nav[aria-label="Primary Navigation"] { padding: 0; & ul { list-style: none; width: 100%; display: flex; flex-direction: row; justify-content: start; align-items: start; gap: 30px; padding: 0; & li { margin: 0; } & ul li { list-style: none; } } }

model equivalency with SAS

Woodstock

HI there,

I have a mixed effects model which I am trying to show equivalence with SAS for... Was hoping to get a better understanding of how to do that...

JMP:

Fit Model(
Y( :log_val ),
Effects( :Group, :Analyst ),
Random Effects( :Sample[:Group] ),
Personality( "Standard Least Squares" ),
Emphasis( "Effect Leverage" ),
Method( "REML" ),
NoBounds( 0 )
)

SAS:

proc mixed data=WEBWORK.MYDATA method=reml plots=(residualPanel) alpha=0.05;
class Group Analyst Sample;
model log_val=Group Analyst / solution;
random Sample(Group) /;
run;

My parameter estimates for fixed params is off by a factor of 2 in some cases

Accepted answers

All comments

ron_horne

Hi @Woodstock

my best guess is that your issue is caused by the way jmp builds the contrast across ordinal and nominal independent variables categories.

have a look at these two links:

https://www.jmp.com/support/help/14-2/coding-for-nominal-effects.shtml

https://www.jmp.com/support/help/14-2/models-with-nominal-and-ordinal-effects.shtml

first compare models with just continuous independents. then, when you add binary independents define them as ordinal in JMP. this should produce parameter estimates that are equivalent to SAS (in the case of 0 and 1).

if you have a nominal effect with many categories, you can produce n-1 dummy variables in JMP (as is done the old fashion way in SAS) and define each one of them as ordinal.

since you mentioned the factor of 2 this is most likely the case.

you can save the prediction formula to the table or ask for the prediction expression. this may help clarify how JMP is estimating the equation.

https://www.jmp.com/support/help/14-2/show-prediction-expression.shtml

also have a look here:

https://community.jmp.com/t5/Discussions/Logit-why-aren-t-my-exponentiated-coefficients-equaling-my-odds/td-p/15661

overall, prediction and conclusions should be identical between SAS and JMP (or any other software for standard procedures).

Mark_Bailey

See Help > Books > Fitting Linear Models > Statistical Details. It explains the differences between the way JMP and SAS parameterize linear models. I would not expect the estimates to be the same. They estimate different parameters.

Woodstock

Hi Mark,

Thanks for the reply. Help me understand what you mean by "I would not expect the estimates to be the same. They estimate different parameters.", Given I need to show equivalency between parameter estimation, what are the differences, and how important are they to the overall, and individual estimations?

I found this: https://www.jmp.com/support/help/14-2/nominal-factors.shtml#65535 from your Help menus breadcrumb suggestion, but I'm not sure how to use it to obtain equivalence between the two platforms.

Thanks for your time

Woodstock

Hi @ron_horne ,

Thanks for your reply. I'm wondering what you mean by the statement: "you can produce n-1 dummy variables in JMP (as is done the old fashion way in SAS) and define each one of them as ordinal."... specifically, how do I do this?

When I export the model specification directly from JMP as SAS code, I don't get the same results. See attached

jmp-export-to-sas.txt

ron_horne

Hi @Woodstock

as for creating dummy variables from a categorical column try the following:

https://community.jmp.com/t5/JMP-Academic-Knowledge-Base/Make-Indicator-Dummy-Variables/ta-p/22636

with respect to the output that you sent, perhaps someone else will answer.

ron

Mark_Bailey

The models are equivalent in their predictions. Given the same predictor values, both models predict the same response. The models are different in their parameterization. I do not expect the same parameter values.

If I fit a simple linear model with one term for a continuous predictor, then center and scale the predictor, and re-fit the model with the transformed predictor, the parameters are necessarily different yet the two models predict the same response. The parameters are different. The interpretation of the parameters is different.

Screen Shot 2019-11-02 at 6.34.10 AM.png

I do not like the previous reference to the "old fashioned" parameters in SAS. They are simply different. They represent a different interpretation and test of the effects of changing the factor level. No parameterization is inherently superior. The best parameterization depends on how you think about the effects. The choice is situational.

This issue is the same as the one about different types of sum of squares. Most of the time I prefer Type III. That doesn't mean that Type I are "old fashioned" or wrong or of no use. There are times when they provide exactly the right information, like the test for poolabilith of multiple batches in determining the expiry of a drug. It just depends.