Undertow Tracker
https://communities.sas.com/kntur85557/tracker
Undertow TrackerThu, 12 Sep 2024 06:52:42 GMT2024-09-12T06:52:42ZRe: Hierarchical logistic regression: choice of a model and convergence problem (proc glimmix)
https://communities.sas.com/t5/Statistical-Procedures/Hierarchical-logistic-regression-choice-of-a-model-and/m-p/358051#M18816
<P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor">Dear <A href="https://communities.sas.com/t5/user/viewprofilepage/user-id/130720" target="_self"><SPAN class="">lkeyes, </SPAN></A></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Many thanks for your detailed answer, which was very helpful, even though I am disappointed by the results.<BR /></SPAN></SPAN></P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">When I run the 2-level unconditional model (without predictors), I have the following covariance parameter estimates: intercept estimate 0.12, StdErr 0.13. Values for the estimate and StdErr of each center's intercept seem to be correctly displayed In the "solution for random effects" table. DF=409.</SPAN></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Thanks to the addition of the "parms /ParmsData= Parms;" statement, I am now able to obtain convergence of the algorithm for the 2-level model containing only a random intercept and the continuous thrombus_length variable (fixed slope). However, in the "solution for random effects", the estimate and StdErr of each center's intercept are now equal to zero. </SPAN></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">I also realized that in the 3-level model described in my earlier post, the estimate and StdErr of each center's intercept (random effects) are also equal to zero. </SPAN></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Every time the thrombus_length variable is entered in a hierarchical logistic model, this problem occurs. Rescaling the variable [0 , 1] or making it categorical (tertiles) did not change the situation. This is the only variable leading to this problem. Unfortunately, it is the most important predictor and therefore it cannot be simply ruled out of any model.<BR /></SPAN></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Could you please confirm that both models are invalid ? Or equivalent to a simple (non-hierarchical) logistic regression ? Is this due to the fact that the number of events (recanalisation=yes) per center is too small ?<BR /></SPAN></SPAN></P><P> </P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Thank you for your response.</SPAN></SPAN></P><P><SPAN class="UserName lia-user-name lia-user-rank-Occasional-Contributor"><SPAN class="">Guillaume</SPAN></SPAN></P><P> </P>Thu, 11 May 2017 21:22:01 GMThttps://communities.sas.com/t5/Statistical-Procedures/Hierarchical-logistic-regression-choice-of-a-model-and/m-p/358051#M18816Undertow2017-05-11T21:22:01ZHierarchical logistic regression: choice of a model and convergence problem (proc glimmix)
https://communities.sas.com/t5/Statistical-Procedures/Hierarchical-logistic-regression-choice-of-a-model-and/m-p/357668#M18782
<P>Good evening everyone,</P><P> </P><P>I am new to hierarchical logistic regression and would like to have your help on a practical problem.</P><P>I need to analyse potential predictors of early recanalisation (dependent variable, binary) in an observational study of 352 patients with acute stroke. Patients had brain imaging and thrombolysis therapy at the nearest stroke center (nearest_center, n=15) and all patients were subsequently transferred to a comprehensive stroke center (comprehensive_center, n=4) to receive endovascular therapy. One important point: sometimes the nearest center was a comprehensive stroke center and secondary transfer was obviously not needed.</P><P> </P><P>I consider using a mixed model to take into account potential heterogeneity across centers. </P><P>When I run a three-level hierarchical model with proc glimmix (level-1 = patient; level-2= nearest center; level-3 = comprehensive center), I have no convergence problem. Here is the code that I used (random intercept model, fixed slopes). Thrombus_length is a continuous variable and fvh is a binary variable:</P><P> </P><P> </P><PRE><CODE class=" language-sas">proc glimmix method=laplace ;
class comprehensive_center nearest_center fvh (ref='0') ;
model recanalisation_complete (event='1') = thrombus_length fvh / dist=binary link=logit solution oddsratio ;
random intercept / subject=comprehensive_center ;
random intercept / subject=nearest_center(comprehensive_center) ;
run; </CODE></PRE><P> </P><P> My first question is: is it correct to use such a model since for some patients the comprehensive center and nearest center are the same ?</P><P> </P><P>I also considered using a simpler model, with two hierarchical levels (level-1=patient; level-2= nearest center):</P><P> </P><PRE><CODE class=" language-sas">proc glimmix method=laplace ;
class nearest_center fvh (ref='0') ;
model recanalisation_complete (event='1') = thrombus_length fvh / dist=binary link=logit solution oddsratio ;
random intercept / subject=nearest_center ;
run; </CODE></PRE><P>However, in this situation I get the following message: "Warning: obtaining minimum var. quadratic unbiased estimates as starting values for the covariance parameters failed." I tried<SPAN style="font-family: arial, helvetica, sans-serif;"> many options proc glimmix offers for controlling the optimization process (<A title="http://support.sas.com/resources/papers/proceedings12/332-2012.pdf" href="http://support.sas.com/resources/papers/proceedings12/332-2012.pdf" target="_blank" rel="nofollow noopener noreferrer">http://support.sas.com/resources/papers/proceedings12/332-2012.pdf</A>), without success. The culprit seems to be the thrombus_length variable, because I get the above error message only when this variable is in the model (even if it is the only independant variable entered).<BR /></SPAN></P><P> </P><P> My second question is: could this error message reflect that the number of patients (or the number of patients who recanalize) in each nearest_center is too small to reasonably allow the use of a hierarchical logistic model ? Is there a rule of thumb regarding this point ? The number of patients in each nearest_center ranges from 6 to 44, and the number of patients who recanalize ranges from 0 to 11 (total number of patients who recanalize: 61).</P><P> </P><P>Many thanks for your help.</P><P>Guillaume</P>Wed, 10 May 2017 20:35:53 GMThttps://communities.sas.com/t5/Statistical-Procedures/Hierarchical-logistic-regression-choice-of-a-model-and/m-p/357668#M18782Undertow2017-05-10T20:35:53ZRe: In proc logistic, how to display the results of the score test for EACH variable ?
https://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308362#M16314
Thank you so much for this elegant solution !Mon, 31 Oct 2016 19:36:59 GMThttps://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308362#M16314Undertow2016-10-31T19:36:59ZRe: In proc logistic, how to display the results of the score test for EACH variable ?
https://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308243#M16311
<P>Thank you for your response. This may be true. However, I doubt it because when you operate a forward (or stepwise) variable selection using proc logistic, the result of the score test for each candidate variable is provided.</P><P>However, in this case, the results of each score test seems to correspond to a comparison of the current model vs. the current model + one variable.</P><P> </P><P>To make it clearer, let's take an example:</P><P>When i run the following SAS code :</P><P>" proc logistic desc; model end4_unexplained = age antiplatelet nihss0h<BR />/ selection=forward slstay=0.10 details; run; "</P><P> </P><P>At step 1, the variable "nihss0h" is entered in the model.</P><P>A the end of step 1, a list of variables eligible for entry is provided, with a Score Chi-Square and corresponding P-value for each candidate variable (please see the image associated to this message). I believe that the P-value for the variable "age" corresponds to a comparison, using the Score Test, of the model with the variables "age" and "nihss0h", versus the model containing only "nihss0h".</P><P>Is this correct ?</P><P> </P><P>If so, back to the main question of this topic (displaying the results of a Score test for each variable using proc logistic): do you think that such a test can not apply to this situation, or that it is simply not provided by proc logistic ?</P><P> </P><BR /><IMG src="https://communities.sas.com/t5/image/serverpage/image-id/13093i6F9F7D6D71799788/image-size/large?v=1.0&px=600" border="0" alt="1.jpg" title="1.jpg" />Mon, 31 Oct 2016 11:33:35 GMThttps://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308243#M16311Undertow2016-10-31T11:33:35ZRe: In proc logistic, how to display the results of the score test for EACH variable ?
https://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308234#M16309
<P>This is not an acronym. I just wanted to underline that I would like to have the result of the test score associated with each variable of the multivariable model, not just the result of the score test for the global null hypothesis (All Beta=0), which is already provided by proc logistic.</P>Mon, 31 Oct 2016 10:23:41 GMThttps://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308234#M16309Undertow2016-10-31T10:23:41ZIn proc logistic, how to display the results of the score test for EACH variable ?
https://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308204#M16307
<P>When using proc logistic for a multivariable binary logistic regression, the results of the Wald Chi-Square and corresponding P-value are displayed for each variable entered in the model ("Analysis of Maximum Likelihood Estimates" section).</P><P> </P><P>Is it possible to display the Score Chi-Square and corresponding P-value for EACH variable instead ? If so, how ?</P><P>Thank you for your help.</P>Mon, 31 Oct 2016 08:13:36 GMThttps://communities.sas.com/t5/Statistical-Procedures/In-proc-logistic-how-to-display-the-results-of-the-score-test/m-p/308204#M16307Undertow2016-10-31T08:13:36Z