Marginal Effects with Proc glimmix

[Jerem]

Many thanks to both of you!

I've been extremely busy over the last days so I couldn't work that much on the model. I've thus need to have a closer look at this later.

Now the model do not converge....I tried different dependent variable and this is always the same problem: "did not converge".

Is it a problem in the modelling or the data?

Best,

Jérémy

[SteveDenham]

"Did not converge" isn't the same as "Will not noverge".  We can address the first with the NLOPTIONS statement, and by adjusting the convergence criterion in the PROC GLIMMIX statement.  In order to be of assistance, though, I need to see code and output for that code, especially the iteration history

Steve Denham

[Jerem]

Hi Steve,

I included the code as well as the SAS output. Is that all you need?

I'm really grateful for your help here!

proc glimmix data=candidates;

nloptions maxiter=100;

  class commune ;

  model  binouverture (event='1') = EurostatA EurostatC sizelog ENP commune

      / solution dist=binary link=logit ddfm=kenwardrogers ; /* Addresses individual commune levels, and adds the Kenward-Rogers adjustment for repeated measures designs.  This also applies the averaged effects of the continuous covariates across all communes */

  random intercept / subject= commune solution residual; /* Converts to a repeated measures design.  Even though this is a RANDOM statement, the residual option means treat the values within a commune as correlated.  */

lsmeans commune/ilink cl; /* Outputs the estimates for each level of commune, on both the logit and original scale */

run;

Jérémy

Le Système SAS

The GLIMMIX Procedure

Data Set	WORK.CANDIDATES
Response Variable	binouverture
Response Distribution	Binary
Link Function	Logit
Variance Function	Default
Variance Matrix Blocked By	Commune
Estimation Technique	Residual PL
Degrees of Freedom Method	Kenward-Roger
Fixed Effects SE Adjustment	Prasad-Rao-Kackar-Harville-Jeske

Commune	252	Aiseau-Presles Amay Amel Andenne Anderlues Anhée Ans Anthisnes Antoing Arlon Assesse Ath Attert Aubange Aubel Awans Aywaille Baelen Bassange Bastogne Beaumont Beauraing Beauvechain Beloeil Berloz Bernissart Bertogne Bertrix Beyne-Heusay Binche Bièvre Blegny Bouillon Boussu Braine-l'Alleud Braine-le-Château Braine-le-Comte Braives Brugelette Brunehaut Burdinne Burg-Reuland Büllingen Bütgenbach Celles Cerfontaine Chapelle-Lez-Herlaimont Charleroi Chastre Chaudfontaine Chaumont-Gistoux Chimay Chiny Chièvres Châtelet Ciney Colfontaine Comblain-au-Pont Comines-Warneton Courcelles Court-St-Etienne Couvin Crisnée Dalhem Dinant Dison Doische Donceel Dour Durbuy Ecaussinnes Eghezée Ellezelles Enghien Engis Erezée Erquelinnes Esneux Estaimpuis Estinnes Etalle Eupen Faimes Farciennes Fauvillers Fernelmont Ferrières Fexhe-le-Haut-Clocher Flemalle Fleurus Flobecq Floreffe Florennes Florenville Fléron Fontaine-l'Evêque Fosses-la-Ville Frameries Frasnes-lez-Anvaing Froidchapelle Gedinne Geer Gembloux Genappe Gerpinnes Gesves Gouvy Grace-Hollogne Grez-Doiceau Habay Ham-sur-Heure-Nalinnes Hamois Hannut Hastière Havelange Hensies Herbeumont Herstal Herve Honnelles Hotton Houffalize Houyet Huy Hélecine Héron Incourt Ittre Jalhay Jemeppe-sur-Sambre Jodoigne Juprelle Jurbise Kelmis La Bruyère La Hulpe La Louvière La Roche-en-Ardenne Lasne Le Roeulx Lens Les-Bons-Villers Lessines Leuze-en-Hainaut Libramont-Chevigny Lierneux Limbourg Lincent Liège Lobbes Lontzen Malmedy Manage Manhay Marche-en-Famenne Marchin Martelange Meix-devant-Virton Merbes-le-Château Messancy Mettet Modave Momignies Mons Mont-St-Guibert Mont-de-l'Enclus Montigny-le-Tilleul Morlanwelz Mouscron Musson Namur Nandrin Nassogne Neufchâteau Neupré Nivelles Ohey Olne Onhaye Oreye Orp-Jauche Ottignies-LLN Ouffet Oupeye Paliseul Pecq Pepinster Perwez Philippeville Plombières Pont-à-Celles Profondeville Péruwelz Quaregnon Quiévrain Quévy Raeren Ramilies Rebecq Remicourt Rixensart Rochefort Rouvroy Rumes Saint Nicolas Saint-Georges-sur-Meuse Saint-Ghislain Saint-Hubert Sainte-Ode Sambreville Sankt Vith Seneffe Seraing Silly Sivry-Rance Soignies Sombreffe Somme-Leuze Soumagne Spa Sprimont Stavelot Stoumont Tellin Theux Thimister-Clermont Thuin Tinlot Tintigny Tournai Trooz Tubize Vaux-sur-Sûre Verlaine Verviers Vielsalm Villers-la-Ville Villers-le-Bouillet Viroinval Virton Visé Vresse-sur-Semois Walcourt Walhain Wanze Waremme Wasseiges Waterloo Wavre Welkenraedt Wellin Yvoir

Number of Observations Read	804
Number of Observations Used	803

1	0	476
2	1	327

R-side Cov. Parameters	1
Columns in X	257
Columns in Z per Subject	0
Subjects (Blocks in V)	252
Max Obs per Subject	14

Optimization Technique	None
Parameters	0
Lower Boundaries	0
Upper Boundaries	0
Fixed Effects	Profiled
Residual Variance	Profiled
Starting From	Data

0	0	0	1.340781E154	2.00000000	.
1	0	0	1.340781E154	1.13896683	.
2	0	0	1.340781E154	2.00000000	.
3	0	0	1.340781E154	2.00000000	.
4	0	0	1.340781E154	0.91900698	.
5	0	0	3460.7243613	2.00000000	.
6	0	0	1.340781E154	2.00000000	.
7	0	0	3728.5751363	2.00000000	.
8	0	0	1.340781E154	2.00000000	.
9	0	0	3993.8654206	2.00000000	.
10	0	0	4127.2354391	2.00000000	.
11	0	0	1.340781E154	2.00000000	.
12	0	0	4394.1129597	2.00000000	.
13	0	0	4525.2340423	2.00000000	.
14	0	0	4658.3500927	2.00000000	.
15	0	0	4791.7156	2.00000000	.
16	0	0	4924.1913727	2.00000000	.
17	0	0	5059.8010277	2.00000000	.
18	0	0	5191.3961115	2.00000000	.
19	0	0	5324.1823992	2.00000000	.
20	0	0	1.340781E154	2.00000000	.
21	0	0	1.340781E154	2.00000000	.
22	0	0	1.340781E154	2.00000000	.
23	0	0	1.340781E154	2.00000000	.
24	0	0	1.340781E154	2.00000000	.
25	0	0	1.340781E154	0.99810273	.
26	0	0	1.340781E154	0.81393176	.
27	0	0	1.340781E154	2.00000000	.
28	0	0	3148.7088811	2.00000000	.
29	0	0	3281.2624666	2.00000000	.
30	0	0	1.340781E154	2.00000000	.
31	0	0	3547.6660221	2.00000000	.
32	0	0	3677.9846413	2.00000000	.
33	0	0	3814.0718304	2.00000000	.
34	0	0	1.340781E154	2.00000000	.
35	0	0	1.340781E154	0.14868120	.
36	0	0	4210.5072254	2.00000000	.
37	0	0	4346.0003696	2.00000000	.
38	0	0	1.340781E154	2.00000000	.
39	0	0	1.340781E154	0.09425822	.
40	0	0	1.340781E154	0.09072392	.
41	0	0	4877.3197427	2.00000000	.
42	0	0	5009.6182614	2.00000000	.
43	0	0	1.340781E154	2.00000000	.
44	0	0	1.340781E154	0.06120650	.
45	0	0	5409.7912614	2.00000000	.
46	0	0	1.340781E154	2.00000000	.
47	0	0	1.340781E154	2.00000000	.
48	0	0	16935.289247	2.00000000	.
49	0	0	1.340781E154	2.00000000	.
50	0	0	1.340781E154	2.00000000	.
51	0	0	1.340781E154	0.99810273	.
52	0	0	1.340781E154	0.81393176	.
53	0	0	1.340781E154	2.00000000	.
54	0	0	3148.7088811	2.00000000	.
55	0	0	3281.2624666	2.00000000	.
56	0	0	1.340781E154	2.00000000	.
57	0	0	3547.6660221	2.00000000	.
58	0	0	3677.9846413	2.00000000	.
59	0	0	3814.0718304	2.00000000	.
60	0	0	1.340781E154	2.00000000	.
61	0	0	1.340781E154	0.14868120	.
62	0	0	4210.5072254	2.00000000	.
63	0	0	4346.0003696	2.00000000	.
64	0	0	1.340781E154	2.00000000	.
65	0	0	1.340781E154	0.09425822	.
66	0	0	1.340781E154	0.09072392	.
67	0	0	4877.3197427	2.00000000	.
68	0	0	5009.6182614	2.00000000	.
69	0	0	1.340781E154	2.00000000	.
70	0	0	1.340781E154	0.06120650	.
71	0	0	5409.7912614	2.00000000	.
72	0	0	1.340781E154	2.00000000	.
73	0	0	1.340781E154	2.00000000	.
74	0	0	16935.289247	2.00000000	.
75	0	0	1.340781E154	2.00000000	.
76	0	0	1.340781E154	2.00000000	.
77	0	0	1.340781E154	0.99810273	.
78	0	0	1.340781E154	0.81393176	.
79	0	0	1.340781E154	2.00000000	.
80	0	0	3148.7088811	2.00000000	.
81	0	0	3281.2624666	2.00000000	.
82	0	0	1.340781E154	2.00000000	.
83	0	0	3547.6660221	2.00000000	.
84	0	0	3677.9846413	2.00000000	.
85	0	0	3814.0718304	2.00000000	.
86	0	0	1.340781E154	2.00000000	.
87	0	0	1.340781E154	0.14868120	.
88	0	0	4210.5072254	2.00000000	.
89	0	0	4346.0003696	2.00000000	.
90	0	0	1.340781E154	2.00000000	.
91	0	0	1.340781E154	0.09425822	.
92	0	0	1.340781E154	0.09072392	.
93	0	0	4877.3197427	2.00000000	.
94	0	0	5009.6182614	2.00000000	.
95	0	0	1.340781E154	2.00000000	.
96	0	0	1.340781E154	0.06120650	.
97	0	0	5409.7912614	2.00000000	.
98	0	0	1.340781E154	2.00000000	.
99	0	0	1.340781E154	2.00000000	.

Did not converge.

Residual (VC)	1.0328	.

[SteveDenham]

I may have spoken too hastily.  This has all the symptoms of quasi-separation--for at least some of the communes, all of the values are either 1 or zero, and all of the continuous covariates are identical.  I think you will either have to relinquish the hope of finding marginal estimates (i.e., consider commune to be a random effect), or find a way to collapse the communes into some larger categorization that eliminates the quasi-separation.  There is a major problem in trying to estimate 258 parameters from only 803 observations.  A rule of thumb for getting decent estimates is at least 5, and preferably 10, observations per parameter.  I think you may be able to build a respectable model if you could collapse the commune variable into 70 or 80 "super-commune" levels.

Steve Denham