TY - JOUR
T1 - One model, several results
T2 - the paradox of the Hosmer-Lemeshow goodness-of-fit test for the logistic regression model.
AU - Bertolini, G.
AU - D'Amico, R.
AU - Nardi, D.
AU - Tinazzi, A.
AU - Apolone, G.
PY - 2000
Y1 - 2000
N2 - BACKGROUND: The Hosmer-Lemeshow test, used extensively to assess the fit of the logistic regression model, is performed by several statistical packages. Recent studies have shown some problems in the use of this test when ties are present. These problems were attributed merely to the test implementation. METHODS: We analysed the order of the observations as an alternative explanation of the problem of ties. Using a data-set of 1393 intensive care unit (ICU) patients we performed the Hosmer-Lemeshow test with all possible subjects dispositions. RESULTS: We obtained about one million different P values, ranging from 0.01 to 0.95. DISCUSSION: It is already known that when the Hosmer-Lemeshow goodness-of-fit test is performed with a number of covariate patterns lower than the number of subjects, its result may be inaccurate. We showed that the extent of this problem could be relevant under particular conditions. We also suggest a strategy for estimating the extent of the problem and subsequent interpretation.
AB - BACKGROUND: The Hosmer-Lemeshow test, used extensively to assess the fit of the logistic regression model, is performed by several statistical packages. Recent studies have shown some problems in the use of this test when ties are present. These problems were attributed merely to the test implementation. METHODS: We analysed the order of the observations as an alternative explanation of the problem of ties. Using a data-set of 1393 intensive care unit (ICU) patients we performed the Hosmer-Lemeshow test with all possible subjects dispositions. RESULTS: We obtained about one million different P values, ranging from 0.01 to 0.95. DISCUSSION: It is already known that when the Hosmer-Lemeshow goodness-of-fit test is performed with a number of covariate patterns lower than the number of subjects, its result may be inaccurate. We showed that the extent of this problem could be relevant under particular conditions. We also suggest a strategy for estimating the extent of the problem and subsequent interpretation.
UR - http://www.scopus.com/inward/record.url?scp=0034567668&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=0034567668&partnerID=8YFLogxK
M3 - Article
C2 - 11055275
AN - SCOPUS:0034567668
VL - 5
SP - 251
EP - 253
JO - Journal of Epidemiology and Biostatistics
JF - Journal of Epidemiology and Biostatistics
SN - 1359-5229
IS - 4
ER -