Unianova discrepancy

[Roland Stark]

SPSS v26's Unianova procedure can produce discrepancies between results that appear in the Tests of Between-Subjects Effects table and results in the Parameter Estimates table.  In the presence of other predictors, including an interaction involving the first predictor, that predictor's main effect will have different P and Partial Eta-Squared results from one table to the other.  I found no such discrepancy if I removed the interaction term.  Thus it doesn't seem to be a matter of deviation vs. indicator contrasts, as stated here:  https://stats.stackexchange.com/questions/209815/why-does-spss-give-different-p-values-in-the-factorial-anova-table-and-the-param .  I would appreciate any help making sense of this.

=====================
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]

In the Parameter Estimates table, the -.924 for agegrp=1 is the simple effect
of age=1 vs age=2 WHEN gender=2.  And the -3.706 for gender=1 is the simple
effect of gender=1 vs gender=2 WHEN agegrp=2.  

In the Tests of Between-Subjects Effects tables, on the other hand, you are
getting tests of the conventional ANOVA main effects--i.e., collapsing
across the levels of the other variable.  

One easy way to see this would be to include a couple of EMMEANS commands as
follows:

 /EMMEANS=TABLES(agegrp*gender) COMPARE(agegrp)
 /EMMEANS=TABLES(agegrp*gender) COMPARE(gender)

COMPARE(agegrp) will give you the simple effects of agegrp at each level of
Gender, and one of those effects will match what you see in the table of
parameter estimates.  

COMPARE(gender) will give you the simple effects of gender at each level of
agegrp, and one of those effects will match what you see in the table of
parameter estimates.

HTH.



Roland Stark wrote

> SPSS v26's Unianova procedure can produce discrepancies between results
> that appear in the Tests of Between-Subjects Effects table and results in
> the Parameter Estimates table.  In the presence of other predictors,
> including an interaction involving the first predictor, that predictor's
> main effect will have different P and Partial Eta-Squared results from one
> table to the other.  I found no such discrepancy if I removed the
> interaction term.  Thus it doesn't seem to be a matter of deviation vs.
> indicator contrasts, as stated here:
> https://stats.stackexchange.com/questions/209815/why-does-spss-give-different-p-values-in-the-factorial-anova-table-and-the-param
> .  I would appreciate any help making sense of this.

-----
--
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.

--
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

[Kirill Orlov]

Will you offer a data and the results which bother you?


29.09.2020 17:01, Roland Stark пишет:

> SPSS v26's Unianova procedure can produce discrepancies between results that appear in the Tests of Between-Subjects Effects table and results in the Parameter Estimates table.  In the presence of other predictors, including an interaction involving the first predictor, that predictor's main effect will have different P and Partial Eta-Squared results from one table to the other.  I found no such discrepancy if I removed the interaction term.  Thus it doesn't seem to be a matter of deviation vs. indicator contrasts, as stated here:  https://stats.stackexchange.com/questions/209815/why-does-spss-give-different-p-values-in-the-factorial-anova-table-and-the-param .  I would appreciate any help making sense of this.
>
> =====================
> 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

[Kirill Orlov]

In the example from the link it is clear that there is a two-factor
ANOVA with 2 levels in each factor. (The factors are AGEGP and GENDER.)
Under dummy (indicator) coding each of the factors are turned into two
variables with values 0 and 1, and under deviation coding - with values
-1 and 1.

If you run linear regression with these variables representing the main
effects you will find that there is no difference in parameter estimates
(= regressional coefficients) in case when a predictor is coded 0 1 and
when it is coded  -1 1. (This is because there is a constant term in the
model which compensates for the difference in the two coding schemes.)

It explains why the results of GLM - in "Parameter estimates" table and
in "Between-subject effects" table the results (the p values) will be
the same under the main-effects-only model.

But if you include, besides the main effects, also the interaction term
between the factors, the discrepancy that is due to the differential
coding schemes - dummy vs deviation - will show through. And you will
observe the "discrepancy" between the two tables in the manner displayed
in the linked question.

To see all it, just do the linear regressions with the below toy
dataset. Y - dependent, F1 F2 - the factors, a1_i b1_i a1b1_i - the
three contrast variables representing the two factors and their
interaction under dummy (indicator) encoding, a1_d b1_d a1b1_d - the
three contrast variables representing the two factors and their
interaction under deviation encoding.

        y       f1       f2     a1_i     b1_i   a1b1_i a1_d     b1_d  
a1b1_d

     1.00        1        1        1        1 1        1        1        1
     2.00        1        1        1        1 1        1        1        1
     3.00        1        1        1        1 1        1        1        1
     4.00        1        1        1        1 1        1        1        1
     4.00        1        1        1        1 1        1        1        1
     5.00        1        1        1        1 1        1        1        1
     3.00        1        1        1        1 1        1        1        1
     1.00        1        2        1        0 0        1       -1       -1
     2.00        1        2        1        0 0        1       -1       -1
     4.00        1        2        1        0 0        1       -1       -1
     3.00        1        2        1        0 0        1       -1       -1
     5.00        1        2        1        0 0        1       -1       -1
     5.00        1        2        1        0 0        1       -1       -1
     6.00        1        2        1        0 0        1       -1       -1
     4.00        1        2        1        0 0        1       -1       -1
     4.00        1        2        1        0 0        1       -1       -1
     5.00        1        2        1        0 0        1       -1       -1
     4.00        2        1        0        1        0 -1        1       -1
     6.00        2        1        0        1        0 -1        1       -1
     7.00        2        1        0        1        0 -1        1       -1
     6.00        2        1        0        1        0 -1        1       -1
     5.00        2        1        0        1        0 -1        1       -1
     8.00        2        1        0        1        0 -1        1       -1
     7.00        2        1        0        1        0 -1        1       -1
     4.00        2        1        0        1        0 -1        1       -1
     1.00        2        1        0        1        0 -1        1       -1
     7.00        2        1        0        1        0 -1        1       -1
     6.00        2        1        0        1        0 -1        1       -1
     6.00        2        2        0        0        0 -1       -1        1
     5.00        2        2        0        0        0 -1       -1        1
     5.00        2        2        0        0        0 -1       -1        1
     6.00        2        2        0        0        0 -1       -1        1
     7.00        2        2        0        0        0 -1       -1        1
     6.00        2        2        0        0        0 -1       -1        1
     5.00        2        2        0        0        0 -1       -1        1
     6.00        2        2        0        0        0 -1       -1        1
     2.00        2        2        0        0        0 -1       -1        1
     5.00        2        2        0        0        0 -1       -1        1
     6.00        2        2        0        0        0 -1       -1        1
     8.00        2        2        0        0        0 -1       -1        1


Number of cases read:  40    Number of cases listed:  40


29.09.2020 17:01, Roland Stark пишет:

> SPSS v26's Unianova procedure can produce discrepancies between results that appear in the Tests of Between-Subjects Effects table and results in the Parameter Estimates table.  In the presence of other predictors, including an interaction involving the first predictor, that predictor's main effect will have different P and Partial Eta-Squared results from one table to the other.  I found no such discrepancy if I removed the interaction term.  Thus it doesn't seem to be a matter of deviation vs. indicator contrasts, as stated here:  https://stats.stackexchange.com/questions/209815/why-does-spss-give-different-p-values-in-the-factorial-anova-table-and-the-param .  I would appreciate any help making sense of this.
>
> =====================
> 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

[Kirill Orlov]

... if you are attentive, you notice how different are the data in the
variables representing the interaction, a1b1_i and a1b1_d.

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