Logistic regression with rare events:

problems and solutions

Georg Heinze
Medical University of Vienna

Supported by the Austrian Science Fund FWF (I2276-N33)

Barcelona, 27 September 2017

[email protected] @Georg__Heinze http://prema.mf.uni-lj.si http://cemsiis.meduniwien.ac.at/en/kb

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 1
Rare events: examples

Medicine:
• Side effects of treatment 1/1000s to fairly common
• Hospital-acquired infections 9.8/1000 pd
• Epidemiologic studies of rare diseases 1/1000 to 1/200,000
Engineering:
• Rare failures of systems 0.1-1/year
Economy:
• E-commerce click rates 1-2/1000 impressions
Political science:
• Wars, election surprises, vetos 1/dozens to 1/1000s
…
Georg Heinze
27.09.2017 2
Problems with rare events

• ‚Big‘ studies needed to observe enough events

• Difficult to attribute events to risk factors

• Low absolute number of events

• Low event rate

Georg Heinze
27.09.2017 3
Our interest

• Models
• for prediction of binary outcomes
• should be interpretable,
i.e., betas should have a meaning
 explanatory models

Georg Heinze
27.09.2017 4
Logistic regression

Pr 1 1 exp

• Leads to odds ratio interpretation of exp :

Pr 1 1 /Pr 0| 1
exp
Pr 1 /Pr 0|

• Likelihood: L | ∏ 1
• Its nth root: Probability of correct prediction

• How well can we estimate if events ( 1) are rare?

Georg Heinze
27.09.2017 5
 Rare event problems…

Not much gain!

Logistic regression with 5 variables:

• estimates are unstable (large MSE) because of few events
• removing some ‚non-events‘ does not affect precision
Georg Heinze
27.09.2017 6
Penalized likelihood regression

∗
log log

Imposes priors on model coefficients, e.g.

• ∑ (ridge: normal prior)
• ∑| | (LASSO: double exponential)
• log det (Firth-type: Jeffreys prior)
in order to
• avoid extreme estimates and stabilize variance (ridge)
• perform variable selection (LASSO)
• correct small-sample bias in (Firth-type)

Georg Heinze
27.09.2017 7
Firth‘s penalization for logistic regression

In exponential family models with canonical parametrization the Firth-type

penalized likelihood is given by
∗ /
det ,

where is the Fisher information matrix and is the likelihood.

Firth-type penalization

• removes the first-order bias of the ML-estimates of ,

• is bias-preventive rather than corrective,

• is available in Software packages such as SAS, R, Stata…

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 8
Firth‘s penalization for logistic regression

In exponential family models with canonical parametrization the Firth-type

penalized likelihood is given by
Jeffreys
∗ /
det , invariant prior
where is the Fisher information matrix and is the likelihood.

Firth-type penalization

• removes the first-order bias of the ML-estimates of ,

• is bias-preventive rather than corrective,

• is available in Software packages such as SAS, R, Stata…

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 9
Firth‘s penalization for logistic regression

In logistic regression, the penalized likelihood is given by

∗
det / , with

diag expit X 1 expit X

diag 1 .

• Firth-type estimates always exist.

is maximised at , i.e. at 0, thus

• predictions are usually pulled towards ,

• coefficients towards zero.
Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 10
Firth‘s penalization for logistic regression

• Separation of outcome classes by covariate values (Figs. from Mansournia et al 2017)

• Firth‘s bias reduction method was proposed as solution to the problem of separation
in logistic regression (Heinze and Schemper, 2002)

• Penalized likelihood has a unique mode

• It prevents infinite coefficients to occur

Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 11
Firth‘s penalization for logistic regression

Bias reduction also leads to reduction in MSE:

• Rainey, 2017: Simulation study of LogReg for political science

‚Firth‘s methods dominates ML in bias and MSE‘

However, the predictions get biased…

• Elgmati et al, 2015

… and anti-shrinkage could occasionally arise:

• Greenland and Mansournia, 2015

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 12
Firth‘s Logistic regression

For logistic regression with one binary regressor*,

Firth’s bias correction amounts to adding 1/2 to each cell:

original augmented
A B A B
Firth-type
Y=0 44 4 0 44.5 4.5
penalization
Y=1 1 1 1 1.5 1.5

event rate 0.04 event rate ~0.058

ORBvsA 11 ORBvsA 9.89
av. pred. prob. 0.054

* Generally: for saturated models

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 13
Example of Greenland 2010
original augmented

A B A B
Y=0 315 5 320 Y=0 315.5 5.5 321
Y=1 31 1 32 Y=1 31.5 1.5 33
346 6 352 346.5 6.5 354

event rate 0.091 event rate 0.093

ORBvsA 2.03 ORBvsA 2.73

Greenland, AmStat 2010

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 14
Greenland example: likelihood, prior, posterior

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 15
Bayesian non-collapsibility:
anti-shrinkage from penalization

• Prior and likelihood modes do not ‚collapse‘:

posterior mode exceeds both
• The ‚shrunken‘ estimate
is larger than ML estimate

• How can that happen???

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 16
An even more extreme example
from Greenland 2010

• 2x2 table X=0 X=1

Y=0 25 5 30
Y=1 5 1 6
30 6 36

• Here we immediately see that the odds ratio = 1 ( 0

• But the estimate from augmented data: odds ratio = 1.26

(try it out!) Greenland, AmStat 2010

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics
17
Simulating the example of Greenland

• We should distinguish BNC in a single data set from a systematic increase

in bias of a method (in simulations)
X=0 X=1
Y=0 315 5 320
Y=1 31 1 32
346 6 352

• Simulation of the example:

• Fixed groups x=0 and x=1, P(Y=1|X) as observed in example

• True log OR=0.709

Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 18
Simulating the example of Greenland

• True value: log OR = 0.709

Parameter ML Jeffreys-
Firth
Bias * +18%
RMSE * 0.86
Bayesian non- 63.7%
collapsibility

* Separation causes to be undefined ( ∞) in 31.7% of the cases

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 19
Simulating the example of Greenland

• To overcome Bayesian non-collapsibility,

Greenland and Mansournia (2015)
proposed not to impose a prior on the intercept

• They suggest a log-F(1,1) prior for all other regression coefficients

• The method can be used with conventional frequentist software

because it uses a data-augmentation prior

Greenland and Mansournia, StatMed 2015

Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 20
logF(1,1) prior (Greenland and Mansournia, 2015)
∗
Penalizing by log-F(1,1) prior gives ⋅ ∏ .

This amounts to the following modification of the data set:

x1 x2 y x1 x2 y
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
1 ∗ ∗ ∗ 1 ∗ ∗ ∗ each assigned a weight of 1
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
1 ∗ ∗ ∗ 1 ∗ ∗ ∗
0 1 0 0
0 1 0 1
0 0 1 0 each assigned a weight of ½
0 0 1 1
• No shrinkage for the intercept, no rescaling of the variables

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 21
Simulating the example of Greenland

• Re-running the simulation with the log-F(1,1) method yields:

Parameter ML Jeffreys- logF(1,1)

Firth
Bias * +18%
RMSE * 0.86
Bayesian non- 63.7% 0%
collapsibility

* Separation causes be undefined ( ∞) in 31.7% of the cases

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 22
Simulating the example of Greenland

• Re-running the simulation with the log-F(1,1) method yields:

Parameter ML Jeffreys- logF(1,1)

Firth
Bias * +18% -52%
RMSE * 0.86 1.05
Bayesian non- 63.7% 0%
collapsibility

* Separation causes be undefined ( ∞) in 31.7% of the cases

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 23
Other, more subtle occurrences
of Bayesian non-collapsibility

• Ridge regression: normal prior around 0

• usually implies bias towards zero,
• But:
• With correlated predictors with different effect sizes,
for some predictors the bias can be away from zero

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 24
Simulation of bivariable log reg models

• , ~Bin 0.5 with correlation 0.8, 50

• 1.5, 0.1, ridge parameter optimized by cross-validation
Parameter ML Ridge (CV ) Log- Jeffreys-
F(1,1) Firth
Bias +40% (+9%*) -26% -2.5% +1.2%
RMSE 3.04 (1.02*) 1.01 0.73 0.79
Bias -451% (+16%*) +48% +77% +16%
RMSE 2.95 (0.81*) 0.73 0.68 0.76
Bayesian non- 25% 28% 23%
collapsibility
*excluding 2.7% separated samples
Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 25
Anti-shrinkage from penalization?
Bayesian non-collapsibility/anti-shrinkage

• can be avoided in univariable models,

but no general rule to avoid it in multivariable models

• Likelihood penalization can often decrease RMSE

(even with occasional anti-shrinkage)

• Likelihood penalization guaranteed shrinkage

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 26
Reason for anti-shrinkage

• We look at the association of X and Y

• We could treat the source of data as a ‚ghost factor‘ G

• G=0 for original table

• G=1 for pseudo data

• We ignore that the conditional association of X and Y is confounded by G

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 27
Example of Greenland 2010 revisited
original augmented

A B A B
Y=0 315 5 320 Y=0 315.5 5.5 321
Y=1 31 1 32 Y=1 31.5 1.5 33
346 6 352 347 7 352

To overcome both the overestimation and anti-shrinkage problems:

• We propose to adjust for the confounding by including the ‚ghost factor‘ G

in a logistic regression model

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 28
FLAC: Firth‘s Logistic regression with Added Covariate
Split the augmented data into the original and pseudo data:
original pseudo
augmented G=0 G=1 Ghost factor
A B A B A B
0 315.5 5.5 0 315 5 0 0.5 0.5
1 31.5 1.5 1 31 1 1 0.5 0.5
ORBvsA 2.03

Define Firth type Logistic regression with Additional

Covariate as an analysis including the ghost factor as
added covariate:

ORBvsA 1.84

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 29
FLAC: Firth‘s Logistic regression with Added Covariate

Beyond 2x2 tables:

Firth-type penalization can be obtained by solving modified score equations:

1
0; 0, … ,
2

/
where the ’s are the diagonal elements of the hat matrix
They are equivalent to:

1
2

1 0
2 2
Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 30
FLAC: Firth‘s Logistic regression with Added Covariate

• A closer inspection yields:

1 0
2 2

The original data

Original data, Data with reversed outcome,
weighted by /2 weighted by /2

Pseudo data

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 31
 FLAC: Firth‘s Logistic regression with Added Covariate

• A closer inspection yields:

1 0
2 2

The original data

Original data, data with reversed outcome,
weighted by /2 weighted by /2

Pseudo data

Ghost factor: G=0 G=1

(‚Added covariate‘)
Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 32
FLAC: Firth‘s Logistic regression with Added Covariate

FLAC estimates can be obtained by the following steps:

1) Define an indicator variable discriminating between original and
pseudo data.
2) Apply ML on the augmented data including the indicator.

unbiased pred. probabilities

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 33
FLIC

Firth type Logistic regression with Intercept Correction:

1. Fit a Firth logistic regression model

2. Modify the intercept in Firth-type estimates such that the average pred. prob.
becomes equal to the observed proportion of events.

unbiased pred. probabilities

effect estimates are the same as in Firth type logistic regression

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 34
Simulation study: the set-up

We investigated the performance of FLIC and FLAC,

simulating 1000 data sets for 45 scenarios with:

• 500, 1000 or 1400 observations,

• event rates of 1%, 2%, 5% or 10%

• 10 covariables (6 cat., 4 cont.),

see Binder et al., 2011

• none, moderate and strong effects

of positive and mixed signs

Main evaluation criteria:

bias and RMSE of predictions and effect estimates

Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 35
Other methods for accurate prediction

In our simulation study, we compared FLIC and FLAC to the following methods:

• weakened Firth-type penalization (Elgmati 2015),

∗
with det , 0.1, WF

• ridge regression, RR

• penalization by log-F(1,1) priors, LF

• penalization by Cauchy priors with scale parameter=2.5. CP

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 36
Cauchy priors (CP)
Cauchy priors (scale=2.5) have heavier tails than log-F(1,1)-priors:

We follow Gelman 2008:

• all variables are centered,
• binary variables are coded to have a range of 1,
• all other variables are scaled to have standard deviation 0.5,
• the intercept is penalized by Cauchy(0,10).
This is implemented in the function bayesglm in the R-package arm.

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 37
Simulation results

• Bias of : clear winner is Firth method

FLAC, logF, CP: slight bias towards 0

• RMSE of :
equal effect sizes: ridge the winner
unequal effect sizes: very good performance of FLAC and CP
closely followed by logF(1,1)

• Calibration: often FLAC the winner; considerable instability of ridge

• Bias and RMSE of : see following slides

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 38
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 39
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 40
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 41
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 42
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 43
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 44
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 45
Predictions: bias RMSE

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 46
Comparison

FLAC Bayesian methods (CP, logF)

• No tuning parameter • CP: in-built standardization,
no tuning parameter
• Transformation-invariant
• logF(m,m): choose m by ’95% prior region’ for
• Often best MSE, calibration
parameter of interest
Ridge m=1 for wide prior, m=2 less vague
• Standardization is standard • (in principle, m could be tuned as in ridge)
• Tuning parameter • logF: easily implemented
– no confidence intervals
• CP and logF are not transformation-invariant
• Not transformation-invariant

• Performance decreases
if effects are very different

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 47
Confidence intervals

It is important to note that:

• With penalized (=shrinkage) methods one cannot achieve nominal coverage over
all possible parameter values

• But one can achieve nominal coverage averaging over the implicit prior

• Prior – penalty correspondence can be a-priori established

if there is no tuning parameter

• Important to use profile penalized likelihood method

• Wald method ( 1.96 depends on unbiasedness of estimate

Gustafson&Greenland, StatScience 2009
Georg Heinze – Logistic regression with rare events
CeMSIIS-Section for Clinical Biometrics 48
Conclusion

• We recommend FLAC for:

• Good performance

• Invariance to transformations or coding

• Cannot be ‘outsmarted’ by creative coding

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 49
References
• Heinze G, Schemper M. A solution to the problem of separation in logistic regression. Statistics in
Medicine 2002
• Mansournia M, Geroldinger A, Greenland S, Heinze G. Separation in logistic regression – causes,
consequences and control. American Journal of Epidemiology, 2017.
• Puhr R, Heinze G, Nold M, Lusa L, Geroldinger A. Firth‘s logistic regression with rare events – accurate
effect estimates and predictions? Statistics in Medicine 2017.

Please cf. the reference lists therein for all other citations of this presentation.

Further references:
• Gustafson P, Greenland S. Interval estimation for messy observational data. Statistical Science 2009,
24:328-342.
• Rainey C. Estimating logit models with small samples. www.carlislerainey.com/papers/small.pdf (27
March 2017)

Georg Heinze – Logistic regression with rare events

CeMSIIS-Section for Clinical Biometrics 50