Discussion: View Thread

I've got a client who wants to do an error-in-measurement model where multiple groups are tested against a standard and he wants a common slope/different intercepts solution.  The request is specific for a SAS solution, so I'm working with proc Calis.  I've built a couple of models, but there are issues with convergence, variance estimates and degrees of freedom, just to name a few.  If anyone has done something like this and would be willing to share the benefit of their experience, I would be appreciative.

Thanks,

-------------------------------------------
Robert Gallavan
-------------------------------------------

I am not an expert on latent variable models nor have I ever used Calis, but I am an experienced statistician with a lot of consulting experience.  I think the key words are "he wants a common slope/different intercepts solution."  Before jumping into fitting models a good consultant will ask the client why he believes each group should have the same model.  If you get a satisfactory answer to that the next thing to do is plot the data and see if you believe that the data is compatible with an equal slopes model or you formally test for equality of slopes.  My guess is that numerical problems with convergence is simply due to the data not fitting a restricted model.
-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------






Original Message:
Sent: 08-26-2011 05:58
From: Robert Gallavan
Subject: Indicator Variables in Latent Variable Modeling

I've got a client who wants to do an error-in-measurement model where multiple groups are tested against a standard and he wants a common slope/different intercepts solution.  The request is specific for a SAS solution, so I'm working with proc Calis.  I've built a couple of models, but there are issues with convergence, variance estimates and degrees of freedom, just to name a few.  If anyone has done something like this and would be willing to share the benefit of their experience, I would be appreciative.

Thanks,

-------------------------------------------
Robert Gallavan
-------------------------------------------

Unfortunately, that ship has already sailed.  My client is also a statistician and he wants what he wants.  Before I fight that battle again, I would like to be sure that my inability to obtain a decent model is due to the fact that it is not appropriate, rather than my inexperience with latent models.

-------------------------------------------
Robert Gallavan
-------------------------------------------






Original Message:
Sent: 08-26-2011 06:58
From: Michael Chernick
Subject: Indicator Variables in Latent Variable Modeling


I am not an expert on latent variable models nor have I ever used Calis, but I am an experienced statistician with a lot of consulting experience.  I think the key words are "he wants a common slope/different intercepts solution."  Before jumping into fitting models a good consultant will ask the client why he believes each group should have the same model.  If you get a satisfactory answer to that the next thing to do is plot the data and see if you believe that the data is compatible with an equal slopes model or you formally test for equality of slopes.  My guess is that numerical problems with convergence is simply due to the data not fitting a restricted model.
-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------

It doesn't matter whether or not the client is a statistician the first thing you should always do is check the assumptions.  It also doesn't matter that you already tried fitting the models.  You can still step back and look at the data to see if teh equal slopes model is appropriate.  If it turns out that it appears to be okay then you can start looking at the details of what the software does.  Doesn't that make sense?  It is never too late to tell the client that he is wrong.
-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------






Original Message:
Sent: 08-26-2011 07:08
From: Robert Gallavan
Subject: Indicator Variables in Latent Variable Modeling

Unfortunately, that ship has already sailed.  My client is also a statistician and he wants what he wants.  Before I fight that battle again, I would like to be sure that my inability to obtain a decent model is due to the fact that it is not appropriate, rather than my inexperience with latent models.

-------------------------------------------
Robert Gallavan
-------------------------------------------

I certainly agree with Michael, It's not the model you want, but the model supported by the data.
  I also wonder if you are using the right method in SAS.
This seems like a natural for proc mixed with a growth curve = random slope + random intercept plus potential fixed group effects.  My experience with proc callis is that it is one of the older SAS procedures and
may not have equivalent quality algorithms.  Although it does seem that did quite a rewrite with SAS 9.3.

Those of of that saw the progression of "different results" with the development of proc mixed, realize how
sensitive some of these methods are to both the data and the algorithms.

-------------------------------------------
Raymond Hoffmann
Professor
Medical College of Wisconsin
-------------------------------------------






Original Message:
Sent: 08-26-2011 07:41
From: Michael Chernick
Subject: Indicator Variables in Latent Variable Modeling


It doesn't matter whether or not the client is a statistician the first thing you should always do is check the assumptions.  It also doesn't matter that you already tried fitting the models.  You can still step back and look at the data to see if teh equal slopes model is appropriate.  If it turns out that it appears to be okay then you can start looking at the details of what the software does.  Doesn't that make sense?  It is never too late to tell the client that he is wrong.
-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------

You could start with something real simple like a plot and an ANCOVA in GLM to test for equal slopes.  If the slopes are clearly unequal or if you see some other pathology then forcing the data as it exists into the clients model is futile.

The views expressed on this Web site are mine alone and do not necessarily reflect the views of my employer.
-------------------------------------------
Emil Friedman
MannKind Corporation
-------------------------------------------






Original Message:
Sent: 08-26-2011 08:10
From: Raymond Hoffmann
Subject: Indicator Variables in Latent Variable Modeling

I certainly agree with Michael, It's not the model you want, but the model supported by the data.
  I also wonder if you are using the right method in SAS.
This seems like a natural for proc mixed with a growth curve = random slope + random intercept plus potential fixed group effects.  My experience with proc callis is that it is one of the older SAS procedures and
may not have equivalent quality algorithms.  Although it does seem that did quite a rewrite with SAS 9.3.

Those of of that saw the progression of "different results" with the development of proc mixed, realize how
sensitive some of these methods are to both the data and the algorithms.

-------------------------------------------
Raymond Hoffmann
Professor
Medical College of Wisconsin
-------------------------------------------






Original Message:
Sent: 08-26-2011 07:41
From: Michael Chernick
Subject: Indicator Variables in Latent Variable Modeling


It doesn't matter whether or not the client is a statistician the first thing you should always do is check the assumptions.  It also doesn't matter that you already tried fitting the models.  You can still step back and look at the data to see if teh equal slopes model is appropriate.  If it turns out that it appears to be okay then you can start looking at the details of what the software does.  Doesn't that make sense?  It is never too late to tell the client that he is wrong.
-------------------------------------------
Michael Chernick
Director of Biostatistical Services
Lankenau Institute for Medical Research
-------------------------------------------