When:
Thursday, April 25, 2019
9:00 AM - 10:00 AM CT
Where: 680 N. Lake Shore Drive, Suite 1400, Stamler Conference Room, Chicago, IL 60611 map it
Audience: Faculty/Staff - Student - Post Docs/Docs - Graduate Students
Contact:
Putri Kusumo
(312) 908-1718
Group: Department of Preventive Medicine
Category: Academic
Pengcheng Lu, MS
Candidate for the Division of Biostatistics/BCC Faculty Position
PhD Candidate in Biostatistics
Graduate Research Assistant
Graduate Teaching Assistant
Department of Biostatistics
University of Kansas Medical Center
Presentation Title
A New Partition-based Goodness-of-fit Test for Logistic Regression Models with Continuous Covariates
Abstract
An important topic in logistic regression modeling exercise is an assessment of model fit, i.e. an assessment of goodness-of-fit, just as in any other models’ processing. It has been well-known that assessing goodness-of-fit in logistic regression models by using the Deviance or Pearson’s chi-square statistics can be problematic when sparse data especially when continuous covariates are presented. The Hosmer-Lemeshow (HL) test was introduced to overcome the issue and it has become the industry standard, however the HL test is criticized for at least two reasons, i.e. the fixed setting of the number of groups as 10 is arbitrary, and the test is sensitive to the chosen number of groups. We propose an asymptotic theorem-driven new partition method to classify observations into distinct groups according to fitted probabilities similarly to that incorporated by the HL test. The new partition-based goodness-of-fit test ensures sufficient expected cell frequencies for valid chi-square testing. A variety of simulations are performed comparing the proposed test to the HL test and two more alternative methods, namely Copas’s unweighted residual sum of squares (USS) and cumulative sums of residuals (CUSUM), some recommendations are provided based on our research.