Why so? Curious
Locked
 
|
This question came up yesterday because I needed the covariance matrix of parameter estimates for a logistic regression. I know that for a (multiple) regression, the Statistics subcommand provides the covariance matrix (keyword: bcov).
However, the Print subcommand for logistic regression provides the correlation matrix (keyword: corr). So, there could be a specific statistical reason that I don’t know anything about for that difference—if so, what is it? On the other hand, it just could
have been a design decision--so there! |
|
|
Well, binary logistic allows you to save the covariance metrics in the model XML file but only the correlation matrix as a Viewer table. With NOMREG, you can have either or both. The correlation matrix is more useful for eyeballing, but the covariance matrix is needed for various post-calculations. :-) On Thu, May 25, 2017 at 10:00 AM, Maguin, Eugene <[hidden email]> wrote:
|
|
|
In reply to this post by Maguin, Eugene
I have no idea why SPSS does not allow one to
print the
covariance matrix of parameter estimates
(ditto for why
Multivariate ANOVA results are always provided
for
repeated measures ANOVA as well as polynomial
contrasts
regardless if they make sense or not) but
since the SPSS
syntax manual for logistic regression (Ver 23)
cites
Hosmer and Lemeshow (2000; Applied Logistic
Regression, 2nd ed)
as the source for the Hosmer-Lemeshow goodness
of fit
test (pp140-145), it is odd that H&L's
presentation of the
covariance matrix of parameter estimates
(on pp104-105)
is ignored.
I have not used Nomreg but it allows one to
obtain the
asymptotic covariance matrix (see p1234 in the
V23 manual)
and I guess one might be able to Nomreg
in order to get
the cova matrix. People more familiar
with how Nomreg
compares to the logistic
regression procedure.
More generally, it is curious why certain
types of output
are provided or not and who makes the decision
for this.
Perhaps it is a programming decision but it
might be an
"executive" decision (i.e., someone in a
higher pay grade
not involved in programming) who makes that
decision.
On an unrelated sidenote, when one does a
repeated measures
ANOVA with GLM and uses multiple comparisons,
there
only three options available (i.e., LSD test,
Bonnferroni corrected
t-test, and Sidak modification of the
Bonferroni correction),
why don't they provide that actual value of
the LSD or other
minimum differences? I use the Gravetter
& Wallnau statistics
test in psych stats and they don't even cover
these tests, so
I show how the hand calculations can be done
(though there
is a debate about whether one should use the
mean square
from the ANOVA table if one is comparing 3 or
more means,
or use the variance errors for the two groups
being compared --
SPSS does not indicate which one it
uses). Who decides
what the output should contain?
-Mike Palij
New York University
===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
|
|
These are not executive or programmer decisions. They are made by statisticians and designers, subject to design review and resource constraints. LOGISTIC is very old, and apparently demand for the covariance matrix has been minimal. I don't recall it ever appearing on a request list, but it could be requested through the suggestion process. Or you could parse it out of the model xml :-) On Thu, May 25, 2017 at 11:22 AM Mike Palij <[hidden email]> wrote:
|
|
|
On Thursday, May 25, 2017 1:31 PM, Jon Peck wrote:
>These are not executive or programmer decisions. Could have fooled me. ;-) >They are made by statisticians and designers, subject to design >review and resource constraints. LOGISTIC is very old, and >apparently demand for the covariance matrix has been minimal. I quote from the table caption for Hosmer & Lemeshow (2000) Table 4.14: Estiates of the Variance and Covariances for the Estimated Parameters Used in the Calculation of the Logit Differences Use to Compute the Odds Ratio for LWD and Smoking Adjusting for AGE [Note: these are variables used in the example being covered] On the prior 2 pages, L&M show how various calculations can be done by hand, most importantly the confidence intervals of various differences. I wonder why a statistician or designer would leave out the covariance matrix given it's role in these calculations. I guess this is just one reason why one would prefer to use BMDP's (RIP) LR regression program which was available back in the late 1980s/early 1990s (depending upon which version of BMDP one had access). >I don't recall it ever appearing on a request list, but it could >be requested through the suggestion process. Then again, it is not always clear what the "suggestion process" is and why suggestions made to SPSS people on SPSSX-L (back in the day, Dave Nichols), wasn't sufficient to get the process started. Still wainting to get a "sorted" graphic representation of a correlation table -- BMDP's cluster analysis of variables prog 1M provided this via /Print Shade. "Shade" was also available for cluster analysis of cases (2M) and factor analysis (4M), using different criterion for ordering the correlations. Boy, do I miss BMDP. :-( >Or you could parse it out of the model xml :-) Yeah, like I don't have anything better to do. ;-) Perhaps I'd be better off trying to get my copy of BMDPPC to run is a DOS box on Windows. -Mike Palij New York University [hidden email] P.S. Don't think I didn't notice your avoidance of my comments about GLM. ;-) On Thu, May 25, 2017 at 11:22 AM Mike Palij <[hidden email]> wrote: I have no idea why SPSS does not allow one to print the covariance matrix of parameter estimates (ditto for why Multivariate ANOVA results are always provided for repeated measures ANOVA as well as polynomial contrasts regardless if they make sense or not) but since the SPSS syntax manual for logistic regression (Ver 23) cites Hosmer and Lemeshow (2000; Applied Logistic Regression, 2nd ed) as the source for the Hosmer-Lemeshow goodness of fit test (pp140-145), it is odd that H&L's presentation of the covariance matrix of parameter estimates (on pp104-105) is ignored. I have not used Nomreg but it allows one to obtain the asymptotic covariance matrix (see p1234 in the V23 manual) and I guess one might be able to Nomreg in order to get the cova matrix. People more familiar with how Nomreg compares to the logistic regression procedure. More generally, it is curious why certain types of output are provided or not and who makes the decision for this. Perhaps it is a programming decision but it might be an "executive" decision (i.e., someone in a higher pay grade not involved in programming) who makes that decision. On an unrelated sidenote, when one does a repeated measures ANOVA with GLM and uses multiple comparisons, there only three options available (i.e., LSD test, Bonnferroni corrected t-test, and Sidak modification of the Bonferroni correction), why don't they provide that actual value of the LSD or other minimum differences? I use the Gravetter & Wallnau statistics test in psych stats and they don't even cover these tests, so I show how the hand calculations can be done (though there is a debate about whether one should use the mean square from the ANOVA table if one is comparing 3 or more means, or use the variance errors for the two groups being compared -- SPSS does not indicate which one it uses). Who decides what the output should contain? ----- Original Message ----- From: Maguin, Eugene Sent: Thursday, May 25, 2017 12:00 PM Subject: Why so? Curious This question came up yesterday because I needed the covariance matrix of parameter estimates for a logistic regression. I know that for a (multiple) regression, the Statistics subcommand provides the covariance matrix (keyword: bcov). However, the Print subcommand for logistic regression provides the correlation matrix (keyword: corr). So, there could be a specific statistical reason that I don’t know anything about for that difference—if so, what is it? On the other hand, it just could have been a design decision--so there! Gene Maguin ===================== To manage your subscription to SPSSX-L, send a message to [hidden email] (not to SPSSX-L), with no body text except the command. To leave the list, send the command SIGNOFF SPSSX-L For a list of commands to manage subscriptions, send the command INFO REFCARD |
Administrator
|
In reply to this post by Jon Peck
You can also use GENLIN (with DISTRIBUTION=BINOMIAL and LINK=LOGIT) to estimate a binary logistic regression model. Both the PRINT and OUTFILE sub-commands have COVB options.
--
Bruce Weaver bweaver@lakeheadu.ca http://sites.google.com/a/lakeheadu.ca/bweaver/ "When all else fails, RTFM." PLEASE NOTE THE FOLLOWING: 1. My Hotmail account is not monitored regularly. To send me an e-mail, please use the address shown above. 2. The SPSSX Discussion forum on Nabble is no longer linked to the SPSSX-L listserv administered by UGA (https://listserv.uga.edu/). |
«
Return to SPSSX Discussion
|
1 view|%1 views
| Free forum by Nabble | Edit this page |