(Tuesday, 7/31/2018) 11:00 AM - 12:15 PM
Room: Karl Dean Ballroom A2
Prediction modeling is distinct from almost all areas in biostatistics. Typically, we apply statistical methods to describe a data set; in prediction modeling, we aren’t interested in the data set we use to create a model, our focus is solely on how well we will be able to predict a future set of observations. This leads to our first golden rule: “Don’t like a coefficient? Feel free to change it to something else�. Of course, advocates of machine learning might tell us that we don’t need models or coefficients at all, that a neural net or random forest will solve all our prediction problems. As it turns out, however, the answer to the question “are machine learning methods better than traditional logistic or Cox models?� is “well, it depends�. This leads to our second golden rule: “No, a road bike isn’t necessarily faster than a mountain bike�. Whatever type of analytic approach is used, analysts must choose how to enter variables into the analysis. This is a particular issue with continuous variables like body mass index, or hemoglobin, which are often dichotomized (e.g. obese / not obese or anemic / not anemic). It turns out that analysts would often be better off if they followed the rule: “It doesn’t matter if the hotel is too expensive�. Our fourth golden rule refers to the need to develop and evaluate models across multiple settings. So golden rule five is: “What happened to Dr. Scardino’s patients in Houston in the 1990s is not the sum total of cancer biology�. One key characteristic of prediction models is that they are explicitly designed to influence decisions. A model to predict adverse reactions in terms of dose, for instance, might influence a decision about the right dose to use for a particular patient. How we evaluate models must therefore take a decision-analytic perspective. Accordingly, our final golden rule is: “Accuracy doesn’t matter�.
Learning Objectives:
1. To understand prediction modelling and what it entails.
2. To know how to evaluate statistical models.
3. To understand how to select the right statistical model for a particular study.
Handouts
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
