How Many Outcome Events Does a Prognostic Study Need?

Is a Large Sample Enough if Very Few Participants Experience the Outcome?

You have assembled a cohort of 2,000 participants for a prognostic study. The dataset looks substantial, and you have several characteristics that might be associated with a future outcome. However, only 40 participants experience that outcome during follow-up.

Can you confidently estimate prognostic associations or develop a multivariable prediction model? Would recruiting more participants solve the problem? And is the familiar recommendation of ten outcome events per variable still a reliable rule?

These questions matter because prognostic information depends not only on how many people enter a study, but also on how much information the observed outcomes provide. A large cohort with few events may support some research objectives while being inadequate for others.

There Is No Universal Minimum Number of Outcome Events

How Outcome Events Determine the Information Available for Prognostic Analysis

What Counts as an Outcome Event?

An outcome event is an occurrence of the endpoint specified in the prognostic research question.

For example, if researchers investigate mortality within one year after diagnosis, deaths occurring within that horizon are the relevant events. If the outcome is first hospital readmission, the event is the first qualifying readmission. In educational research, an event might be withdrawal from an academic program.

The definition must be established before calculating event requirements. A participant experiencing several hospitalizations does not automatically contribute several independent events to a first-hospitalization analysis.

Likewise, events occurring after the specified prediction horizon do not count as events within that horizon.

The outcome definition, starting point, and follow-up period therefore determine what event information is available.

Why Is the Number of Events Often More Important Than the Total Sample?

For binary and time-to-event outcomes, the number and distribution of observed events strongly influence how precisely certain associations and model parameters can be estimated.

Consider two hypothetical studies containing 1,000 participants each. In one, 300 experience the outcome; in the other, only 20 do.

The second study has the same number of participants but considerably less event information for estimating associations involving the outcome.

Its regression coefficients may be imprecise or unstable, particularly when several predictors are included.

However, events alone do not determine information. The number of non-events, predictor distributions, censoring, correlation among predictors, and outcome variability also matter.

For continuous outcomes, the relevant sample size considerations differ because the analysis does not rely on a count of discrete outcome events.

Do All Prognostic Studies Need the Same Number of Events?

No. The requirements differ substantially according to the research objective.

Applying the same numerical event threshold to all four objectives overlooks their different statistical requirements.

What Is the Events-per-Variable Rule?

Events per variable, commonly abbreviated EPV, is a traditional measure comparing the number of outcome events with the number of parameters or variables included in a regression model.

The frequently cited rule of ten events per variable emerged from simulation research examining instability in particular logistic regression settings.

It subsequently became a convenient heuristic in observational and prognostic research.

However, later methodological work demonstrated that the required information depends on considerably more than a single event-to-variable ratio.

Some models may perform adequately with fewer than ten events per parameter under favorable conditions, while others may remain unstable with considerably more.

The historical rule is therefore not a general sample size calculation.

Why Can Ten Events Per Variable Be Misleading?

The rule ignores important features of the study, including the expected strength of predictor-outcome relationships, predictor distributions, correlation among predictors, the number of candidate parameters, and the intended modeling strategy.

For example, a model containing highly correlated predictors may be difficult to estimate even when the nominal EPV appears acceptable.

A rare binary predictor may provide little information if very few events occur among participants possessing that characteristic.

Conversely, a relatively simple model with strong, prespecified predictors may require less information than a highly flexible model involving nonlinear terms and interactions.

Researchers should not interpret EPV as a direct measure of model validity or generalizability.

Should Events Be Counted Per Variable or Per Parameter?

When EPV is used descriptively, the number of predictor parameters is generally more informative than the number of named variables.

A continuous predictor represented by one linear term may contribute one regression coefficient. Modeling it with several spline terms requires additional parameters.

A categorical predictor with four categories ordinarily requires three indicator parameters under conventional reference-category coding.

Interactions introduce additional parameters, while flexible transformations can increase complexity further.

Consequently, a model containing five named predictors may require considerably more than five predictor coefficients.

Counting only variable names can therefore understate model complexity.

Why Must Candidate Predictors Be Considered?

Researchers sometimes calculate EPV using only the predictors retained in the final model.

This can be misleading when many candidate predictors were examined and a data-driven selection procedure was used to choose the final subset.

The selection process itself uses outcome information and may introduce optimism or unstable coefficient estimates.

For prediction-model development, sample size planning should therefore consider the full candidate predictor parameter set and the modeling procedure, not merely the number of coefficients ultimately reported.

The question of how many candidate predictors the available data can support requires explicit attention to model complexity.

How Is Sample Size Determined When Estimating Overall Prognosis?

When the objective is to estimate an outcome proportion at a fixed horizon with complete ascertainment, sample size can be planned around the desired confidence interval precision.

For an illustrative simple random sample and a binary outcome, an approximate planning formula is:

This illustration concerns precision of a single outcome proportion. It does not establish that 246 participants would be sufficient for a multivariable prognostic model.

For survival probabilities, researchers should use planning methods that account for follow-up duration, censoring, and the intended estimator.

How Should Prognostic Factor Studies Be Planned?

Prognostic factor studies often focus on estimating the association between a particular characteristic and a future outcome.

Sample size planning should consider the anticipated association, the distribution of the factor, the number and relationships of adjustment variables, outcome frequency, and desired precision or statistical power.

For example, examining a rare exposure may require a larger sample than examining a characteristic that varies widely across the population.

Researchers should specify the association measure and clinically meaningful effect before conducting a power or precision calculation.

Exploratory analyses involving many candidate factors require additional caution because unstable estimates and selective reporting may become more likely.

The aim should be to estimate the association with sufficient precision, not merely to obtain a statistically significant p-value.

Why Does Prediction-Model Development Need a Different Calculation?

Prediction-model development aims to estimate future outcomes accurately using several predictors.

Sample size requirements therefore concern the stability of the complete prediction function, not simply the precision of one regression coefficient.

For binary and time-to-event prediction models, modern sample size approaches consider factors such as anticipated model performance, the number of candidate predictor parameters, overall outcome risk, and the desired degree of overfitting control.

Riley and colleagues proposed criteria that include limiting overfitting, controlling optimism in apparent model fit, and estimating the overall outcome risk with sufficient precision.

The largest sample size required across the relevant criteria should guide planning.

Such methods are preferable to treating ten events per parameter as an automatic design target.

What Is Overfitting, and Why Do Few Events Increase the Risk?

Overfitting occurs when a model captures sample-specific variation that does not generalize adequately to new participants.

When relatively few events are available compared with model complexity, regression coefficients may become unstable and predicted risks may be overly extreme.

The model can appear to discriminate well in its development data while performing less reliably in independent data.

Penalization, shrinkage, and appropriate internal validation can reduce or assess some aspects of overfitting.

However, these methods do not make sample size irrelevant.

A model developed with inadequate information may still produce uncertain predictions even when sophisticated regularization methods are used.

Does External Validation Require Its Own Sample Size Calculation?

Yes. External validation has a different objective from model development.

Its purpose is to estimate how accurately a fixed model performs in new data, particularly in terms of calibration and discrimination.

The required sample size depends on the precision sought for these performance measures, anticipated outcome frequency, and distribution of predicted risks.

Historical recommendations based on fixed numbers of events and non-events may be useful as rough reference points, but they should not replace a calculation tailored to the validation objective.

Researchers should distinguish external validation requirements from the sample size needed to develop a new model.

How Do Censoring and Follow-Up Affect Event Requirements?

In time-to-event studies, the amount of information depends on when events occur and how long participants remain observable.

Administrative censoring and loss to follow-up can reduce the number of events observed within the intended horizon.

For example, a registry containing 5,000 participants may provide limited information about five-year outcomes if most participants have only one year of observation.

Researchers should therefore estimate expected event accrual using realistic recruitment and follow-up assumptions.

Extending follow-up may increase event numbers, but it does not necessarily improve precision at an earlier, fixed prediction horizon.

The decision about how long participants should be followed must be coordinated with sample size planning.

What Happens When Competing Events Are Common?

Competing events can prevent participants from experiencing the outcome of interest.

For example, death before a nonfatal complication reduces the number of individuals who can subsequently experience that complication.

Sample size planning based only on the target event rate may therefore be misleading when competing events are frequent.

Researchers should consider the cumulative incidence of the target outcome, competing-event probabilities, and the intended analysis.

Cause-specific hazard models, cumulative incidence models, and prediction models may have different information requirements.

The influence of competing events on prognostic design should be incorporated into planning rather than addressed only after data collection.

Can More Non-Events Compensate for Very Few Events?

Not indefinitely. Both event and non-event information matter for binary outcome modeling, but adding large numbers of non-events does not automatically resolve instability caused by very few events.

For example, a dataset with 20 events and 20,000 non-events may still provide limited information for estimating several associations among participants who experienced the outcome.

Rare-event methods and penalized regression may be appropriate in particular circumstances, but they require careful evaluation.

Researchers should consider the balance of event and non-event information, predictor distributions, and intended model complexity.

What If the Available Dataset Is Too Small?

The appropriate response depends on the research objective and the degree of information deficiency.

Researchers may consider recruiting additional participants, extending appropriate follow-up, reducing unnecessary model complexity, or focusing on a narrower prespecified question.

Combining compatible datasets may also increase information, provided measurement and population differences are addressed.

For prediction research, validating an existing model may be more defensible than developing a new model from inadequate data.

Researchers should avoid selecting predictors solely to achieve a favorable EPV after inspecting the outcome data.

If the available information remains inadequate, the study should report exploratory estimates with appropriate uncertainty rather than making unsupported confirmatory claims.

How Should Sample Size Justification Be Reported?

The protocol should identify the primary objective, outcome definition, prediction horizon, anticipated event frequency, and relevant statistical assumptions.

For regression analyses, researchers should report the number of candidate predictor parameters and the rationale for the chosen modeling approach.

For prediction-model development, the sample size method and anticipated predictive performance assumptions should be documented.

For external validation, the intended precision of performance estimates should be specified.

Reports should distinguish the sample size planned before data collection from the actual number of participants and observed events.

Transparent justification is more informative than simply stating that the study satisfied a conventional events-per-variable rule.

Why 2,000 Participants May Still Be Insufficient for a Complex Prediction Model

The example demonstrates why a large participant count does not necessarily provide adequate information for complex prognostic modeling.

It also shows why researchers should not declare a model acceptable or unacceptable solely from its EPV. A defensible conclusion requires a calculation aligned with the actual modeling objective.

Common Mistakes When Planning Outcome-Event Requirements

How to Determine Whether Your Prognostic Study Has Enough Events

Begin by identifying the primary research objective and the statistical quantity you need to estimate.

Then determine the anticipated outcome frequency at the relevant horizon and the amount of information required for an acceptably precise or stable result.

Document the assumptions underlying the calculation, including event frequency, follow-up, missingness, and model complexity.

For prediction-model studies, use established sample size methods or appropriate software implementing those methods rather than relying exclusively on a simple EPV ratio.

Ultimately, the relevant question is not how many events a conventional rule recommends, but how much outcome information is necessary to support the specific inference the study intends to make.

Before Finalizing Prognostic Study Sample Size

Frequently Asked Questions About Outcome Events and Sample Size

Enough Events Means Enough Information for the Intended Inference

How to Cite This Guide

This guide is intended to be read, shared, and used in research, teaching, and academic work. If you draw on its ideas, explanations, or other content, please acknowledge the source by citing the guide. Doing so gives appropriate credit and helps your readers locate the original resource.