Dummy Coding & Screwy Parameter Estimates

I'm using generalized estimating equations to test the relationship between
two dichotomous independent variables with a dichotomous dependent variable.
I would like to use the parameter estimates B to report an odds ratio (which,
if I recall, should be euler's number ^ B).

Problem: When I enter my IVs as predictors and don't include an interaction,
the significance tests and chi-square statistics in the "parameter estimates"
match the "tests of model effects."  However, when I INCLUDE an interaction
(which I need to do), the parameter estimates stuff does not match the tests
of model effects.  I don't really know why this is, but I know it has
something to do with dummy coding, or lack thereof.

To get accurate parameter estimates for the interaction and main effects, do I
need to dummy code?  If so, is this what I would do?  Right now, both
variables are coded 0, 1.  I'm adding in the interaction on the GEE "model
tab."

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

[Bruce Weaver]

Simon - [e-mail] wrote

I'm using generalized estimating equations to test the relationship between
two dichotomous independent variables with a dichotomous dependent variable.
I would like to use the parameter estimates B to report an odds ratio (which,
if I recall, should be euler's number ^ B).

Problem: When I enter my IVs as predictors and don't include an interaction,
the significance tests and chi-square statistics in the "parameter estimates"
match the "tests of model effects."  However, when I INCLUDE an interaction
(which I need to do), the parameter estimates stuff does not match the tests
of model effects.  I don't really know why this is, but I know it has
something to do with dummy coding, or lack thereof.

To get accurate parameter estimates for the interaction and main effects, do I
need to dummy code?  If so, is this what I would do?  Right now, both
variables are coded 0, 1.  I'm adding in the interaction on the GEE "model
tab."

I suspect a lot of us are not sure what your output looks like.  Could you post some sample data and your GEE code?

Sure.  So my data looks like this:

DYADID          DYAD#        IV1    IV2   DV
1                 1           1      1    0
1                 2           0      1    1
2                 1           1      0    0
2                 2           1      0    0
3                 1           0      1    1
3                 2           0      1    1
4                 1           0      0    0
4                 2           1      0    0

The IVs and DV are all categorical variables, coded 0, 1.

So, in Scenario 1, I leave out the interaction, and I get:

 Tests of Model Effects
  Type III
Source      Wald Chi-Square df Sig.
(Intercept) 148.218 1 .000
IV1           7.333 1 .007
IV2       .175 1 .676
Dependent Variable: DV
Model: (Intercept), IV1, IV2

   Parameter Estimates
   95% Wald Confidence Interval
Hypothesis Test
Parameter B   Std. Error Lower  Upper Wald Chi-Square df Sig.
(Intercept) 1.814 .3431 1.141 2.486 27.951 1 .000
[IV1=0]   1.052 .3885 .291 1.814 7.333 1 .007
[IV1=1]  0a . . . . . .
[IV2=0]  .164 .3929 -.606 .934 .175 1 .676
[IV2=1]  0a . . . . . .
(Scale)           1
Dependent Variable: DV
Model: (Intercept), IV1, IV2
a. Set to zero because this parameter is redundant.

So I realize this is messy, but you'll notice the Wald-chi square and the
significance levels = the output in test of model effects.  In Scenario2, I
add an IV1 x IV2 interaction term.

 Tests of Model Effects
  Type III
Source Wald Chi-Square df Sig.
(Intercept) 145.361 1 .000
IV1 7.132 1 .008
IV2 .170 1 .680
IV1*IV2 .001 1 .976
Dependent Variable: DV
Model: (Intercept), IV1, IV2, IV1 * IV2

   Parameter Estimates
   95% Wald Confidence Interval
Hypothesis Test
Parameter B  Std. Error Lower Upper Wald Chi-Square df Sig.
(Intercept) 1.808 .3849 1.054 2.563 22.077 1 .000
[IV1=0]  1.063 .4944 .094 2.032 4.627 1 .031
[IV1=1]  0a . . . . . .
[IV2=0]         .178 .6285 -1.054 1.410 .080 1 .777
[IV2=1]  0a . . . . . .
[IV1=0] * [IV2=0] -.024 .7874 -1.567 1.519 .001 1 .976
[IV1=0] * [IV2=1] 0a . . . . . .
[IV1=1] * [IV2=0] 0a . . . . . .
[IV1=1] * [IV2=1] 0a . . . . . .
(Scale) 1
Dependent Variable: DV
Model: (Intercept), IV1, IV2, IV1 * IV2
a. Set to zero because this parameter is redundant.

So in this example, the Wald Ch-Square and significance level for the
parameter estimates are not equal to the test of type III model effects.

I had this problem once before in LMM.  It's not a GEE-specific issue, at
least I don't think.





On Wed, 15 Sep 2010 14:41:21 -0700, Bruce Weaver <[hidden email]>
wrote:

>Simon - [hidden email] wrote:
>>
>> I'm using generalized estimating equations to test the relationship
>> between
>> two dichotomous independent variables with a dichotomous dependent
>> variable.
>> I would like to use the parameter estimates B to report an odds ratio
>> (which,
>> if I recall, should be euler's number ^ B).
>>
>> Problem: When I enter my IVs as predictors and don't include an
>> interaction,
>> the significance tests and chi-square statistics in the "parameter
>> estimates"
>> match the "tests of model effects."  However, when I INCLUDE an
>> interaction
>> (which I need to do), the parameter estimates stuff does not match the
>> tests
>> of model effects.  I don't really know why this is, but I know it has
>> something to do with dummy coding, or lack thereof.
>>
>> To get accurate parameter estimates for the interaction and main effects,
>> do I
>> need to dummy code?  If so, is this what I would do?  Right now, both
>> variables are coded 0, 1.  I'm adding in the interaction on the GEE

"model

>> tab."
>>
>>
>
>I suspect a lot of us are not sure what your output looks like.  Could you
>post some sample data and your GEE code?
>
>
>
>-----
>--
>Bruce Weaver
>[hidden email]
>http://sites.google.com/a/lakeheadu.ca/bweaver/
>
>"When all else fails, RTFM."
>
>NOTE: My Hotmail account is not monitored regularly.
>To send me an e-mail, please use the address shown above.
>
>--
>View this message in context: http://spssx-

discussion.1045642.n5.nabble.com/Dummy-Coding-Screwy-Parameter-Estimates-
tp2841406p2841431.html
>Sent from the SPSSX Discussion mailing list archive at 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