Marginal model in R and SPSS

[Emperor3]

Hi all,

I have 2 datasets which i analyze both with R and SPSS (for professional
reasons). Specifically, i use a marginal model with an AR(1) covariance
structure. I have 2 covariates, namely the main effect of time ( days ) and
the interaction between time and treatment (days_trtm).

In R, i do that with nlme::gls and the results are :

Generalized least squares fit by REML
Model: log_volume ~ days + days_trtm

Correlation Structure: Continuous AR(1)
Formula: ~days | ID
Parameter estimate(s):
   Phi
0.8602442

Coefficients:
                      Value   Std.Error     t-value       p-value
(Intercept)  3.730806 0.23320201 15.998174   0e+00
days           0.225310 0.02611839  8.626483   0e+00
days_trtm   -0.098717 0.02542601 -3.882526   2e-04


Residual standard error: 0.8417871
Degrees of freedom: 109 total; 106 residual



In SPSS i use the mixed command and the (somewhat cleaned) results are :



Fixed Effects


Estimates of Fixed Effects
Parameter   Estimate     Std. Error     df          t      Sig.    

Intercept   3,741469     ,229005    24,173  16,338  ,000    
days         ,226131      ,025627     39,376    8,824  ,000    
days_trtm   -,101737    ,024903    26,396  -4,085  ,000    -



Covariance Parameters



Estimates of Covariance Parameters
                                 Parameter    Estimate    Std. Error
Repeated Measures     AR1 diagonal    ,691944  ,156614
                               AR1 rho         ,698854  ,070333
 


As you can see, i get different results, and especially for the correalation
parameter (0.86 vs 0.69)... My gut says that they are parametarized in a
different way, but i couldn't find the answer..

If someone can help me with that, i would be glad!

Thanks!



--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
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

[Maguin, Eugene]

Others, I think, can comment on differences between R and SPSS-Mixed in the estimation routines, but REML was used on the R solution; was REML used on the mixed solution? Not to minimize the differences but they seem pretty small, I think, except for days_trtm: 0.29% on the intercept, 0.36% on days, and 3.06% on days_trtm. By the way, what is the phi value in the R solution?
Gene Maguin


-----Original Message-----
From: SPSSX(r) Discussion <[hidden email]> On Behalf Of Emperor3
Sent: Monday, August 27, 2018 10:06 AM
To: [hidden email]
Subject: Marginal model in R and SPSS

Hi all,

I have 2 datasets which i analyze both with R and SPSS (for professional reasons). Specifically, i use a marginal model with an AR(1) covariance structure. I have 2 covariates, namely the main effect of time ( days ) and the interaction between time and treatment (days_trtm).

In R, i do that with nlme::gls and the results are :

Generalized least squares fit by REML
Model: log_volume ~ days + days_trtm

Correlation Structure: Continuous AR(1)
Formula: ~days | ID
Parameter estimate(s):
   Phi
0.8602442

Coefficients:
                      Value   Std.Error     t-value       p-value
(Intercept)  3.730806 0.23320201 15.998174   0e+00
days           0.225310 0.02611839  8.626483   0e+00
days_trtm   -0.098717 0.02542601 -3.882526   2e-04


Residual standard error: 0.8417871
Degrees of freedom: 109 total; 106 residual



In SPSS i use the mixed command and the (somewhat cleaned) results are :



Fixed Effects


Estimates of Fixed Effects
Parameter   Estimate     Std. Error     df          t      Sig.    

Intercept   3,741469     ,229005    24,173  16,338  ,000    
days         ,226131      ,025627     39,376    8,824  ,000    
days_trtm   -,101737    ,024903    26,396  -4,085  ,000    -



Covariance Parameters



Estimates of Covariance Parameters
                                 Parameter    Estimate    Std. Error
Repeated Measures     AR1 diagonal    ,691944  ,156614
                               AR1 rho         ,698854  ,070333
 


As you can see, i get different results, and especially for the correalation
parameter (0.86 vs 0.69)... My gut says that they are parametarized in a
different way, but i couldn't find the answer..

If someone can help me with that, i would be glad!

Thanks!



--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
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

=====================
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

[Emperor3]

Thanks for your answer!

I used REML also in SPSS! Actually that was also my first checking!

Yes, i am not that concerned about the differences in the predictors, but
mostly about the correlation parameter. Even the residual variation is
identical (In R the standard deviation is reported instead of the variance
in SPSS).

From what i know, Phi denotes the correlation between 2 adjacent
measurements, and i assumed that the same is true also for SPSS...

Best,
John



--
Sent from: http://spssx-discussion.1045642.n5.nabble.com/

=====================
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