Question about using MANCOVA with differing covariates

[jiupui]

Hello,

I'm wondering if anyone can provide me with some guidance regarding MANCOVA vs. using multiple ANCOVAs.

I'm testing the impact of any intervention of two groups and examining group differences at post intervention.  However, I am controlling for the baseline scores using ANCOVA (for various reasons I am confident that this is the way to go for my purposes).  However, I recognize that running a separate ANCOVA for each outcome was dangerous as it drives up the type I error risk.

My question is can I use MANCOVA instead- is it even possible under these conditions?  I'm not sure if you can do it with different covariates per analysis.  What I mean by that is for each outcome, I controlled for the baseline score by entering it as a covariate (as recommended).  BUT if I did a MANCOVA then wouldn't ALL the baseline scores get controlled for in each outcome?  I'm not sure if that is the way to go given that not all of the baseline scores correlate with one another.

Any guidance would be much appreciated.  Also if this question does not make sense I am happy to clarify.

Thanks in advance to anyone who can help!

[Bruce Weaver]

You need to read the classic article by Huberty & Morris.

https://bitbucket.org/mhunter/readinglists/src/d8e8010f0b0d/ReadingList_NotreDame/HubertyMorris1989MultivariateVsUnivariate.pdf

HTH.

jiupui wrote

Hello,

I'm wondering if anyone can provide me with some guidance regarding MANCOVA vs. using multiple ANCOVAs.

I'm testing the impact of any intervention of two groups and examining group differences at post intervention.  However, I am controlling for the baseline scores using ANCOVA (for various reasons I am confident that this is the way to go for my purposes).  However, I recognize that running a separate ANCOVA for each outcome was dangerous as it drives up the type I error risk.

My question is can I use MANCOVA instead- is it even possible under these conditions?  I'm not sure if you can do it with different covariates per analysis.  What I mean by that is for each outcome, I controlled for the baseline score by entering it as a covariate (as recommended).  BUT if I did a MANCOVA then wouldn't ALL the baseline scores get controlled for in each outcome?  I'm not sure if that is the way to go given that not all of the baseline scores correlate with one another.

Any guidance would be much appreciated.  Also if this question does not make sense I am happy to clarify.

Thanks in advance to anyone who can help!

[Poes, Matthew Joseph]

You are correct on multiple fronts.  You can't use a MANCOVA in this case, it wouldn't make sense.  The reason to use it would be that it would give you an omnibus test, not really because it would aid in reducing type 1 error.  The covariate control of baseline is the best approach for change scores, so you don't need to defend that to us.  As for the multiple tests, if you are running multiple tests on the same population with the same outcome, but dividing up the population into subgroups, or running different predictors, then you should be using a Bonferoni (or alternative) correction.  If you are using the same population, the same predictors, but only changing the outcome (and its baseline control) then you do not need to correct for this as a multiple test.

HTH

Matthew J Poes
Research Data Specialist
Center for Prevention Research and Development
University of Illinois
510 Devonshire Dr.
Champaign, IL 61820
Phone: [phone]
email: [hidden email]


-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of jiupui
Sent: Sunday, September 09, 2012 11:06 AM
To: [hidden email]
Subject: Question about using MANCOVA with differing covariates

Hello,

I'm wondering if anyone can provide me with some guidance regarding MANCOVA vs. using multiple ANCOVAs.

I'm testing the impact of any intervention of two groups and examining group differences at post intervention.  However, I am controlling for the baseline scores using ANCOVA (for various reasons I am confident that this is the way to go for my purposes).  However, I recognize that running a separate ANCOVA for each outcome was dangerous as it drives up the type I error risk.

My question is can I use MANCOVA instead- is it even possible under these conditions?  I'm not sure if you can do it with different covariates per analysis.  What I mean by that is for each outcome, I controlled for the baseline score by entering it as a covariate (as recommended).  BUT if I did a MANCOVA then wouldn't ALL the baseline scores get controlled for in each outcome?  I'm not sure if that is the way to go given that not all of the baseline scores correlate with one another.

Any guidance would be much appreciated.  Also if this question does not make sense I am happy to clarify.

Thanks in advance to anyone who can help!



--
View this message in context: http://spssx-discussion.1045642.n5.nabble.com/Question-about-using-MANCOVA-with-differing-covariates-tp5714994.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 paper that Bruce Weaver links to is excellent.  And wordy.

In response to particular questions -

Yes, you can do MANCOVA.  Problem 1 -- it is wasteful of power,
because (as you say) it uses every covariate for every outcome.
The loss of power is not especially damaging to the test of rational
hypothesis, even though "not all the baseline scores correlate."
 
But the "rational hypothesis" is only the overall test, "Is there a
difference?"  Problem 2 -- Rejecting it does not even say that one
group is better than the other; only, that they differ in some pattern
of outcomes.  I always consider *that* to be a big waste of power,
not having a focused, 1 d.f.  outcome. (Where 1 variable gives you
a 2-tailed test, 3 variables (say) gives you an 8-tail test: any pattern
may occur.  And so on.)

And the loss of power may be huge if there are many variables.
(Are all your Outcomes really of equal importance?  Should you combine
some, selectively, or drop some?  Consider these questions.)
You could easily find that there is less power for MANCOVA at 5%
than what you get *after* doing Bonferroni correction for the separate
Ancovas an a nominal 1% or smaller.  (The Ancovas also point to where
the difference is seen, which the MANCOVA is not assured to do.)

If they are "distinct enough", you often want separate tests. 
But if you have the hypothesis, "one group has better outcome"
as your over-riding concern, then *I* have always suggested that
you create one composite score to reflect outcome, and test on *that*.

Or choose the single variable that reflects outcome.  In my own experience,
we had an outpatient followup of schizophrenic outpatients, comparing
injectible vs oral meds, which collected a dozen rating scales.  The single
outcome of "superiority" in the protocol was the fact that the "blind" had
to be broken because of clinical worsening.  This would be after
re-hospitalization or when the clinician decided that re-hospitalization
was imminent.

Whether or not you correct for multiple testing depends on your overall
narrative.  That depends partly on what your academic area expects,
and it depends partly on how surprising or all of the outcomes were,
before the experiment was done. 

--
Rich Ulrich

> Date: Sun, 9 Sep 2012 09:06:19 -0700

> From: [hidden email]
> Subject: Question about using MANCOVA with differing covariates
> To: [hidden email]
>
> Hello,
>
> I'm wondering if anyone can provide me with some guidance regarding MANCOVA
> vs. using multiple ANCOVAs.
>
> I'm testing the impact of any intervention of two groups and examining group
> differences at post intervention. However, I am controlling for the
> baseline scores using ANCOVA (for various reasons I am confident that this
> is the way to go for my purposes). However, I recognize that running a
> separate ANCOVA for each outcome was dangerous as it drives up the type I
> error risk.
>
> My question is can I use MANCOVA instead- is it even possible under these
> conditions? I'm not sure if you can do it with different covariates per
> analysis. What I mean by that is for each outcome, I controlled for the
> baseline score by entering it as a covariate (as recommended). BUT if I did
> a MANCOVA then wouldn't ALL the baseline scores get controlled for in each
> outcome? I'm not sure if that is the way to go given that not all of the
> baseline scores correlate with one another.
>
> Any guidance would be much appreciated. Also if this question does not make
> sense I am happy to clarify.
>
> Thanks in advance to anyone who can help!
>
> ...

[Bruce Weaver]

One of the key points I remember taking away from the Huberty & Morris article was this:  If you have a multivariate question use a multivariate test; but if you have a series of univariate questions (i.e., questions about individual outcome variables), use a series of univariate analyses (and correct for multiple tests if necessary).  

I also remember being disabused of the notion that the multivariate test somehow controls for FW error in a manner similar to what happens with Fisher's LSD (where the omnibus F-test must be significant before one proceeds to the pair-wise t-tests).  I don't have time to go dig out the exact quote right now, but I think H&M are pretty explicit about that.

Cheers,
Bruce

Rich Ulrich-2 wrote

The paper that Bruce Weaver links to is excellent.  And wordy.

In response to particular questions -

Yes, you can do MANCOVA.  Problem 1 -- it is wasteful of power,
because (as you say) it uses every covariate for every outcome.
The loss of power is not especially damaging to the test of rational
hypothesis, even though "not all the baseline scores correlate."
 
But the "rational hypothesis" is only the overall test, "Is there a
difference?"  Problem 2 -- Rejecting it does not even say that one
group is better than the other; only, that they differ in some pattern
of outcomes.  I always consider *that* to be a big waste of power,
not having a focused, 1 d.f.  outcome. (Where 1 variable gives you
a 2-tailed test, 3 variables (say) gives you an 8-tail test: any pattern
may occur.  And so on.)

And the loss of power may be huge if there are many variables.
(Are all your Outcomes really of equal importance?  Should you combine
some, selectively, or drop some?  Consider these questions.)
You could easily find that there is less power for MANCOVA at 5%
than what you get *after* doing Bonferroni correction for the separate
Ancovas an a nominal 1% or smaller.  (The Ancovas also point to where
the difference is seen, which the MANCOVA is not assured to do.)

If they are "distinct enough", you often want separate tests.  
But if you have the hypothesis, "one group has better outcome"
as your over-riding concern, then *I* have always suggested that
you create one composite score to reflect outcome, and test on *that*.

Or choose the single variable that reflects outcome.  In my own experience,
we had an outpatient followup of schizophrenic outpatients, comparing
injectible vs oral meds, which collected a dozen rating scales.  The single
outcome of "superiority" in the protocol was the fact that the "blind" had
to be broken because of clinical worsening.  This would be after
re-hospitalization or when the clinician decided that re-hospitalization
was imminent.

Whether or not you correct for multiple testing depends on your overall
narrative.  That depends partly on what your academic area expects,
and it depends partly on how surprising or all of the outcomes were,
before the experiment was done.  

--
Rich Ulrich

> Date: Sun, 9 Sep 2012 09:06:19 -0700
> From: [hidden email]
> Subject: Question about using MANCOVA with differing covariates
> To: [hidden email]
>
> Hello,
>
> I'm wondering if anyone can provide me with some guidance regarding MANCOVA
> vs. using multiple ANCOVAs.
>
> I'm testing the impact of any intervention of two groups and examining group
> differences at post intervention.  However, I am controlling for the
> baseline scores using ANCOVA (for various reasons I am confident that this
> is the way to go for my purposes).  However, I recognize that running a
> separate ANCOVA for each outcome was dangerous as it drives up the type I
> error risk.
>
> My question is can I use MANCOVA instead- is it even possible under these
> conditions?  I'm not sure if you can do it with different covariates per
> analysis.  What I mean by that is for each outcome, I controlled for the
> baseline score by entering it as a covariate (as recommended).  BUT if I did
> a MANCOVA then wouldn't ALL the baseline scores get controlled for in each
> outcome?  I'm not sure if that is the way to go given that not all of the
> baseline scores correlate with one another.
>
> Any guidance would be much appreciated.  Also if this question does not make
> sense I am happy to clarify.
>
> Thanks in advance to anyone who can help!
>
> ...

[SR Millis-3]

In Regresson Modeling Strategies (2001), Harrell discusses the use of the logistic model for the simultaneous comparison of multiple factors/covariates between 2 groupss"

"Since the logistic model makes no asumption regarding the distribution of the descriptor variables, it can easily test for simultaneous group differences involving a mixture of continuous, binary, or nominal variables. In observational studies, such models for treatment received or exposure (propensity score models) hold great promise for adjusting for confounding." (p. 227)

Scott Millis

 

---

From: Bruce Weaver <[hidden email]>
To: [hidden email]
Sent: Monday, September 10, 2012 1:43 PM
Subject: Re: Question about using MANCOVA with differing covariates

One of the key points I remember taking away from the Huberty & Morris
article was this:  If you have a multivariate question use a multivariate
test; but if you have a series of univariate questions (i.e., questions
about individual outcome variables), use a series of univariate analyses
(and correct for multiple tests if necessary).

I also remember being disabused of the notion that the multivariate test
somehow controls for FW error in a manner similar to what happens with
Fisher's LSD (where the omnibus F-test must be significant before one
proceeds to the pair-wise t-tests).  I don't have time to go dig out the
exact quote right now, but I think H&M are pretty explicit about that.

Cheers,
Bruce



Rich Ulrich-2 wrote

>
> The paper that Bruce Weaver links to is excellent.  And wordy.
>
> In response to particular questions -
>
> Yes, you can do MANCOVA.  Problem 1 -- it is wasteful of power,
> because (as you say) it uses every covariate for every outcome.
> The loss of power is not especially damaging to the test of rational
> hypothesis, even though "not all the baseline scores correlate."
>
> But the "rational hypothesis" is only the overall test, "Is there a
> difference?"  Problem 2 -- Rejecting it does not even say that one
> group is better than the other; only, that they differ in some pattern
> of outcomes.  I always consider *that* to be a big waste of power,
> not having a focused, 1 d.f.  outcome. (Where 1 variable gives you
> a 2-tailed test, 3 variables (say) gives you an 8-tail test: any pattern
> may occur.  And so on.)
>
> And the loss of power may be huge if there are many variables.
> (Are all your Outcomes really of equal importance?  Should you combine
> some, selectively, or drop some?  Consider these questions.)
> You could easily find that there is less power for MANCOVA at 5%
> than what you get *after* doing Bonferroni correction for the separate
> Ancovas an a nominal 1% or smaller.  (The Ancovas also point to where
> the difference is seen, which the MANCOVA is not assured to do.)
>
> If they are "distinct enough", you often want separate tests.
> But if you have the hypothesis, "one group has better outcome"
> as your over-riding concern, then *I* have always suggested that
> you create one composite score to reflect outcome, and test on *that*.
>
> Or choose the single variable that reflects outcome.  In my own
> experience,
> we had an outpatient followup of schizophrenic outpatients, comparing
> injectible vs oral meds, which collected a dozen rating scales.  The
> single
> outcome of "superiority" in the protocol was the fact that the "blind" had
> to be broken because of clinical worsening.  This would be after
> re-hospitalization or when the clinician decided that re-hospitalization
> was imminent.
>
> Whether or not you correct for multiple testing depends on your overall
> narrative.  That depends partly on what your academic area expects,
> and it depends partly on how surprising or all of the outcomes were,
> before the experiment was done.
>
> --
> Rich Ulrich
>
>> Date: Sun, 9 Sep 2012 09:06:19 -0700
>> From: jenstein@
>> Subject: Question about using MANCOVA with differing covariates
>> To: SPSSX-L@.UGA
>>
>> Hello,
>>
>> I'm wondering if anyone can provide me with some guidance regarding
>> MANCOVA
>> vs. using multiple ANCOVAs.
>>
>> I'm testing the impact of any intervention of two groups and examining
>> group
>> differences at post intervention.  However, I am controlling for the
>> baseline scores using ANCOVA (for various reasons I am confident that
>> this
>> is the way to go for my purposes).  However, I recognize that running a
>> separate ANCOVA for each outcome was dangerous as it drives up the type I
>> error risk.
>>
>> My question is can I use MANCOVA instead- is it even possible under these
>> conditions?  I'm not sure if you can do it with different covariates per
>> analysis.  What I mean by that is for each outcome, I controlled for the
>> baseline score by entering it as a covariate (as recommended).  BUT if I
>> did
>> a MANCOVA then wouldn't ALL the baseline scores get controlled for in
>> each
>> outcome?  I'm not sure if that is the way to go given that not all of the
>> baseline scores correlate with one another.
>>
>> Any guidance would be much appreciated.  Also if this question does not
>> make
>> sense I am happy to clarify.
>>
>> Thanks in advance to anyone who can help!
>>
>> ...
>

-----
--
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/Question-about-using-MANCOVA-with-differing-covariates-tp5714994p5715009.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

[Rich Ulrich]

This comment always annoys me --  "Since the logistic model makes
no assumption regarding the distribution of the descriptor variables ...."

What is wrong with it is that it draws an implicit contrast to
OLS, which sometimes states "assumptions";  but it conceals
from the user how seldom there is benefit from that contrast.
And it wrongly suggests that you can get away with lousy
distributions in the predictors.

The important consideration that I have in mind is that LR and
OLS each generate a prediction equation that is linear in the
predictors.  Both of these can accept a "mixture of continuous,
binary, or nominal variables" (dummy coded, for the latter). 

It does seem to me that if you want to create a prediction equation
that is going to be robust and reproducible across samples, you
want the same sort of distributional assumptions in the IVs in both
cases.

 - After you have that firmly in mind, they you can start to pay
attention to the distinctions in the testing...  where LR is better for
highly successful prediction (large R^2 for OLS), but for robustness,
requires that the sample be large enough that you don't end up
with perfect (or near-perfect) discrimination of the two samples. 

--
Rich Ulrich

---

Date: Mon, 10 Sep 2012 12:46:07 -0700
From: [hidden email]
Subject: Re: Question about using MANCOVA with differing covariates
To: [hidden email]

In Regresson Modeling Strategies (2001), Harrell discusses the use of the logistic model for the simultaneous comparison of multiple factors/covariates between 2 groupss"

"Since the logistic model makes no asumption regarding the distribution of the descriptor variables, it can easily test for simultaneous group differences involving a mixture of continuous, binary, or nominal variables. In observational studies, such models for treatment received or exposure (propensity score models) hold great promise for adjusting for confounding." (p. 227)

Scott Millis

 

[... snip, previous]

[SR Millis-3]

You're missing the point, Rich.

The contrast is made with MANOVA---NOT OLS.  

Scott Millis

---

From: Rich Ulrich <[hidden email]>
To: [hidden email]
Sent: Monday, September 10, 2012 11:19 PM
Subject: Re: Question about using MANCOVA with differing covariates

This comment always annoys me --  "Since the logistic model makes
no assumption regarding the distribution of the descriptor variables ...."

What is wrong with it is that it draws an implicit contrast to
OLS, which sometimes states "assumptions";  but it conceals
from the user how seldom there is benefit from that contrast.
And it wrongly suggests that you can get away with lousy
distributions in the predictors.

The important consideration that I have in mind is that LR and
OLS each generate a prediction equation that is linear in the
predictors.  Both of these can accept a "mixture of continuous,
binary, or nominal variables" (dummy coded, for the latter). 

It does seem to me that if you want to create a prediction equation
that is going to be robust and reproducible across samples, you
want the same sort of distributional assumptions in the IVs in both
cases.

 - After you have that firmly in mind, they you can start to pay
attention to the distinctions in the testing...  where LR is better for
highly successful prediction (large R^2 for OLS), but for robustness,
requires that the sample be large enough that you don't end up
with perfect (or near-perfect) discrimination of the two samples. 

--
Rich Ulrich

---

Date: Mon, 10 Sep 2012 12:46:07 -0700
From: [hidden email]
Subject: Re: Question about using MANCOVA with differing covariates
To: [hidden email]

In Regresson Modeling Strategies (2001), Harrell discusses the use of the logistic model for the simultaneous comparison of multiple factors/covariates between 2 groupss"

"Since the logistic model makes no asumption regarding the distribution of the descriptor variables, it can easily test for simultaneous group differences involving a mixture of continuous, binary, or nominal variables. In observational studies, such models for treatment received or exposure (propensity score models) hold great promise for adjusting for confounding." (p. 227)

Scott Millis

 

[... snip, previous]

[Rich Ulrich]

If I was missing the point... I'm afraid I still miss it.

The comment you cite is about logistic regression, which has
a dichotomous criterion.  If the contrast is MANOVA, it does
seem to me that the subset of MANOVA invoked is 2-group
discriminant function, which is fully equivalent to OLS
regression with a 0/1 criterion.  The assumptions and
characteristics of OLS should be what pertain.

What am I missing?

--
Rich Ulrich

---

Date: Tue, 11 Sep 2012 05:11:07 -0700
From: [hidden email]
Subject: Re: Question about using MANCOVA with differing covariates
To: [hidden email]

You're missing the point, Rich.

The contrast is made with MANOVA---NOT OLS.  

Scott Millis

---

From: Rich Ulrich <[hidden email]>
To: [hidden email]
Sent: Monday, September 10, 2012 11:19 PM
Subject: Re: Question about using MANCOVA with differing covariates

This comment always annoys me --  "Since the logistic model makes
no assumption regarding the distribution of the descriptor variables ...."

What is wrong with it is that it draws an implicit contrast to
OLS, which sometimes states "assumptions";  but it conceals
from the user how seldom there is benefit from that contrast.
And it wrongly suggests that you can get away with lousy
distributions in the predictors.

The important consideration that I have in mind is that LR and
OLS each generate a prediction equation that is linear in the
predictors.  Both of these can accept a "mixture of continuous,
binary, or nominal variables" (dummy coded, for the latter). 

It does seem to me that if you want to create a prediction equation
that is going to be robust and reproducible across samples, you
want the same sort of distributional assumptions in the IVs in both
cases.

 - After you have that firmly in mind, they you can start to pay
attention to the distinctions in the testing...  where LR is better for
highly successful prediction (large R^2 for OLS), but for robustness,
requires that the sample be large enough that you don't end up
with perfect (or near-perfect) discrimination of the two samples. 

--
Rich Ulrich

---

Date: Mon, 10 Sep 2012 12:46:07 -0700
From: [hidden email]
Subject: Re: Question about using MANCOVA with differing covariates
To: [hidden email]

In Regresson Modeling Strategies (2001), Harrell discusses the use of the logistic model for the simultaneous comparison of multiple factors/covariates between 2 groupss"

"Since the logistic model makes no asumption regarding the distribution of the descriptor variables, it can easily test for simultaneous group differences involving a mixture of continuous, binary, or nominal variables. In observational studies, such models for treatment received or exposure (propensity score models) hold great promise for adjusting for confounding." (p. 227)

Scott Millis

 

[... snip, previous]

[John Morris]

Yep, a perfect capture.

Huberty, C. J, & Morris, J. D. (1989). Multivariate analysis versus multiple univariate analyses. Psychological Bulletin, 105, 302-308.

Dan Morris
=================

-----Original Message-----
From: SPSSX(r) Discussion [mailto:[hidden email]] On Behalf Of Bruce Weaver
Sent: Monday, September 10, 2012 1:43 PM
To: [hidden email]
Subject: Re: Question about using MANCOVA with differing covariates

One of the key points I remember taking away from the Huberty & Morris article was this:  If you have a multivariate question use a multivariate test; but if you have a series of univariate questions (i.e., questions about individual outcome variables), use a series of univariate analyses (and correct for multiple tests if necessary).

I also remember being disabused of the notion that the multivariate test somehow controls for FW error in a manner similar to what happens with Fisher's LSD (where the omnibus F-test must be significant before one proceeds to the pair-wise t-tests).  I don't have time to go dig out the exact quote right now, but I think H&M are pretty explicit about that.

Cheers,
Bruce



Rich Ulrich-2 wrote

>
> The paper that Bruce Weaver links to is excellent.  And wordy.
>
> In response to particular questions -
>
> Yes, you can do MANCOVA.  Problem 1 -- it is wasteful of power,
> because (as you say) it uses every covariate for every outcome.
> The loss of power is not especially damaging to the test of rational
> hypothesis, even though "not all the baseline scores correlate."
>
> But the "rational hypothesis" is only the overall test, "Is there a
> difference?"  Problem 2 -- Rejecting it does not even say that one
> group is better than the other; only, that they differ in some pattern
> of outcomes.  I always consider *that* to be a big waste of power, not
> having a focused, 1 d.f.  outcome. (Where 1 variable gives you a
> 2-tailed test, 3 variables (say) gives you an 8-tail test: any pattern
> may occur.  And so on.)
>
> And the loss of power may be huge if there are many variables.
> (Are all your Outcomes really of equal importance?  Should you combine
> some, selectively, or drop some?  Consider these questions.) You could
> easily find that there is less power for MANCOVA at 5% than what you
> get *after* doing Bonferroni correction for the separate Ancovas an a
> nominal 1% or smaller.  (The Ancovas also point to where the
> difference is seen, which the MANCOVA is not assured to do.)
>
> If they are "distinct enough", you often want separate tests.
> But if you have the hypothesis, "one group has better outcome"
> as your over-riding concern, then *I* have always suggested that you
> create one composite score to reflect outcome, and test on *that*.
>
> Or choose the single variable that reflects outcome.  In my own
> experience, we had an outpatient followup of schizophrenic
> outpatients, comparing injectible vs oral meds, which collected a
> dozen rating scales.  The single outcome of "superiority" in the
> protocol was the fact that the "blind" had to be broken because of
> clinical worsening.  This would be after re-hospitalization or when
> the clinician decided that re-hospitalization was imminent.
>
> Whether or not you correct for multiple testing depends on your
> overall narrative.  That depends partly on what your academic area
> expects, and it depends partly on how surprising or all of the
> outcomes were, before the experiment was done.
>
> --
> Rich Ulrich
>
>> Date: Sun, 9 Sep 2012 09:06:19 -0700
>> From: jenstein@
>> Subject: Question about using MANCOVA with differing covariates
>> To: SPSSX-L@.UGA
>>
>> Hello,
>>
>> I'm wondering if anyone can provide me with some guidance regarding
>> MANCOVA vs. using multiple ANCOVAs.
>>
>> I'm testing the impact of any intervention of two groups and
>> examining group differences at post intervention.  However, I am
>> controlling for the baseline scores using ANCOVA (for various reasons
>> I am confident that this is the way to go for my purposes).  However,
>> I recognize that running a separate ANCOVA for each outcome was
>> dangerous as it drives up the type I error risk.
>>
>> My question is can I use MANCOVA instead- is it even possible under
>> these conditions?  I'm not sure if you can do it with different
>> covariates per analysis.  What I mean by that is for each outcome, I
>> controlled for the baseline score by entering it as a covariate (as
>> recommended).  BUT if I did a MANCOVA then wouldn't ALL the baseline
>> scores get controlled for in each outcome?  I'm not sure if that is
>> the way to go given that not all of the baseline scores correlate
>> with one another.
>>
>> Any guidance would be much appreciated.  Also if this question does
>> not make sense I am happy to clarify.
>>
>> Thanks in advance to anyone who can help!
>>
>> ...
>

-----
--
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/Question-about-using-MANCOVA-with-differing-covariates-tp5714994p5715009.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