topic I am receiving Lost DFs, Biased and Zeroed parameter estimate messages, what do they mean? in Discussions

I am receiving Lost DFs, Biased and Zeroed parameter estimate messages, what do they mean?

escheenstra — Mon, 06 Nov 2017 16:04:43 GMT

First post here as I'm relatively new to JMP. I've been trying to run an 3 way effects test in the fit model option and am getting errors saying "Lost DFs". Also parameter estimates gives "biased" and "zeroed" results. In fit model I'm basically selecting my 3 effects which are year (2013, 2014), type (heirloom or standard) and variety (20 varieties with 4 samples each) and then choosing "full factorial" and then running the standard least squares with effect leverage selected as the emphasis. The data itself consists of 20 bean varieties with different test results like yield, height, weights, etc.... When I run a 2-way with year*type or year*variety there's no problem but when adding the third effect is when the problems happen... I guess I'm just wanting to know more about these errors and if there is something else I should be doing that I'm not.

Re: Lost DFs?

Peter_Bartell — Mon, 10 Aug 2015 16:44:26 GMT

Basically I think what is happening when you see the "biased" and "zeroed" results is you are specifying more terms in the model than you have sufficient degrees of freedom to estimate. So JMP gives you these messages...kind of a gentle reminder to insure that the model you are specifying can indeed be estimated by the data you have collected. You may want to check out the "Fitting Linear Models" JMP Help book devoted to this very broad topic.

Re: Lost DFs?

escheenstra — Mon, 10 Aug 2015 17:13:06 GMT

Thanks for the help, I will definitely check that out. It doesn't sound like running that model is possible with my current data set as-is unless there are some other settings or options that will provide some kind of a "work around".

Re: Lost DFs?

Peter_Bartell — Mon, 10 Aug 2015 18:11:31 GMT

escheenstra: From your original post/question it sounds like you initially tried to use the Analyze -> Fit Model -> Standard Least Squares modeling subpersonality. There are alternative regression based modeling techniques that can be used when you don't have the luxury of sufficient degrees of freedom to estimate all the terms you'd like to estimate under that path that still might provide the ultimate answers to the practical questions at hand. For example, partial least squares is a technique well suited for the wide and shallow problem...or if you are running JMP Pro then definitely check out the Generalized Regression personality.

Re: Lost DFs?

Phil_Kay — Tue, 11 Aug 2015 12:57:06 GMT

Just to add to what Peter said, by my reckoning the full model (for 2 * 2-level and 1 * 20-level) with intercept, main effects, 2-factor interactions and 3-factor interactions will have 80 parameters in it. From your description it sounds like you have 80 observations. So you will have a saturated model. In addition, depending on how the data was designed, it may not be possible to estimate some of these parameters because certain combinations of factor settings were not observed.

Re: Lost DFs?

jenkins_macedo — Mon, 17 Aug 2015 14:51:59 GMT

I would recommend that you do series of Y by X fit on each of the variable and see what the outputs give you. Bear in mind that depending on your data type (numeric,,,continuous, character...nominal, etc) the selected analysis that will be conducted will vary and that depends on your data type.

Re: Lost DFs?

David_Burnham — Tue, 01 Sep 2015 17:36:01 GMT

There is useful information in the "biased" and "zeroed" messages. Lets say there are two terms in your model A and B (I know you are looking at 3-way interactions but this is just easier for explanation). If the 2 terms are collinear (aliased, correlated, take your pick of terminology) then statistically the effects of A and B can not be estimated independently of each other. This problem can be overcome by putting either A or B (but not both) in the model. Effectively this is what JMP is doing for you - if it puts A in the model and excludes B then A will be labelled as biased and B will be labelled as zeroed.