09-22-2017 08:49 AM
I have 25 raw data files that describe five vehicle tests of a set of five lubricant formulations where each automated test is 1,000 miles on a vehicle dynamometer. The combined dataset is nearly 300,000 rows of data. In the dataset, there are over 30 recorded operating parameters from various sensors on the vehicle. To fairly compare the lubricant formulations, it is desirable to create a model that removes the effects of the operating parameters that vary from test to test (experimental error). Of the 30 parameters recorded, some may have essentially the same effect, in a model, as they may be collinear or correlated.
What method in EG would best select the terms of the model, thereby allowing a fair comparison of the five lubricant formulations?
At present, with some subject matter expertise, I have manually selected terms and inspected their Pr > F values. With such a large dataset, these terms all show <.0001. However, I am concerned that I could be overspecifying the model. If there were a way to let the data define the model first, I would feel more comfortable.
How would you recommend evaluating the data?
09-22-2017 08:59 AM
Suggestion: instead of working from the point of view of selecting model terms, work from the point of view that empirically there is no logical way to select from correlated predictor variables. In other words, use a statistical methodology that accounts for these correlations between predictors and develops a good predictive model using ALL predictor variables.
This method is Partial Least Squares regression, or in SAS it is PROC PLS. Studies have shown you get a better predictive model (better meaning lower root mean square error of the predicted values and model coefficients) than if you do some sort of model selection technique. Reference: http://amstat.tandfonline.com/doi/abs/10.1080/00401706.1993.10485033#.WcUJHmdMqJA
09-25-2017 01:56 PM
Thank you for your informative reply and reference.
I am wondering here, if in EG 7.1, that I choose a PLS model rather than a linear model. I will see if I can create a branch off the data flow to another modeling method. I do not use SAS, so it has to be something I can see in EG.