Bergtold, Jason S.Pokharel, Krishna P.Featherstone, Allen M.Mo, Lijia2022-07-142022-07-142018https://hdl.handle.net/2097/42354The numerical reliability of statistical software packages was examined for logistic regression models, including SAS 9.4, MATLAB R2015b, R 3.3.1., Stata/IC 14, and LIMDEP 10. Thirty unique benchmark datasets were created by simulating alternative conditional binary choice processes examining rare events, near-multicollinearity, quasi-separation and nonlinear transformation of variables. Certified benchmark estimates for parameters and standard errors of associated datasets were obtained following standards set-out by the National Institute of Standards and Technology. The logarithm of relative error was used as a measure of accuracy for numerical reliability. The paper finds that choice of software package and procedure for estimating logistic regressions will impact accuracy and use of default settings in these packages may significantly reduce reliability of results in different situations.This version of the article has been accepted for publication, after peer review (when applicable) and is subject to Springer Nature’s AM terms of use, but is not the Version of Record and does not reflect post-acceptance improvements, or any corrections.https://rightsstatements.org/vocab/InC/1.0/https://perma.cc/KDW9-RWNUAccuracyBenchmark datasetsEconometric softwareLogistic regressionMaximum likelihood estimationText