SPSSX Discussion

Mediation and Moderation analysis with Multilevel analysis

Classic

List

Threaded

4 messages Options

Stacey-14

Mediation and Moderation analysis with Multilevel analysis

Hi,

I wondered if someone might help. I'm testing if T7items mediates or
moderates the relationship beween TF and ERole (variables correspond with
syntax below). Given that my data are nested (139 individuals nested in 33
supervisors nested in 5 departments), I've been asked to do multilevel
analysis. Based on Tabachnick and Fidell's book, I've written the syntax
command below, which is directed at testing moderation with department as
a fixed factor at level two. I don't use syntax and so I'm not sure if
what I've written is correct (I'm guessing that it probably isn't. Also,
I'm not sure if moderation and mediation with multilevel analysis is as
simple as entering a list of variables, which in effect, is what I've
done). When I run the syntax command, I get a message that the model
failed to converge but the results are presented nevertheless. With syntax
problems/errors aside, I wasn't sure if I was getting this error message
because my data set is too small?

Any advice on the warning message or the syntax below would be greatly
appreciated. I've been trying to run this analysis for a few days and feel
little further forward than when I first started.

Thanks,

Stacey

MIXED
ERole WITH TF T7items moderatorCentred Department
/CRITERIA = CIN(95) MXITER (100) MXSTEP(5) SCORING(1)
SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
PCONVERGE(0.000001, ABSOLUTE)
/FIXED = TF T7items moderatorCentred Department | SSTYPE(3)
/METHOD = ML
/PRINT = SOLUTION TESTCOV
/RANDOM INTERCEPT TF T7items moderatorCentred | SUBJECTS
(supervisornumber) COVTYPE (UN).

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

peter link

Re: Mediation and Moderation analysis with Multilevel analysis

Stacy -

A few comments. First, the model failing to converge might have to do with
a small sample and the number of parameters that you are trying to estimate.
You may want to first, increase mxstep and mxiter. These are the max number
of step-halvings and iterations, respectively. If the model still does not
converge, consider not having so many random parameters. If the parameters
have a small amount of random variation, maybe they should not be random in
the model. There is a strong likelihood that this is the cause of the
problem.

As for mediation/moderation in multilevel models see the following,

Krull & MacKinnon (2001), Multilevel Modeling of Individual and Group Level
Mediated Effects, Multivariate Behavioral Research, 36(2), 249-277.

All the best,

Peter Link
VA San Diego Healthcare System

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
Stacey
Sent: Tuesday, November 27, 2007 5:38 AM
To: [hidden email]
Subject: Mediation and Moderation analysis with Multilevel analysis

Hi,

I wondered if someone might help. I'm testing if T7items mediates or
moderates the relationship beween TF and ERole (variables correspond with
syntax below). Given that my data are nested (139 individuals nested in 33
supervisors nested in 5 departments), I've been asked to do multilevel
analysis. Based on Tabachnick and Fidell's book, I've written the syntax
command below, which is directed at testing moderation with department as
a fixed factor at level two. I don't use syntax and so I'm not sure if
what I've written is correct (I'm guessing that it probably isn't. Also,
I'm not sure if moderation and mediation with multilevel analysis is as
simple as entering a list of variables, which in effect, is what I've
done). When I run the syntax command, I get a message that the model
failed to converge but the results are presented nevertheless. With syntax
problems/errors aside, I wasn't sure if I was getting this error message
because my data set is too small?

Any advice on the warning message or the syntax below would be greatly
appreciated. I've been trying to run this analysis for a few days and feel
little further forward than when I first started.

Thanks,

Stacey

MIXED
ERole WITH TF T7items moderatorCentred Department
/CRITERIA = CIN(95) MXITER (100) MXSTEP(5) SCORING(1)
SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
PCONVERGE(0.000001, ABSOLUTE)
/FIXED = TF T7items moderatorCentred Department | SSTYPE(3)
/METHOD = ML
/PRINT = SOLUTION TESTCOV
/RANDOM INTERCEPT TF T7items moderatorCentred | SUBJECTS
(supervisornumber) COVTYPE (UN).

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

Stacey-14

Re: Mediation and Moderation analysis with Multilevel analysis

In reply to this post by Stacey-14

Peter,

Thanks for your suggestions. I've manged to do the analysis without
convergence problems when one IV is in the model, but not when two or more
are entered. As you say, my sample size might be too small for the number
of IVs in the model, but I'm not sure how to determine this? I've come
across a reference to Mass and Hox (2005), which states that approx 50
cases should be used at level two to avoid biased estimates. Based on this,
my 33 cases would be too small to conclude anything reliable, especially
with so many IVs in the model. However, I'm not sure that this citation
alone is enough to argue that my sample is too small. However, with the
lack of convergence it looks more convincing.

Also, when I run different models (1IV/2IVs/2IV's plus moderator) I get a
warning telling me that the syntax for RANDOM effects has changed since
version 11 -- implying that I need to update my command. As you know, I
don't use syntax and so I don't know what version mine is written in or if
it needs to change? Further, for some models, I get a message telling me
that a covariance parameter is redundant and so it's not computed. This is
especially the case when I include the moderator variable with the two IVs
that it's based on. I'm not sure how to interpret this -- no unique
variance explained by the moderator, lack of available df to estimate it,
something else -- as the moderator variable is significant when it's
entered alone.

Sorry for the many questions. I hope you, or someone else can help.

Thanks,

Stacey

On Tue, 27 Nov 2007 08:55:41 -0800, peter link <[hidden email]> wrote:

>Stacy -
>
>A few comments. First, the model failing to converge might have to do with
>a small sample and the number of parameters that you are trying to
estimate.
>You may want to first, increase mxstep and mxiter. These are the max
number

>of step-halvings and iterations, respectively. If the model still does not
>converge, consider not having so many random parameters. If the parameters
>have a small amount of random variation, maybe they should not be random in
>the model. There is a strong likelihood that this is the cause of the
>problem.
>
>As for mediation/moderation in multilevel models see the following,
>
>Krull & MacKinnon (2001), Multilevel Modeling of Individual and Group Level
>Mediated Effects, Multivariate Behavioral Research, 36(2), 249-277.
>
>All the best,
>
>Peter Link
>VA San Diego Healthcare System
>
>-----Original Message-----
>From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
>Stacey
>Sent: Tuesday, November 27, 2007 5:38 AM
>To: [hidden email]
>Subject: Mediation and Moderation analysis with Multilevel analysis
>
>
>Hi,
>
>I wondered if someone might help. I'm testing if T7items mediates or
>moderates the relationship beween TF and ERole (variables correspond with
>syntax below). Given that my data are nested (139 individuals nested in 33
>supervisors nested in 5 departments), I've been asked to do multilevel
>analysis. Based on Tabachnick and Fidell's book, I've written the syntax
>command below, which is directed at testing moderation with department as
>a fixed factor at level two. I don't use syntax and so I'm not sure if
>what I've written is correct (I'm guessing that it probably isn't. Also,
>I'm not sure if moderation and mediation with multilevel analysis is as
>simple as entering a list of variables, which in effect, is what I've
>done). When I run the syntax command, I get a message that the model
>failed to converge but the results are presented nevertheless. With syntax
>problems/errors aside, I wasn't sure if I was getting this error message
>because my data set is too small?
>
>Any advice on the warning message or the syntax below would be greatly
>appreciated. I've been trying to run this analysis for a few days and feel
>little further forward than when I first started.
>
>Thanks,
>
>Stacey
>
>
>MIXED
>ERole WITH TF T7items moderatorCentred Department
>/CRITERIA = CIN(95) MXITER (100) MXSTEP(5) SCORING(1)
>SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
>PCONVERGE(0.000001, ABSOLUTE)
>/FIXED = TF T7items moderatorCentred Department | SSTYPE(3)
>/METHOD = ML
>/PRINT = SOLUTION TESTCOV
>/RANDOM INTERCEPT TF T7items moderatorCentred | SUBJECTS
>(supervisornumber) COVTYPE (UN).
>
>=====================
>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

peter link

Re: Mediation and Moderation analysis with Multilevel analysis

Hi Stacey -

Syntax should look something like this (SPSS v11.5)

MIXED
global WITH group month
/CRITERIA = CIN(95) MXITER(100) MXSTEP(5) SCORING(1)
SINGULAR(0.000000000001) HCONVERGE(0, ABSOLUTE) LCONVERGE(0, ABSOLUTE)
PCONVERGE(0.000001, ABSOLUTE)
/FIXED = group month group*month | SSTYPE(3)
/METHOD = REML
/RANDOM INTERCEPT month | SUBJECT(labid) COVTYPE(UN) .

I am suggesting that you should have fewer random parameters. On your
/RANDOM statement, you should consider removing some of the variables
listed. Should they all be random? Generally speaking, level-1 parameters
(intercept & slope(s)) may be considered random. Are all of the variables
listed on the /RANDOM statement level-1 variables? Should all of the slopes
be considered random? My guess, this is the main cause of problems for you.
The warning about covariance parameters being redundant adds evidence to
this. This warning is essentially telling you that the covariance parameter
cannot be estimated. This is most likely due to it being near the boundary
(close to zero). If the parameter in question is a covariance (not
variance) changing the covariance matrix to variance components, for
example, rather than unstructured may help. This fixes the covariance to
zero, and it therefore no longer needs to be estimated. If the parameter in
question is a variance, consider removing that variable from the /RANDOM
statement and not having that slope be random. This should help.

Peter Link
VA San Diego Healthcare System

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]]On Behalf Of
Stacey
Sent: Thursday, November 29, 2007 4:21 AM
To: [hidden email]
Subject: Re: Mediation and Moderation analysis with Multilevel analysis

Peter,

Thanks for your suggestions. I've manged to do the analysis without
convergence problems when one IV is in the model, but not when two or more
are entered. As you say, my sample size might be too small for the number
of IVs in the model, but I'm not sure how to determine this? I've come
across a reference to Mass and Hox (2005), which states that approx 50
cases should be used at level two to avoid biased estimates. Based on this,
my 33 cases would be too small to conclude anything reliable, especially
with so many IVs in the model. However, I'm not sure that this citation
alone is enough to argue that my sample is too small. However, with the
lack of convergence it looks more convincing.

Also, when I run different models (1IV/2IVs/2IV's plus moderator) I get a
warning telling me that the syntax for RANDOM effects has changed since
version 11 -- implying that I need to update my command. As you know, I
don't use syntax and so I don't know what version mine is written in or if
it needs to change? Further, for some models, I get a message telling me
that a covariance parameter is redundant and so it's not computed. This is
especially the case when I include the moderator variable with the two IVs
that it's based on. I'm not sure how to interpret this -- no unique
variance explained by the moderator, lack of available df to estimate it,
something else -- as the moderator variable is significant when it's
entered alone.

Sorry for the many questions. I hope you, or someone else can help.

Thanks,

Stacey

On Tue, 27 Nov 2007 08:55:41 -0800, peter link <[hidden email]> wrote:

>Stacy -
>
>A few comments. First, the model failing to converge might have to do with
>a small sample and the number of parameters that you are trying to
estimate.
>You may want to first, increase mxstep and mxiter. These are the max
number

Hi Peter,

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