I'm trying to generate a prediction equation based on the model selection. If I have non-significant lack of fit on model 1(linear) and model 2 (quadratic), but model 2 has significant quadratic terms, which model should I select to generate prediction equation?
Hi
If the quadratic term is significant, it means that the relationship between exposure and outcome is non-linear, so you should keep the model with the quadratic term.
Hope it helps.
Hello Nischal,
you have the mathematics which give an objective basis to make decisions. But a lot of times the mathematics don't say it all. You also have to use some background of physiological information- if you are dealing with biological materials. In other words would you expect a linear relationship or non-linear between the predictor and response variables. Part of the answer is also in taking a look at the shape of your data on your graphics landscape.
Check for non linearity using BDS test.if non linearity is confirmed go for predictive non linearity model. Do confirm that all the tests like coefficient diagonistics; residual diagonistics and stability diagonistics with the linear model is done before procee
Plot the raw data. does it look linear or quadratic?
Are the residuals well behaved in both cases?
Does it matter? I could get an equation Y=B0 + B1*x1 + B2*x1^2 where B1=57 and B2=0.0000000007. Even if B2 is significant, it is not really adding a whole lot to the answer. I probably could not detect the effect by eye. Of course if making a mistake will cost hundreds of millions or kill thousands of people, then I might still leave the quadratic term in the model.
Is it really significant? Is the p-value for the quadratic term 0.049999, or is it 0.0000003?
Is the model with the quadratic really better? Do you get a greater adjusted R squared value? Would AIC values support including the quadratic term?
- Badges
- Science topic
- Similar topics
- Psychology