strange behaviour of SPSS mixed with random effects of dichotomous predictors

[Christopher Stride]

Hi all

When I have a dichotomous predictor in any form of regression analysis I
almost always set the codes to 0 vs 1 and treat it as continuous (i.e. I
do the dummy coding myself)


However, when playing around with something in SPSS mixed I decided to
let SPSS do this coding for me by treating my dichotomous predictor, X,
as categorical, using the 'by' keyword, i.e. entering the code as

MIXED Y by X...

as opposed to

MIXED Y with X...

And in doing so I ran up against something very odd...


If I just look at the fixed effect of X I get the same result if I do
the dummy coding or if SPSS does it e.g.

MIXED Y with X  /fixed = X /random = intercept | subject(TEAMID) /print
= solution testcov.

gives the same interpretation as

MIXED Y by X  /fixed = X /random = intercept | subject(TEAMID) /print =
solution testcov.

This is exactly as it should be


But... if I also include X as a random effect, and allow for covarying
intercepts and slopes e.g.

MIXED Y with X  /fixed = X /random = intercept X | subject(TEAMID)
covtype(UN)  /print = solution testcov.

DOES NOT give the same results as when I use SPSS to do the dummy
coding, e.g.

MIXED Y by X  /fixed = X /random = intercept X | subject(TEAMID)
covtype(UN)  /print = solution testcov.


The former code seems OK in terms of the results given - but the latter
code, i.e the automated spss way of dummy coding X, gives a very odd
result, in that, when calculating the random effects, it seems to treat
the reference category of X as a variable too, and gives a 3x3
variance-covariance matrix, albeit with un1,3), un(2,3) and un(3,3)
unable to be estimated, and the usual warning re: the Hessian matrix!?


Has anyone else run into this issue?! Any idea what is going on?


cheers

Chris




--

--

Dr Chris Stride, C. Stat, Statistician, Institute of Work Psychology,
University of Sheffield
Telephone: [phone]
Fax: [phone]

"Figure It Out"
Statistical Consultancy and Training Service for Social Scientists

Visit www.figureitout.org.uk for details of my consultancy services, and
forthcoming training courses, which are also available on an in-house basis:

  - Data management using SPSS syntax
  - Advanced SPSS syntax and SPSS macros
  - Testing for Mediation and Moderation using SPSS
  - Multi-level Modelling using SPSS
  - Introduction to Structural Equation Modelling using Mplus
  - Testing for Mediation and Moderation using Mplus
  - Multi-level Modelling using Mplus
  - Latent Growth Curve Modelling using Mplus

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

[Ryan]

Chris,

Do you have exactly two measurements on each subject "TEAM_ID"? Is the binary predictor variable analogous to a time predictor variable in which each subject is measured on the dependent variable y at time t=0 and t=1? If I am correct so far, you get no additional information by adding a random slope term.

 Ryan

---------- Forwarded message ----------
From: Dr Chris Stride <[hidden email]>
Date: Fri, Sep 1, 2017 at 5:33 PM
Subject: strange behaviour of SPSS mixed with random effects of dichotomous predictors
To: [hidden email]


Hi all

When I have a dichotomous predictor in any form of regression analysis I almost always set the codes to 0 vs 1 and treat it as continuous (i.e. I do the dummy coding myself)


However, when playing around with something in SPSS mixed I decided to let SPSS do this coding for me by treating my dichotomous predictor, X, as categorical, using the 'by' keyword, i.e. entering the code as

MIXED Y by X...

as opposed to

MIXED Y with X...

And in doing so I ran up against something very odd...


If I just look at the fixed effect of X I get the same result if I do the dummy coding or if SPSS does it e.g.

MIXED Y with X  /fixed = X /random = intercept | subject(TEAMID) /print = solution testcov.

gives the same interpretation as

MIXED Y by X  /fixed = X /random = intercept | subject(TEAMID) /print = solution testcov.

This is exactly as it should be


But... if I also include X as a random effect, and allow for covarying intercepts and slopes e.g.

MIXED Y with X  /fixed = X /random = intercept X | subject(TEAMID) covtype(UN)  /print = solution testcov.

DOES NOT give the same results as when I use SPSS to do the dummy coding, e.g.

MIXED Y by X  /fixed = X /random = intercept X | subject(TEAMID) covtype(UN)  /print = solution testcov.


The former code seems OK in terms of the results given - but the latter code, i.e the automated spss way of dummy coding X, gives a very odd result, in that, when calculating the random effects, it seems to treat the reference category of X as a variable too, and gives a 3x3 variance-covariance matrix, albeit with un1,3), un(2,3) and un(3,3) unable to be estimated, and the usual warning re: the Hessian matrix!?


Has anyone else run into this issue?! Any idea what is going on?


cheers

Chris




--

--

Dr Chris Stride, C. Stat, Statistician, Institute of Work Psychology,
University of Sheffield
Telephone: [phone]
Fax: [phone]

"Figure It Out"
Statistical Consultancy and Training Service for Social Scientists

Visit www.figureitout.org.uk for details of my consultancy services, and
forthcoming training courses, which are also available on an in-house basis:

 - Data management using SPSS syntax
 - Advanced SPSS syntax and SPSS macros
 - Testing for Mediation and Moderation using SPSS
 - Multi-level Modelling using SPSS
 - Introduction to Structural Equation Modelling using Mplus
 - Testing for Mediation and Moderation using Mplus
 - Multi-level Modelling using Mplus
 - Latent Growth Curve Modelling using Mplus

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

[Christopher Stride]

No... I have > 20 people in each team, some of whom will s core 1 on X, and others who will score 2 on X.

So fitting random effect of X is not an issue as far as I'm aware... and works fine, ie intercept, slope variances and intercept-slope covariance are estimated if I recode X to 0 vs 1 and treat X as a continuous var, ie place it after WITH.

---

From: [hidden email]
Sent: ‎04/‎09/‎2017 00:00
To: [hidden email]
Subject: Fwd: strange behaviour of SPSS mixed with random effects ofdichotomous predictors

Chris,

Do you have exactly two measurements on each subject "TEAM_ID"? Is the binary predictor variable analogous to a time predictor variable in which each subject is measured on the dependent variable y at time t=0 and t=1? If I am correct so far, you get no additional information by adding a random slope term.

 Ryan

---------- Forwarded message ----------
From: Dr Chris Stride <[hidden email]>
Date: Fri, Sep 1, 2017 at 5:33 PM
Subject: strange behaviour of SPSS mixed with random effects of dichotomous predictors
To: [hidden email]


Hi all

When I have a dichotomous predictor in any form of regression analysis I almost always set the codes to 0 vs 1 and treat it as continuous (i.e. I do the dummy coding myself)


However, when playing around with something in SPSS mixed I decided to let SPSS do this coding for me by treating my dichotomous predictor, X, as categorical, using the 'by' keyword, i.e. entering the code as

MIXED Y by X...

as opposed to

MIXED Y with X...

And in doing so I ran up against something very odd...


If I just look at the fixed effect of X I get the same result if I do the dummy coding or if SPSS does it e.g.

MIXED Y with X  /fixed = X /random = intercept | subject(TEAMID) /print = solution testcov.

gives the same interpretation as

MIXED Y by X  /fixed = X /random = intercept | subject(TEAMID) /print = solution testcov.

This is exactly as it should be


But... if I also include X as a random effect, and allow for covarying intercepts and slopes e.g.

MIXED Y with X  /fixed = X /random = intercept X | subject(TEAMID) covtype(UN)  /print = solution testcov.

DOES NOT give the same results as when I use SPSS to do the dummy coding, e.g.

MIXED Y by X  /fixed = X /random = intercept X | subject(TEAMID) covtype(UN)  /print = solution testcov.


The former code seems OK in terms of the results given - but the latter code, i.e the automated spss way of dummy coding X, gives a very odd result, in that, when calculating the random effects, it seems to treat the reference category of X as a variable too, and gives a 3x3 variance-covariance matrix, albeit with un1,3), un(2,3) and un(3,3) unable to be estimated, and the usual warning re: the Hessian matrix!?


Has anyone else run into this issue?! Any idea what is going on?


cheers

Chris




--

--

Dr Chris Stride, C. Stat, Statistician, Institute of Work Psychology,
University of Sheffield
Telephone: [phone]
Fax: [phone]

"Figure It Out"
Statistical Consultancy and Training Service for Social Scientists

Visit www.figureitout.org.uk for details of my consultancy services, and
forthcoming training courses, which are also available on an in-house basis:

 - Data management using SPSS syntax
 - Advanced SPSS syntax and SPSS macros
 - Testing for Mediation and Moderation using SPSS
 - Multi-level Modelling using SPSS
 - Introduction to Structural Equation Modelling using Mplus
 - Testing for Mediation and Moderation using Mplus
 - Multi-level Modelling using Mplus
 - Latent Growth Curve Modelling using Mplus

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