The brglm2 package

Along with methods for improving the estimation of generalized linear models (see iteration vignette), brglm2 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 analyse data from a study on endometrial cancer. Agresti (2015, Section 5.7) provide details on the data set. Below, we fit a probit regression model with the same linear predictor as the logistic regression model in Heinze and Schemper (2002).

library("brglm2")
data("endometrial", package = "brglm2")
modML <- glm(HG ~ NV + PI + EH, family = binomial("probit"), data = endometrial)
theta_mle <- coef(modML)
summary(modML)
#>
#> Call:
#> glm(formula = HG ~ NV + PI + EH, family = binomial("probit"),
#>     data = endometrial)
#>
#> Deviance Residuals:
#>      Min        1Q    Median        3Q       Max
#> -1.47007  -0.67917  -0.32978   0.00008   2.74898
#>
#> Coefficients:
#>              Estimate Std. Error z value Pr(>|z|)
#> (Intercept)   2.18093    0.85732   2.544 0.010963 *
#> NV            5.80468  402.23641   0.014 0.988486
#> PI           -0.01886    0.02360  -0.799 0.424066
#> EH           -1.52576    0.43308  -3.523 0.000427 ***
#> ---
#> Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
#>
#> (Dispersion parameter for binomial family taken to be 1)
#>
#>     Null deviance: 104.90  on 78  degrees of freedom
#> Residual deviance:  56.47  on 75  degrees of freedom
#> AIC: 64.47
#>
#> Number of Fisher Scoring iterations: 17

As is the case for the logistic regression in Heinze and Schemper (2002), 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(modML))["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 IWLS iterations at east step. At east step the estimated asymptotic standard errors are 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 modML.

check_infinite_estimates(modML)
#>       (Intercept)           NV       PI       EH
#>  [1,]    1.000000 1.000000e+00 1.000000 1.000000
#>  [2,]    1.320822 1.710954e+00 1.352585 1.523591
#>  [3,]    1.410578 5.468665e+00 1.483265 1.617198
#>  [4,]    1.413505 3.225927e+01 1.490753 1.618462
#>  [5,]    1.413559 3.311464e+02 1.490923 1.618484
#>  [6,]    1.413560 5.786011e+03 1.490924 1.618484
#>  [7,]    1.413560 1.704229e+05 1.490924 1.618484
#>  [8,]    1.413560 3.232598e+06 1.490924 1.618484
#>  [9,]    1.413560 5.402789e+06 1.490924 1.618484
#> [10,]    1.413560 9.295902e+06 1.490924 1.618484
#> [11,]    1.413560 2.290248e+07 1.490924 1.618484
#> [12,]    1.413560 3.953686e+07 1.490924 1.618484
#> [13,]    1.413560 3.953686e+07 1.490924 1.618484
#> [14,]    1.413560 3.953686e+07 1.490924 1.618484
#> [15,]    1.413560 3.953686e+07 1.490924 1.618484
#> [16,]    1.413560 3.953686e+07 1.490924 1.618484
#> [17,]    1.413560 3.953686e+07 1.490924 1.618484
#> [18,]    1.413560 3.953686e+07 1.490924 1.618484
#> [19,]    1.413560 3.953686e+07 1.490924 1.618484
#> [20,]    1.413560 3.953686e+07 1.490924 1.618484

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 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

endometrial_sep <- glm(HG ~ NV + PI + EH, data = endometrial,
family = binomial("logit"),
method = "detect_separation")
endometrial_sep
#> 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(modML) + endometrial_sep$betas #> (Intercept) NV PI EH #> 2.18092821 Inf -0.01886444 -1.52576146 and the estimated standard errors are coef(summary(modML))[, "Std. Error"] + abs(endometrial_sep$betas)
#> (Intercept)          NV          PI          EH
#>  0.85732428         Inf  0.02359861  0.43307925

Citation

If you found this vignette or brglm2, in general, useful, please consider citing brglm2 and the associated paper. You can find information on how to do this by typing citation("brglm2").