Uniform shrinkage
Uniform shrinkage multiplies fitted slopes by one factor, then updates the intercept. It can reduce overfitting but cannot rescue a model built from too little data.
Fit
When to use it, and when not to
Use it when
- You are developing a prediction model and want one shrinkage factor for all fitted slopes.
- You can estimate shrinkage and check the full modeling procedure with resampling.
Not when
- You need a penalty estimated during fitting. Compare ridge, lasso and elastic net. Read: Choosing a penalty
- A logistic model has separation. Consider Firth logistic regression. Read: Firth logistic regression
- You want a causal effect. Choose adjustment from causal knowledge, not by shrinkage or selection. Firth's correction still suits separation there. Read: Directed acyclic graphs
Before you start
What you need
- That the model is for predictiontargetPrediction or prognosis for new patients, judged by calibration and discrimination.
- The outcome, and when it is measuredoutcomeTypeSuch as death within 30 days. The model and its sample size depend on whether it is yes or no, a time or a measurement.
- The candidate predictors, counted as parametersA spline, or a category with several levels, counts as several. Choose them from knowledge and from what will be known at prediction time.
- The sample size, from Riley's criteriaThey use the outcome rate, the expected model performance and the number of parameters. Mallard's sizing catalog has no form for this.
- How performance will be checkedResampling must repeat all model-building steps and estimate the shrinkage factor without using future validation outcomes.
- The development data resemble the setting where predictions will be used.How to check: Validate predictions in new patients. Shrinkage does not fix a change in setting.
- The data can estimate a useful shrinkage factor.How to check: Use outcome-specific sample-size criteria. The factor can be very uncertain in small samples.
- One factor is suitable for the fitted slopes.How to check: Check calibration after shrinkage, and update the intercept to match the average outcome.
Plain language
How it works
Overfitting makes predictions too extreme in new patients. Uniform shrinkage applies one estimated factor to all fitted slopes after the initial fit. For logistic prediction, update the intercept with the shrunken predictor fixed as an offset. This restores average calibration in the development sample, not necessarily elsewhere.
Estimate the factor with a stated method, such as bootstrap validation that repeats the modeling steps. The factor itself is uncertain. Simulation studies show improved calibration on average but sometimes poor calibration in individual datasets. Shrinkage does not replace adequate sample size or validation.
Worked example
A worked example
- Setting
- A logistic prediction model was developed for hospital readmission.
- Question
- What does a common shrinkage factor do to a fitted slope?
- Approach
- The fitted slope is 0.5 and the estimated shrinkage factor is 0.8. Multiply each slope by that factor, then update the intercept.
- Result
- The illustrated slope falls from 0.5 to 0.4, a change of -0.1.
- What it means
- The factor changes every slope by the same proportion. Its uncertainty and performance in new patients still need assessment.
In practice
Recent studies PubMed search · not vetted
Watch out
Common mistakes
The one to remember
A common shrinkage factor is estimated with uncertainty. It cannot guarantee calibration in new patients or compensate for too few outcome events.
- Multiplying the intercept by the same factor as the slopes.Instead: Shrink the slopes, then estimate the intercept again so average predictions match the outcome frequency.
- Treating the estimated factor as certain.Instead: Repeat the full procedure in resampling and assess calibration in new patients.
- Sizing a prediction model by events per variable alone.Instead: Events per variable is only a floor. Use outcome-specific Riley criteria for the planned model.
Inside Mallard
How Mallard uses this
For binary prediction, Mallard may offer penalized model development with bootstrap optimism correction. Uniform shrinkage is a related method, not a separately selectable option.
analysis.penalized_regression, the id a plan cue records.Related Learn pages
Related designs and methods
Evidence
Sources & further reading
- Penalization and shrinkage methods produced unreliable clinical prediction models especially when sample size was small (opens in a new tab) · J Clin Epidemiol · doi 10.1016/j.jclinepi.2020.12.005
- Regression shrinkage methods for clinical prediction models do not guarantee improved performance: Simulation study (opens in a new tab) · Stat Methods Med Res · doi 10.1177/0962280220921415
- Minimum sample size for developing a multivariable prediction model: PART II - binary and time-to-event outcomes (opens in a new tab) · Stat Med · doi 10.1002/sim.7992