Choose Language Hide Translation Bar
JBeierle
Occasional Contributor

What does it mean if both my predicted plot P value and my Prob > F are both signficant?

Hi,

I'm running datasets where I'm getting really nice RMSE (0.86-0.90) and P value < 0.0001 for my predicted plots, but my Prob > F is still signficant for Lack of Fit. What does this mean? Example pasted below.

 

Screen Shot 2019-07-31 at 11.01.51 AM.png

0 Kudos
1 ACCEPTED SOLUTION

Accepted Solutions

Re: What does it mean if both my predicted plot P value and my Prob > F are both signficant?

A model can be significant and explain a lot of variability and still exhibit lack of fit. For example, the model may be missing an important term, as shown in the attached simple example.

Example.PNG

 

A significant lack of fit usually indicates that the model form is not correct and usually indicates that there is a curvilinear relationship. Residual plots are the best and easiest tool to use to try and determine why there is a lack of fit. A good residual plot will show random scatter. Any pattern will indicate a systematic problem with your chosen model. For my simple example, here is a residual plot (there is more than one residual plot, so look at all of them to help determine the problem with your model).

Resid.PNG

There is an obvious pattern here indicating that I need a quadratic term in the model to estimate the curvature. I THINK this is a problem with your model as well, but the residual plots should help you.

Dan Obermiller

View solution in original post

2 REPLIES 2
JBeierle
Occasional Contributor

Re: What does it mean if both my predicted plot P value and my Prob > F are both signficant?

*sorry for the typo* - Rsq are really nice. NOT RMSE.
0 Kudos

Re: What does it mean if both my predicted plot P value and my Prob > F are both signficant?

A model can be significant and explain a lot of variability and still exhibit lack of fit. For example, the model may be missing an important term, as shown in the attached simple example.

Example.PNG

 

A significant lack of fit usually indicates that the model form is not correct and usually indicates that there is a curvilinear relationship. Residual plots are the best and easiest tool to use to try and determine why there is a lack of fit. A good residual plot will show random scatter. Any pattern will indicate a systematic problem with your chosen model. For my simple example, here is a residual plot (there is more than one residual plot, so look at all of them to help determine the problem with your model).

Resid.PNG

There is an obvious pattern here indicating that I need a quadratic term in the model to estimate the curvature. I THINK this is a problem with your model as well, but the residual plots should help you.

Dan Obermiller

View solution in original post