Model Averaging



København og omegn

When several models are under consideration, the standard approach is to select a single model based on some principle. Classical principles are likelihood testing (which is the same as backwards stepwise); AIC or alternatively forward stepwise. However, there may be some value in using several models in parallel and averaging the results (like the estimates of regression coefficients). Weights may depend on fit or be chosen beforehand. This may give a better prediction. Example 1: Regression with several potential covariates. Suppose we want to predict the body weight of a child and have information on sex, age and height. Using all three covariates should in theory give the best prediction, but could have extra uncertainty due to the high correlation between age and height. So one could fit three models; sex and age; sex and height; sex, age and height and then average the three predictions. Example 2: Dose finding with alternative dose-response models. In this case, we may have alternative models, say as means f(x) at dose x, the model could be of the Emax type (f(x)=α + β x /(x+γ)) or the exponential type (f(x)= α + β (1- exp(-x γ)). Suppose conventional fit comparisons do not give a clear conclusion as to which is best. Averaging the two predictions could be better than either one. This research area has been pioneered by Claeskens and Hjort.

The project:

The project consists of studying the statistical properties of the predictions. This includes how to weight two or more models to give the most meaningful prediction. A second question is to which extent this approach can be used for testing (that is, whether it satisfies the intended significance level).

Download materiale

Søg i opslag


DTU er et teknisk eliteuniversitet med international rækkevidde og standard. Vores mission er at udvikle og nyttiggøre naturvidenskab og teknisk videnskab til gavn for samfundet. 10.000 studerende uddanner sig her til fremtiden, og 5.700 medarbejdere har hver dag fokus på uddannelse, forskning, myndighedsrådgivning og innovation, som bidrager til øget vækst og velfærd.

Find os her

Anker Engelunds Vej 1
Bygning 101A
2800 Kgs. Lyngby

45 25 25 25

CVR-nr. 30 06 09 46

Liste over EAN Numre

Job på DTU

Se alle jobs