Created:
| Last Modified:
If you refer to Hierarchical Linear Regression then I don't think this a separate platform in JMP. That could be because there are fairly simple alternatives (at least based on my understanding of Hierarchical Linear Regression):
- adding/removing terms in the standard leas squares report,
- stepwise regression and,
- ensemble modeling.
Note that for ensemble modeling you can extend this beyond just linear regression, you could decide whether to include a partition after a liner regression.
Add/Remove Terms
The effect summary section of the least squares report has options to add/remove terms, you can compare your summary statistics just as is described in the stepwise section.
Stepwise
Instead of adding variables in blocks, you can add them individually in any order and just stop after adding each block. Using fit model add all of the possible terms in every group as model effects. For example, if you want to predict age in the big class data set from height and weight and then decide if you should add the height*weight term, you would add all three terms in the model launch dialog box. Then enter the terms in your first block, record model fit parameter's, and then add your second block and compare. It is possible that you can use the settings in the Stepwise Regression Control section to automatically get you to the model you would have chosen yourself.
In Fit Model launch add all model effects and select stepwise:
Add the first block's variables (I am ignoring the P value for weight):
Now add the second block and notice how the RSquare went up slightly, but so did the AIC and BIC, likely indicating that you would not want to include the second block.
Ensemble Modeling
In fit model, build a model using the first block and save the residuals. Then try to predict the residuals using subsequent models.
Build a model using the first block: Here I converted age to continuous to make the example simpler.
Save the predictions and the residuals:
Now launch fit model again and predict the residual from the first model using the second block, remember that neither of these needed to be linear models. You might get a message that you are missing an effect because JMP (correctly) doesn't want to build a model from the cross term without each individual term. In this example they were already included in the previous step but you could add it again.
Now you can use the results in this model to decide whether to include the second block in your final model. In this case the p value for both terms is very high, indicating that you should not include block 2.
To get the prediction using both models, just add the predictions from each:
