Is it okay to transform categorical data into weighted percentages?

keffinger

I am comparing the efficiency of ovens by measuring consistency in cookie baking (well, not really, I'm just using this for an example). After summarizing data, I have two responses, "Cooked Evenly" and "Cooked Unevenly".

 

Can I create a new column "% Cooked Evenly", run Fit X by Y where Y = %cooked evenly, X = oven, and Freq = Batches of Cookies? In doing this, I think I'm giving weight (from number of batches) to the percentages. Running data this way allows me to run means/anova and student t test for connecting letter's report, making it easy to compare several ovens at the same time. 

 

Normally, I stack the "Cooked evenly" and "Cooked Unevenly" and run a contingency analysis (observing proportions) on two ovens at a time, finding out which are significantly different from each other. I often find it difficult to get a nice statistics overview when I use contingency analysis on multiple ovens at the same time.

 

After running data in both scenarios, I get the same overview of which ovens are significantly outperforming other ovens, but I get the information in a fraction of the time by giving "weight" to a percentage.

 

Is this an acceptable way to approach analyzing categorical data?? Any and all criticisms are welcome! Thanks!

Oven Cooking Test.jmp

Mark_Bailey

You have one response with two outcomes. Why not use logistic regression? This method is based on the proportions/probabilities as you describe but you don't have to manually transform the data. Make sure that the response column uses the nominal modeling type, then select Analyze > Fit Model and enter this column in the Y role. It will automatically switch to nominal logistic regression. Then build the linear predictor with any effects that you want to test.

The analysis platform will provide all the features that you need including the Prediction Profiler.

So to be clear, you do not need to summarize the outcomes and transform the response to proportions. Logistic regression will take care of everything for you.