Table of Contents  

INSTRUCTIONS: For optimal educational effect, we suggest reading this paper first and answering the questions for yourself, then comparing your response with our answers. However, the questions can also serve as an illustration of a topic of personal interest, for example "missing values" or "validation."

This paper is quite extensive and provides a lot of statistical detail on several predictive modeling issues for logistic regression. To guide the reader through, we consider the 7 steps in the development of a prognostic model as applied in this paper and list a number of questions.

Click on any question below to see the answer.

1. Preliminary steps

a.	What was the aim of the analysis?
b.	How was an impression of the data under study obtained?
c.	How were missing values treated for the analysis?
d.	Survival was analyzed, but the regression model was a log-normal model rather than a Cox proportional hazards model. Why?

2. Coding of covariables

a.	How were diagnostic categories ('reason for admission') coded?
b	How were continuous covariables modeled?

3. Selection of covariables

a.	How were covariables selected for inclusion in the multivariable model?
b.	Were interaction terms assessed?

4. Estimation of regression coefficients

a.	How were the regression coefficients estimated?

5. Evaluation of performance

a.	Which performance measures were used in the paper?

6. Validation

a.	What type of validation was performed?
b.	What modeling steps were considered in the validation process?
c.	What modeling steps were NOT considered in the validation process?
d.	Was external validity studied?

7. Presentation of model

a.	How was the final model presented?
b.	What is the predicted 1-year survival for an 85-year-old patient as described in the text (p S21: Other disease, weight loss, APS 30, ADL 1), but GCS 9 instead of 14?

In conclusion, this study is a nice illustration of many of the concepts related to prognostic modeling, as discussed before in the chapter. Especially important is the fact that the selection of predictors was prespecified, which limited the model uncertainty. As the authors remark (p S21, S23), external validity of the model should be studied to show the value of the model in a more recent time and in other centers. Also, it would be interesting to know whether the model is indeed applied in practice for e.g. access to hospice services, insurance-related decisions, and medical decisions concerning diagnosis and treatment.