Mismatch between PROC GLIMMIX output predicted probability and results

confooseddesi89 — Thu, 19 Nov 2020 21:14:37 GMT

Hello,

I am running an analysis examining the association between sleep offset, aka rest end (disaggregated into within- and between-person components: restend_PMC and restend_PM, respectively) and odds of breakfast consumption (with adjustment for a covariate) in PROC GLIMMIX. See below the results:

Solutions for Fixed Effects
Effect	Estimate	Standard	DF	t Value	Pr > \|t\|
Error
Intercept	1.5619	0.6444	578	2.42	0.0157
restend_PMC	-0.06121	0.03114	2658	-1.97	0.0494
restend_PM	-0.0994	0.04668	578	-2.13	0.0337
Covariate	-0.6233	0.1477	2658	-4.22	<.0001

Odds Ratio Estimates
restend_PMC	restend_PM	Covariate	_restend_PMC	_restend_PM	_Covariate	Estimate	DF	95% Confidence Limits
1.0038	8.3449	5.00E-18	0.0038	8.3449	5.00E-18	0.941	2658	0.885	1
0.0038	9.3449	5.00E-18	0.0038	8.3449	5.00E-18	0.905	578	0.826	0.992
0.0038	8.3449	1	0.0038	8.3449	5.00E-18	0.536	2658	0.401	0.716

Type III Tests of Fixed Effects
Effect	Num DF	Den DF	F Value	Pr > F
restend_PMC	1	2658	3.86	0.0494
restend_PM	1	578	4.53	0.0337
Covariate	1	2658	17.81	<.0001

The odds ratios for restend_PMC and restend_PM (bolded above) are below 1, meaning that with each unit increase in either variable, the odds of breakfast go down (negative relationship). I've used "output out" within the code to produce predicted probability and standard error of the outcome within the model in order to graph my results. Below is my code:

Proc glimmix data=Data NOCLPRINT NOITPRINT METHOD= RSPL namelen=40;
CLASS idnum;model breakfast(ref='0')=restend_PMC restend_PM covariate / ODDSRATIO SOLUTION DIST=BINARY ddfm=bw;
output out=Sleep_offset_Break
pred(noblup ilink)= Sleep_offset_Break_Pred_PA
STDERR(noblup ilink)= Sleep_offset_Break_STDERR_PA;
RANDOM Intercept / TYPE=AR(1) Subject=idnum;
Title 'sleep_offset predicting breakfast, adjusted';run;title;run;

I graphed rest_end_PMC on the x-axis and the predicted probability of breakfast consumption (output from output out) on the y-axis, with dotted lines indicating the standard error (see image below):

To my surprise, the line is slightly upward, indicating a positive relationship (really, the line is almost straight). I have not had this issue when graphing other predicted results using the "output out" command in proc glimmix, including graphing the association between rest_end_PM and the predicted probability (the graph indicated a negative relationship as expected). Can anyone help me understand the issue?

Thanks.

confooseddesi89

Re: Mismatch between PROC GLIMMIX output predicted probability and results

SASJedi — Fri, 20 Nov 2020 16:16:41 GMT

Moved to Analytical Procedures forum where it should get better visibility

Re: Mismatch between PROC GLIMMIX output predicted probability and results

PaigeMiller — Fri, 20 Nov 2020 16:35:40 GMT

You don't say how you did the plot.

Nevertheless, you probably want to use the STORE command in PROC GLIMMIX to save the model so that PROC PLM can do the plot for you (using the EFFECTPLOT statement). I'm sure that will produce a more reasonable plot.

topic Re: Mismatch between PROC GLIMMIX output predicted probability and results in Statistical Procedures

Mismatch between PROC GLIMMIX output predicted probability and results

Re: Mismatch between PROC GLIMMIX output predicted probability and results

Re: Mismatch between PROC GLIMMIX output predicted probability and results