Successfully performed Ordinary Least Squares - Whats next?

mben

Hi, 

 

Here is a picture of a snippet of the dataset I am working with:

 

p and k are assumed to be dependent variables, dependent on the rest of the parameters. My goal is to get the best possible polynomial p = f(x1,x2,x3...) to model the relationships between my data. Same thing for x.

 

So far, I followed a cookie cutter tutorial on how to perform OLS and got a very basic linear polynomial for my data. The Adj. RSquared ended up being around 0.18, which is pretty poor. I want to improve this!

 

I have very little statistical experience - so could someone please point me in the right direction for learning how to improve the polynomial I already have. In a previous post, a commenter mentioned transforming my data. I did a little searching, but could not figure out the best way to determine what transformations to execute to improve my polynomial. 

 

Thank you for your time!

 

P.S. If anyone would be willing to discord screenshare for like 30 minutes to quickly work with me and teach me related JMP/stats stuff so I can learn to solve my problem, I would compensate you for your time.

 

 

 

statman

Welcome to the JMP community.  I don't understand your situation, but it sounds like you are trying to build a model based just on the data?  You don't provide any information about how the data was obtained.  And now I see you have started multiple threads on the same topic???

Here are my thoughts for developing a model:

1. Before any data has been collected, develop hypotheses, rooted in the sciences, the explain why there would be a potential causal relationship between the independent variables and the dependent variables.

2. Design multiple sampling plans or experiments to provide insight to those hypotheses considering how each would vary in resources necessary, what effects are possible to estimate, what will be confounded and what will be restricted.

3. Predict ALL possible outcomes of such a data collection strategies and what action you will take for each prediction.  Then choose the one that balances the knowledge gain with the resource investment.

4. Collect the data noting any observations of interest and perform analysis. Always Practical, then graphical and lastly quantitative.

How the data was acquired has an effect on how you build the model.  For example, I typically do a subtractive approach when the data is from an experiment (start with a saturated model and remove terms with an eye on is it reasonable from an engineering or scientific perspective?, Rsquare-Rsquare adjusted delta, RMSE, p-values and residuals).  If it is from a nested study, I start at the bottom and work my way up.  If it is just observational data, I might take a more additive approach (e.g., step wise of start with a first order and add terms) and of course you need to worry about multicollinearity.