ANCOVA doubt

[marg88]

Hi,

Can anyone tell me if I can make an ANCOVA with an ordinal dependent variable?

Independent variable: diagnostic group (I have 3 groups)
Dependent variable: result on a specific test (not continuuos, one can only score 0; 0,5; 1; 2 or 3)
Covariates: age and educational level

Thanks in advance.

MV

[Art Kendall]

I would suggest that you try CATREG.  It allows you to test whether it makes a difference whether you treat the DV as strictly ordinal or not too discrepant from interval level.

What is your DV?  Are those the only results the test is designed to yield or or those the only values your study came up with?

Did you look at a 3D scatterplot? perhaps using colors for diagnostic group?


Do you have enough cases to go beyond a simple ancova and include interactions the involve age and education?

How is education measured?
On 1/13/2013 3:23 PM, marg88 wrote:

Hi,

Can anyone tell me if I can make an ANCOVA with an ordinal dependent
variable?

Independent variable: diagnostic group (I have 3 groups)
Dependent variable: result on a specific test (not continuuos, one can only
score 0; 0,5; 1; 2 or 3)
Covariates: age and educational level

Thanks in advance.

MV



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/ANCOVA-doubt-tp5717348.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

[Rich Ulrich]

The absence of dozens of scores is not the important criterion
in deciding whether a outcome is suited for ANOVA methods -- the
matter of "equal intervals" counts a lot more.  So, for instance, a
dichotomous variable has an "equal interval" and can used as a
criterion in either a discriminant function or a regression, and it
will produce exactly the same tests (although some coefficients, etc.,
are standardized differently). 

When I see a criterion specified as (0, 0.5, 1, 2, 3), it ordinarily looks
to me like *someone*  has had their own criterion of "equal intervals"
in mind when they set up the points.  Thus, I would expect that the
regression on those scores would give the best, desired results.  - Any
rank-based testing replaces those numbers with average-ranks at each
level, and essentially performs ANOVA with those numbers while pretending
(for the sake of assumptions) that these (tied) rank-scores are equally spaced. 

Another poster suggested a way to look at the implied spacing of scores.

Another way to do a variance-based analysis would be to perform a
discriminant analysis - SPSS probably wants the groups to be re-scored
as integers for that. 

--
Rich Ulrich

> Date: Sun, 13 Jan 2013 12:23:23 -0800

> From: [hidden email]
> Subject: ANCOVA doubt
> To: [hidden email]
>
> Hi,
>
> Can anyone tell me if I can make an ANCOVA with an ordinal dependent
> variable?
>
> Independent variable: diagnostic group (I have 3 groups)
> Dependent variable: result on a specific test (not continuuos, one can only
> score 0; 0,5; 1; 2 or 3)
> Covariates: age and educational level
>

...

[marg88]

Thank you for your answer.
I'm not familiar with CATREG, and I have a very basic knowledge of SPSS.

My DV is a total result on a specific test and answers can only be 0; 0,5; 1; 2 or 3. It was designed this way. The interval is not always the same so I treat it like an ordinal variable. Subjects score all these results, depending on severity of the disease.
I have nearly 300 cases, 100 in each of the 3 diagnostic groups (IV).
Education is measured in two ways: number of completed schooling years and in intervals (0; 1-4 years; 5-9 years; beyond 10). For the ANCOVA the interval variable is more suitable for covariate right?

Thank you once again,

MV

[Bruce Weaver]

Depending on your license, you may not have access to CATREG.  (The license at my university does not include it.)  

Have you looked at ordinal logistic regression (i.e., the PLUM procedure)?  If the proportional odds assumption is untenable, you could fall back on multinomial logistic regression (NOMREG).  

HTH.

marg88 wrote

Thank you for your answer.
I'm not familiar with CATREG, and I have a very basic knowledge of SPSS.

My DV is a total result on a specific test and answers can only be 0; 0,5; 1; 2 or 3. It was designed this way. The interval is not always the same so I treat it like an ordinal variable. Subjects score all these results, depending on severity of the disease.
I have nearly 300 cases, 100 in each of the 3 diagnostic groups (IV).
Education is measured in two ways: number of completed schooling years and in intervals (0; 1-4 years; 5-9 years; beyond 10). For the ANCOVA the interval variable is more suitable for covariate right?

Thank you once again,

MV

[marg88]

Thank you Dr. Ulrich.

[Weinberg, Jerry]



Due to the flu I will be away from my office.

[marg88]

Hi,

The ordinal logistic regression is this one?

http://www.youtube.com/watch?v=Ak_t86zm_sQ

Can I do it with SPSS?

Thanks for the advice,
MV

[marg88]

I forgot to say that my groups are not equivalent in terms of age and education, is this going to be a problem when interpreting the results of regression?

[Bruce Weaver]

Nope.  That is binary logisitic regression.  Videos for ordinal logistic regression are found here:

http://www.youtube.com/watch?v=fRfb3XWJfx4
http://www.youtube.com/watch?v=rHOR2JR3SY4

NOTE:  I have not watched the videos, so cannot comment on their quality.

marg88 wrote

Hi,

The ordinal logistic regression is this one?

http://www.youtube.com/watch?v=Ak_t86zm_sQ

Can I do it with SPSS?

Thanks for the advice,
MV

[marg88]

Great! Thank you.

[marg88]

Warnings
There are 141 (67,1%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.
Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.




After entering the logistic regression, I got this warning, can I present this analyses anyway?

[Bruce Weaver]

From marg88's first post:

Independent variable: diagnostic group (I have 3 groups)
Dependent variable: result on a specific test (not continuuos, one can only score 0; 0,5; 1; 2 or 3)
Covariates: age and educational level

From a later post:

I have nearly 300 cases, 100 in each of the 3 diagnostic groups (IV).
Education is measured in two ways: number of completed schooling years and in intervals (0; 1-4 years; 5-9 years; beyond 10). For the ANCOVA the interval variable is more suitable for covariate right?


Which version of Education did you use?  Is Age continuous or categorical?  Please post your PLUM syntax.  

I'm wondering if you treated a continuous variable as if it is categorical.  On the other hand, I think when NOMREG first become available, it gave a similar warning about empty cells any time the model included a continuous explanatory variable, even when it was treated as continuous.  

marg88 wrote

Warnings
There are 141 (67,1%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.
Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.




After entering the logistic regression, I got this warning, can I present this analyses anyway?

[marg88]

I used the continuous education and age type of variable (it has to be, right?), but I've tried with the ordinal ones and I get the same warning (however, the results are better).

On the first video you posted, we can see the same warning in the output, but it only mentions the number of cells, while "my" warning presents more issues.

PLUM CDRtotal1 BY Group WITH Escolaridade IdadeAvaliação1
  /CRITERIA=CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5) PCONVERGE(1.0E-6) SINGULAR(1.0E-8)
  /LINK=LOGIT
  /PRINT=FIT PARAMETER SUMMARY TPARALLEL.

[Bruce Weaver]

From an earlier post:

Warnings
There are 141 (67,1%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.
Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.


Given that both explanatory variables are continuous, I don't understand why the first warning (about the percentage of empty cells) is given.  With continuous variables, one would not expect to have observations in every "cell" when crosstabulating continuous predictors with the outcome variable.  "Cells" isn't even the right word here, IMO.  So I would pretty much ignore that first warning.

The other two warnings, however, are more concerning.  I'm not entirely certain what they signify, but I might try combining some categories of the outcome variable, if there is a way to do it that makes sense.  I think I would also try multinomial logistic regression (NOMREG procedure) to see if it gives similar warnings.  NOMREG treats the outcome variable as nominal rather than ordinal.

HTH.

marg88 wrote

I used the continuous education and age type of variable (it has to be, right?), but I've tried with the ordinal ones and I get the same warning (however, the results are better).

On the first video you posted, we can see the same warning in the output, but it only mentions the number of cells, while "my" warning presents more issues.

PLUM CDRtotal1 BY Group WITH Escolaridade IdadeAvaliação1
  /CRITERIA=CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5) PCONVERGE(1.0E-6) SINGULAR(1.0E-8)
  /LINK=LOGIT
  /PRINT=FIT PARAMETER SUMMARY TPARALLEL.

[marg88]

Could it be because of the values of the DV (0; 0,5; 1, ...) ?

I've tried the NOMREG and it gives me the same warning, but suggests the same as you did as far as I understood (to merge some categories).

Warnings
There are 899 (78,2%) cells (i.e., dependent variable levels by subpopulations) with zero frequencies.
Unexpected singularities in the Hessian matrix are encountered. This indicates that either some predictor variables should be excluded or some categories should be merged.
The NOMREG procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.


Another try:

Warnings
There are 568 (77,3%) cells (i.e., dependent variable levels by subpopulations) with zero frequencies.
There is possibly a quasi-complete separation in the data. Either the maximum likelihood estimates do not exist or some parameter estimates are infinite.
The NOMREG procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.


I've also tried combining two groups, instead of analysing 3 at the same time, but nothing changes... I'm starting to think SPSS hates me x)

Thanks a lot!

[David Marso]

I suspect in the first case that the warning wizard is suffering from cranial flatulence!
Second case?  Some matrix can't be inverted!

What does a XTAB of CDRtotal1 BY Group look like?
What about PLUM CDRtotal1 BY Group ?
---
Below!
" I'm starting to think SPSS hates me x) "
OK:  Draw a circle on the ground... Paint a 5 pointed star... Light some candles, incense and your desk on fire and repeat after me......
$#$##$$%$%&**(!!!!!!!!
----
Feel better?  CIT is a horrible thing to waste on your cube mates! Must share with HR management.
--
CIT: Computer Induced Tourettes
Coprolalia here I come eeeehaaahh!!!!
---

Bruce Weaver wrote

From an earlier post:

Warnings
There are 141 (67,1%) cells (i.e., dependent variable levels by combinations of predictor variable values) with zero frequencies.
Unexpected singularities in the Fisher Information matrix are encountered. There may be a quasi-complete separation in the data. Some parameter estimates will tend to infinity.
The PLUM procedure continues despite the above warning(s). Subsequent results shown are based on the last iteration. Validity of the model fit is uncertain.


Given that both explanatory variables are continuous, I don't understand why the first warning (about the percentage of empty cells) is given.  With continuous variables, one would not expect to have observations in every "cell" when crosstabulating continuous predictors with the outcome variable.  "Cells" isn't even the right word here, IMO.  So I would pretty much ignore that first warning.

The other two warnings, however, are more concerning.  I'm not entirely certain what they signify, but I might try combining some categories of the outcome variable, if there is a way to do it that makes sense.  I think I would also try multinomial logistic regression (NOMREG procedure) to see if it gives similar warnings.  NOMREG treats the outcome variable as nominal rather than ordinal.

HTH.

marg88 wrote

I used the continuous education and age type of variable (it has to be, right?), but I've tried with the ordinal ones and I get the same warning (however, the results are better).

On the first video you posted, we can see the same warning in the output, but it only mentions the number of cells, while "my" warning presents more issues.

PLUM CDRtotal1 BY Group WITH Escolaridade IdadeAvaliação1
  /CRITERIA=CIN(95) DELTA(0) LCONVERGE(0) MXITER(100) MXSTEP(5) PCONVERGE(1.0E-6) SINGULAR(1.0E-8)
  /LINK=LOGIT
  /PRINT=FIT PARAMETER SUMMARY TPARALLEL.