03 · What You Need to Know
Understanding Outcome Events and Logistic Regression Sample-Size Requirements
What Counts as an Outcome Event?
Binary logistic regression models the probability of an outcome with two categories, commonly coded zero and one.
An event is conventionally the outcome category coded one. For example, in a student-retention study, dropout might be coded one and continued enrollment zero.
The event count is the number of observations with the event outcome. The remaining observations are nonevents.
However, the coding choice is not inherently a statement about which category provides all the statistical information. Both categories are relevant to estimating the relationship between predictors and outcome probabilities.
When one category is uncommon, its scarcity can become an important limitation.
Why Does Logistic Regression Need Information From Both Outcome Categories?
Logistic regression estimates coefficients by relating predictor values to the probability of an event.
Under conventional maximum-likelihood estimation, information about regression coefficients depends partly on the fitted event probabilities and the distribution of predictors.
For a binary observation with fitted probability π, the familiar Bernoulli variance term is π(1 − π). This quantity is relatively small when the predicted probability approaches zero or one.
In matrix notation, the information matrix for a conventional logistic regression model has the form:
The distribution and correlation of predictors also influence the information matrix. Two datasets with the same numbers of events and nonevents may therefore provide substantially different information for a particular coefficient.
What Is the Events-per-Variable Ratio?
Events per variable, or EPV, describes the number of outcome events relative to the number of predictor variables or parameters.
For a more meaningful assessment of model complexity, the denominator should reflect the candidate predictor parameters rather than merely the number of named variables.
Is Ten Events per Variable Still the Recommended Minimum?
The ten-EPV recommendation became influential following simulation research by Peduzzi and colleagues (1996), who investigated problems associated with limited events in logistic regression.
Subsequent investigations found that model performance depends on more than this ratio.
Vittinghoff and McCulloch (2007) showed that fewer than ten events per variable could sometimes produce acceptable results under particular conditions. Van Smeden and colleagues (2019) demonstrated important limitations of EPV-based criteria for binary prediction models.
These findings do not establish a universally safe alternative of five, fifteen, or twenty EPV.
The limitations of the ten events per variable rule arise because the ratio omits important information about predictor distributions, model complexity, and intended performance.
Why Can a Large Dataset Still Have Too Few Events?
When the outcome is rare, total sample size can substantially overstate the information available for binary-outcome modeling.
Consider a dataset of 2,000 students in which only 1% experience dropout. The expected number of dropout events is approximately 20.
That may be insufficient for a complex model involving numerous predictors, categorical levels, and interactions.
By contrast, a dataset of 400 students with 30% dropout prevalence would contain approximately 120 events on average.
Neither example establishes adequacy by itself, but it demonstrates why total rows should not be the sole planning measure.
The same issue is important when evaluating whether thousands of observations are sufficient for machine-learning prediction.
How Do You Convert an Event Requirement Into a Recruitment Target?
Once an event target has been justified using an appropriate planning method, researchers can translate it into an expected total sample using the anticipated event proportion.
When a minimum realized event count is essential, a probability-based recruitment calculation may be more appropriate than dividing the event target by expected prevalence.
Researchers should also account for missing outcomes and anticipated unusable records.
Why Might the Smaller Outcome Category Be More Relevant?
Suppose an outcome occurs in 95% of participants. Counting only the 95% event category may make the dataset appear information-rich, even though very few nonevents are available.
Examining the smaller category is a useful diagnostic because sparse outcome categories can create instability regardless of which category is coded as the event.
However, a minimum-category-count ratio remains a heuristic. It does not account for how outcome categories are distributed across predictor values.
Modern planning should consider the anticipated outcome prevalence together with the full model specification.
What Happens When Events Are Sparse Across Predictor Categories?
Even a seemingly adequate overall event count may conceal sparse combinations of predictor values and outcomes.
Suppose a university dropout model includes a five-category academic program variable. One program may contain very few students, with only one observed dropout.
The coefficient for that category may be estimated imprecisely despite the model having many events overall.
Similarly, interactions between predictors can create small subgroups with limited outcome information.
Planning should therefore consider the distribution of candidate predictors, especially rare categorical levels, nonlinear terms, and interactions.
What Is Complete Separation?
Complete separation occurs when a linear combination of predictors perfectly separates the two outcome categories.
Under ordinary maximum-likelihood logistic regression, complete separation prevents finite maximum-likelihood estimates for certain coefficients. Quasi-complete separation creates related difficulties.
For example, if every observed student with a particular rare characteristic drops out, the associated coefficient may become extremely large or fail to converge to a finite maximum-likelihood estimate.
Separation is especially concerning with sparse outcome-predictor combinations, although it is not determined solely by the total event count.
Heinze and Schemper (2002) describe a bias-reduction approach that addresses separation through penalized likelihood.
Watch Out
Obtaining finite coefficients through penalized regression does not establish that the available sample is sufficiently informative. Penalization may improve estimation behavior, but confidence intervals and predictions can remain uncertain.
How Do Requirements Differ Between Explanatory and Predictive Logistic Regression?
Explanatory logistic regression may focus on a particular adjusted association, such as the relationship between prior academic failure and subsequent dropout after controlling for prespecified confounders.
In this setting, sample-size planning can focus on the precision or power of the primary regression coefficient. Relevant assumptions include outcome prevalence, predictor prevalence, anticipated association, and correlation with adjustment variables.
Prediction model development has a different objective: producing reliable predictions for new observations.
Prediction planning must consider overfitting, calibration, anticipated predictive performance, candidate parameter complexity, and internal validation.
Riley and colleagues (2019) propose criteria for minimum sample-size planning in binary prediction model development that go beyond a fixed event-to-parameter ratio.
The sample size needed for prediction model development should therefore not be inferred solely from requirements for testing one explanatory coefficient.
How Many Events Are Needed for a Reliable Odds Ratio?
There is no universal event count for estimating an odds ratio with useful precision.
Precision depends on the number of observations across combinations of predictor and outcome categories, the magnitude of the association, and the adjustment structure.
A rare exposure can produce a wide confidence interval even when the total event count is relatively large.
For a prespecified adjusted odds ratio, a coefficient-specific precision calculation or simulation may be more appropriate than an EPV threshold.
The researcher should specify the smallest or largest odds ratio that matters for the research question and the confidence-interval width or power required to investigate it.
How Does the Number of Candidate Parameters Affect the Event Requirement?
Additional predictor parameters generally increase model complexity and can raise the information requirements for stable estimation or prediction.
However, the relationship is not a universal linear multiplier.
A categorical predictor with five levels may require four coefficients. Nonlinear transformations and interactions can introduce further parameters.
Researchers should count all candidate parameters that the modeling process is allowed to consider, not merely the predictors retained in the final fitted model.
Otherwise, data-driven selection can make the apparent model complexity misleadingly small.
Can Penalization Reduce the Number of Events Needed?
Penalized logistic regression, including ridge, lasso, and bias-reduced estimation methods, can improve some aspects of estimation in limited samples.
However, the effect depends on the penalty, tuning procedure, underlying model, and performance criterion.
Penalization cannot recover information that the data never contained. A model with extremely few events may still produce uncertain predictions or associations.
Researchers should evaluate penalized procedures using appropriate validation and, where possible, simulation-based planning rather than assuming that regularization eliminates sample-size requirements.
Can Simulation Help Determine an Appropriate Event Count?
Yes. Simulation is useful when the model includes sparse categories, nonlinear terms, interactions, or complex predictor distributions.
Researchers can generate hypothetical datasets under plausible event rates and predictor relationships, fit the intended model, and evaluate repeated-sample performance.
For explanatory models, criteria may include coefficient bias, standard error, confidence-interval coverage, and power. For prediction models, calibration, optimism, and prediction error may be more relevant.
The resulting event requirement depends on the simulated conditions. A sample-size sensitivity analysis should examine uncertainty in important planning parameters.
What If the Available Dataset Has Too Few Events?
First, reconsider whether the proposed model complexity is justified by the available information.
Possible responses include restricting the analysis to prespecified essential parameters, improving measurement, combining additional comparable data, or using suitable penalized estimation.
Removing important confounders merely to satisfy an EPV threshold may introduce bias and should not be treated as a routine solution.
If adequate estimation remains infeasible, the study may need to adopt a narrower objective and explicitly report the limitations of the available data.
Researchers should avoid presenting exploratory findings from a sparse model as definitive evidence of stable associations or reliable predictions.
