Detect/check for separation and infinite maximum likelihood estimates in logistic regression

Ioannis Kosmidis and Dirk Schumacher

5 January 2020

The detectseparation package

detectseparation provides pre-fit and post-fit methods for the detection of separation and of infinite maximum likelihood estimates in binomial response generalized linear models.

The key methods are detect_separation and check_infinite_estimates and this vignettes describes their use.

Checking for infinite estimates

Heinze and Schemper (2002) used a logistic regression model to analyze data from a study on endometrial cancer (see, Agresti 2015, Section 5.7 or ?endometrial for more details on the data set). Below, we refit the model in Heinze and Schemper (2002) in order to demonstrate the functionality that detectseparation provides.

library("detectseparation")
data("endometrial", package = "detectseparation")
endo_glm <- glm(HG ~ NV + PI + EH, family = binomial(), data = endometrial)
theta_mle <- coef(endo_glm)
summary(endo_glm)
#> 
#> Call:
#> glm(formula = HG ~ NV + PI + EH, family = binomial(), data = endometrial)
#> 
#> Deviance Residuals: 
#>      Min        1Q    Median        3Q       Max  
#> -1.50137  -0.64108  -0.29432   0.00016   2.72777  
#> 
#> Coefficients:
#>               Estimate Std. Error z value Pr(>|z|)    
#> (Intercept)    4.30452    1.63730   2.629 0.008563 ** 
#> NV            18.18556 1715.75089   0.011 0.991543    
#> PI            -0.04218    0.04433  -0.952 0.341333    
#> EH            -2.90261    0.84555  -3.433 0.000597 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> (Dispersion parameter for binomial family taken to be 1)
#> 
#>     Null deviance: 104.903  on 78  degrees of freedom
#> Residual deviance:  55.393  on 75  degrees of freedom
#> AIC: 63.393
#> 
#> Number of Fisher Scoring iterations: 17

The maximum likelihood (ML) estimate of the parameter for NV is actually infinite. The reported, apparently finite value is merely due to false convergence of the iterative estimation procedure. The same is true for the estimated standard error, and, hence the value r round(coef(summary(endo_glm))["NV", "z value"], 3) for the \(z\)-statistic cannot be trusted for inference on the size of the effect for NV.

Lesaffre and Albert (1989, Section 4) describe a procedure that can hint on the occurrence of infinite estimates. In particular, the model is successively refitted, by increasing the maximum number of allowed iteratively re-weighted least squares iterations at east step. The estimated asymptotic standard errors from each step are, then, divided to the corresponding ones from the first fit. If the sequence of ratios diverges, then the maximum likelihood estimate of the corresponding parameter is minus or plus infinity. The following code chunk applies this process to endo_glm.

(inf_check <- check_infinite_estimates(endo_glm))
#>       (Intercept)           NV       PI       EH
#>  [1,]    1.000000 1.000000e+00 1.000000 1.000000
#>  [2,]    1.424352 2.092407e+00 1.466885 1.672979
#>  [3,]    1.590802 8.822303e+00 1.648003 1.863563
#>  [4,]    1.592818 6.494231e+01 1.652508 1.864476
#>  [5,]    1.592855 7.911035e+02 1.652591 1.864492
#>  [6,]    1.592855 1.588973e+04 1.652592 1.864493
#>  [7,]    1.592855 5.298760e+05 1.652592 1.864493
#>  [8,]    1.592855 2.332822e+07 1.652592 1.864493
#>  [9,]    1.592855 2.332822e+07 1.652592 1.864493
#> [10,]    1.592855 2.332822e+07 1.652592 1.864493
#> [11,]    1.592855 2.332822e+07 1.652592 1.864493
#> [12,]    1.592855 2.332822e+07 1.652592 1.864493
#> [13,]    1.592855 2.332822e+07 1.652592 1.864493
#> [14,]    1.592855 2.332822e+07 1.652592 1.864493
#> [15,]    1.592855 2.332822e+07 1.652592 1.864493
#> [16,]    1.592855 2.332822e+07 1.652592 1.864493
#> [17,]    1.592855 2.332822e+07 1.652592 1.864493
#> [18,]    1.592855 2.332822e+07 1.652592 1.864493
#> [19,]    1.592855 2.332822e+07 1.652592 1.864493
#> [20,]    1.592855 2.332822e+07 1.652592 1.864493
#> attr(,"class")
#> [1] "inf_check"
plot(inf_check)

Clearly, the ratios of estimated standard errors diverge for NV.

Detecting separation

detect_separation tests for the occurrence of complete or quasi-complete separation in datasets for binomial response generalized linear models, and finds which of the parameters will have infinite maximum likelihood estimates. detect_separation relies on the linear programming methods developed in the 2017 PhD thesis by Kjell Konis (Konis 2007).

detect_separation is pre-fit method, in the sense that it does not need to estimate the model to detect separation and/or identify infinite estimates. For example

endo_sep <- glm(HG ~ NV + PI + EH, data = endometrial,
                family = binomial("logit"),
                method = "detect_separation")
endo_sep
#> Implementation: ROI | Solver: lpsolve 
#> Separation: TRUE 
#> Existence of maximum likelihood estimates
#> (Intercept)          NV          PI          EH 
#>           0         Inf           0           0 
#> 0: finite value, Inf: infinity, -Inf: -infinity

The detect_separation method reports that there is separation in the data, that the estimates for (Intercept), PI and EH are finite (coded 0), and that the estimate for NV is plus infinity. So, the actual maximum likelihood estimates are

coef(endo_glm) + coef(endo_sep)
#> (Intercept)          NV          PI          EH 
#>   4.3045178         Inf  -0.0421834  -2.9026056

and the estimated standard errors are

coef(summary(endo_glm))[, "Std. Error"] + abs(coef(endo_sep))
#> (Intercept)          NV          PI          EH 
#>  1.63729861         Inf  0.04433196  0.84555156

We can also use the glpk solver for solving the linear program for separation detection

update(endo_sep, solver = "glpk")
#> Implementation: ROI | Solver: glpk 
#> Separation: TRUE 
#> Existence of maximum likelihood estimates
#> (Intercept)          NV          PI          EH 
#>           0         Inf           0           0 
#> 0: finite value, Inf: infinity, -Inf: -infinity

or use the implementation using lpSolveAPI directly

update(endo_sep, implementation = "lpSolveAPI")
#> Implementation: lpSolveAPI | Linear program: primal | Purpose: find 
#> Separation: TRUE 
#> Existence of maximum likelihood estimates
#> (Intercept)          NV          PI          EH 
#>           0         Inf           0           0 
#> 0: finite value, Inf: infinity, -Inf: -infinity

See ?detect_separation_control for more options.

As proven in (Kosmidis and Firth 2021), an estimator that is always finite, regardless whether separation occurs on not, is the reduced-bias estimator of (Firth 1993), which is implemented in the brglm2 R package.

library("brglm2")
#> Registered S3 method overwritten by 'brglm2':
#>   method                  from            
#>   print.detect_separation detectseparation
#> 
#> Attaching package: 'brglm2'
#> The following objects are masked from 'package:detectseparation':
#> 
#>     checkInfiniteEstimates, check_infinite_estimates, detectSeparation,
#>     detectSeparationControl, detect_separation,
#>     detect_separation_control
summary(update(endo_glm, method = "brglm_fit"))
#> 
#> Call:
#> glm(formula = HG ~ NV + PI + EH, family = binomial(), data = endometrial, 
#>     method = "brglm_fit")
#> 
#> Deviance Residuals: 
#>     Min       1Q   Median       3Q      Max  
#> -1.4740  -0.6706  -0.3411   0.3252   2.6123  
#> 
#> Coefficients:
#>             Estimate Std. Error z value Pr(>|z|)    
#> (Intercept)  3.77456    1.48869   2.535 0.011229 *  
#> NV           2.92927    1.55076   1.889 0.058902 .  
#> PI          -0.03475    0.03958  -0.878 0.379914    
#> EH          -2.60416    0.77602  -3.356 0.000791 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#> 
#> (Dispersion parameter for binomial family taken to be 1)
#> 
#>     Null deviance: 104.903  on 78  degrees of freedom
#> Residual deviance:  56.575  on 75  degrees of freedom
#> AIC: 64.575
#> 
#> Number of Fisher Scoring iterations: 6

Citation

If you found this vignette or detectseparation useful, please consider citing detectseparation. You can find information on how to do this by typing citation("detectseparation").

References

Agresti, A. 2015. Foundations of Linear and Generalized Linear Models. Wiley Series in Probability and Statistics. Wiley.
Firth, D. 1993. “Bias Reduction of Maximum Likelihood Estimates.” Biometrika 80 (1): 27–38. https://doi.org/10.1093/biomet/80.1.27.
Heinze, G., and M. Schemper. 2002. “A Solution to the Problem of Separation in Logistic Regression.” Statistics in Medicine 21: 2409–19.
Konis, K. 2007. “Linear Programming Algorithms for Detecting Separated Data in Binary Logistic Regression Models.” DPhil, University of Oxford. https://ora.ox.ac.uk/objects/uuid:8f9ee0d0-d78e-4101-9ab4-f9cbceed2a2a.
Kosmidis, I., and D. Firth. 2021. “Jeffreys-Prior Penalty, Finiteness and Shrinkage in Binomial-Response Generalized Linear Models.” Biometrika 108 (1): 71–92.
Lesaffre, E., and A. Albert. 1989. “Partial Separation in Logistic Discrimination.” Journal of the Royal Statistical Society. Series B (Methodological) 51 (1): 109–16. http://www.jstor.org/stable/2345845.