Chapter 5 Flashcards
State the assumptions of linear-regression modelling
- There is a linear relationship between the outcome and predictor variables
- The variance of residuals is the same for any value of the predictor variable (this is known as homoscedasticity)
- There is independence between observations
- Residuals at any predictor value are normally distributed (normality)
What plot can be used to check linearity and homoscedasticity
Residuals vs Fitted plot:
plot of fitted values vs residuals
What plot can be used to check normality
Normal Q-Q plot
Describe DV vs PRED diagnostic plots for a non-linear mixed-effects model
Dependent variable (DV) vs Population predictions (PRED):
DV vs PRED takes predictions based on population parameters (i.e. excluding πβs)
Describe DV vs IPRED diagnostic plots for a non-linear mixed-effects model
Dependent variable (DV) vs Individual predictions (IPRED):
DV vs IPRED bases predictions on individual values for model parameters (including πβs).
DV vs IPRED takes into account random error associated with an individual.
Describe CWRES
Conditionally-weighted residuals (CWRES)
Normalising residuals through a weighted step.
This is important because residuals might be influenced by the magnitude of the observation (e.g. you might find a greater range of residuals around πΆπππ₯ compared to around πΆπππ). Weighting the residuals in order to standardise them makes it easier to think about what is happening with model predictions across the full range of observations. We will use conditionally weighted residuals (CWRES).
How is the objective function value useful in quantitative evaluation of model fit
Using qualitative metric of improved model fit: OFV is a measure of model fit.
Objective function can be used to compare two models.
A lower OFV indicates a superior model fit.
Define the Akaike Information Criteria (AIC) and describe how it may be used to select a superior model
The Akaike Information Criteria (AIC) is a way of determining whether the magnitude of a drop in objective function is worthy of retaining the more complex model.
π΄πΌπΆ = (ππ΅ππ πππππ β ππ΅ππππππππ₯) + 2π
Describe how AIC may be used to select a superior model
π΄πΌπΆ = (ππ΅ππ πππππ β ππ΅ππππππππ₯) + 2π
A negative π΄πΌπΆ provides (some) justification for retaining the more complex model, alongside improved diagnostic plots.
An ideal DV vs PRED
A degree of even spread of points above and below the line of unity.
A line of best fit that tracks evenly along the line of unity
(Any groups of data points deviating away from this line of unity might make us suspicious that there is some underlying variability not being modelled).