the statistical test that fits your design first.
EPV 10 comes from one influential logistic-regression simulation study (Peduzzi et al., 1996), not from a theorem — it is a planning floor, and reasonable people disagree on how strict it should be. It also does not by itself account for penalized regression (LASSO, ridge, Firth), which relaxes the constraint by shrinking coefficients rather than estimating them freely; a model built that way can carry more predictors than a naive EPV count suggests. And EPV budgets the number of parameters, not the number of variables — a categorical variable with four levels spends three degrees of freedom, and a spline or interaction term spends more than one.
What is events per variable (EPV)?
The ratio of the number of events in the rarer outcome category to the number of candidate predictor variables in a logistic or Cox regression model. It is a planning heuristic for how many covariates a model can support before its coefficient estimates become unstable — built on events, not on total sample size.
Why 10 events per variable?
Peduzzi and colleagues' 1996 simulation study on logistic regression found that below roughly 10 events per variable, coefficient estimates grew biased and unstable. EPV 10 became the conventional planning floor, but it is a rule of thumb from one set of simulations, not a law — some methodologists argue for a stricter 15–20, and the right number depends on the model and the effect sizes involved.
What if I exceed my covariate budget?
Treat it as a flag, not a stop sign: reduce the candidate list to variables with the strongest prior clinical justification (not the smallest p-value), consider a penalized model such as LASSO, ridge, or Firth logistic regression, which relax the EPV constraint, or gather more events before finalizing the model. Which option fits depends on your data and question — this calculator surfaces the tradeoff, it does not resolve it for you.