TY - JOUR
T1 - Revising a prognostic index developed for classification purposes
T2 - An application to gastric cancer data
AU - Miceli, Rosalba
AU - Lusa, Lara
AU - Mariani, Luigi
PY - 2004/8
Y1 - 2004/8
N2 - A prognostic index (PI) is usually derived from a regression model as a weighted mean of the covariates, with weights (partial scores) proportional to the parameter estimates. When a PI is applied to patients other than those considered for its development, the issue of assessing its validity on the new case series is crucial. For this purpose, Van Houwelingen (2000) proposed a method of validation by calibration, which limits overfitting by embedding the original model into a new one, so that only a few parameters will have to be estimated Here we address the problem of PI validation and revision with the above approach when the PI has classification purposes and it represents the linear predictor of a Weibull model, derived from an accelerated failure time parameterization instead of a proportional hazards one, as originally described by Van Houwelingen. We show that the Van Houwelingen method can be applied in a straightforward manner, provided that the parameterization originally used in the PI model is appropriately taken into account. We also show that model validation and revision can be carried out by modifying the cut-off values used for prognostic grouping without affecting the partial scores of the original PI. This procedure can be applied to simplify the clinician's use of an established PI for classification purposes.
AB - A prognostic index (PI) is usually derived from a regression model as a weighted mean of the covariates, with weights (partial scores) proportional to the parameter estimates. When a PI is applied to patients other than those considered for its development, the issue of assessing its validity on the new case series is crucial. For this purpose, Van Houwelingen (2000) proposed a method of validation by calibration, which limits overfitting by embedding the original model into a new one, so that only a few parameters will have to be estimated Here we address the problem of PI validation and revision with the above approach when the PI has classification purposes and it represents the linear predictor of a Weibull model, derived from an accelerated failure time parameterization instead of a proportional hazards one, as originally described by Van Houwelingen. We show that the Van Houwelingen method can be applied in a straightforward manner, provided that the parameterization originally used in the PI model is appropriately taken into account. We also show that model validation and revision can be carried out by modifying the cut-off values used for prognostic grouping without affecting the partial scores of the original PI. This procedure can be applied to simplify the clinician's use of an established PI for classification purposes.
KW - Gastric cancer
KW - Prognostic index
KW - Survival analysis
KW - Weibull model
UR - http://www.scopus.com/inward/record.url?scp=4444235490&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=4444235490&partnerID=8YFLogxK
U2 - 10.1080/0266476042000214510
DO - 10.1080/0266476042000214510
M3 - Article
AN - SCOPUS:4444235490
VL - 31
SP - 817
EP - 830
JO - Journal of Applied Statistics
JF - Journal of Applied Statistics
SN - 0266-4763
IS - 7
ER -