Transforming variables

[Benjamin Spivak (Med)]

Hello everybody,

I would like to get people's general opinion on transforming variables. Is it a bad thing?, does it complicate interpretation?

I ask because I am performing a MANOVA for some jury comprehension research. I have 3 DV's, application, comprehension and recognition. One of my DV's is skewed and has heterogenous variance. When I apply a square root transformation, the problem is fixed. However, I'm worried that interpretation may be difficult with one transformed dv and two untransformed DV's. Any opinions would be much appreciated.

Cheers,

Ben.

[Rich Ulrich]

 - see below -

---

Date: Fri, 19 Aug 2011 17:07:52 +1000
From: [hidden email]
Subject: Transforming variables
To: [hidden email]

> Hello everybody,

> I would like to get people's general opinion on transforming variables. Is it a bad thing?, does it complicate interpretation?

Yes, transformation can complicate both presentation and interpretation.

So, if you are dealing with a single use of a particular variable, and it
does *not* make any important difference in the modeling or tests,
than using a transformation can be a bad thing because it distracts
from the discussion.

However, if you are using the variable often, or the untransformed
version is damaging to the model or the tests, then you really ought
to use a transformation.  (But what is "damaging"?)

> I ask because I am performing a MANOVA for some jury comprehension research. I have 3 DV's, application,
> comprehension and recognition. One of my DV's is skewed and has heterogenous variance. When I apply a
> square root transformation, the problem is fixed. However, I'm worried that interpretation may be difficult
> with one transformed dv and two untransformed DV's. Any opinions would be much appreciated.

Skewness and heterogeneous variance are useful warning signs that
a natural measurement is not being taken in its most useful units, but
that is not sufficient reason to transform.  Is the transformation a
"natural" one for the variable?  - I eventually decided that looking at
the scatter of Pre-Post, or any two replications, was useful for deciding
whether the measurement itself has heterogeneous error, or if
heterogeneity was a feature of the sampling.  However, no matter what
the variable itself looks like, linearity with the immediate Dependent Variable
is more important than heterogeneity for constructing a good model, even
if that means that you need to do something special for the testing.

 - Since you describe these as DVs, not IVs, it could be that you *should*
be concerned with the sampling error itself.  When I read "comprehension
and recognition", it occurs to me to wonder if these are both min-max-
limited ranges, which might be properly scaled as 0-100% and transformed
to logit, if one wished to check out the "natural" scaling against each other,
or against selected predictors.  Again - returning to my initial comment -
that could distract from the discussion and modeling if it makes no difference
for the data in hand.

For your data on hand:  You do not mention the N.  Tests for skewness,
etc., have too much power for large N, considering the robustness of
least squares.  That is to say, your tests *might*  still be in fine shape
even without transforming.  So, how extreme is the situation?  Are the
numbers skewed enough that you feel uncomfortable using the mean(s) as
the description? (That's a pretty good sign that transformation is desirable.)

Your square-root transformation is not as strong as taking log or reciprocal,
though the effective strength also depends on the range being close to zero. 
My point here is that a weaker transform is less likely to matter to the tests
or model than a stronger one.  "Square root" is more likely to be ignorable
than "reciprocal". 

The ultimate test of whether a transformation matters for your data is
to try it both ways and see, for this analysis, "Does it matter?"  If it
doesn't matter, then you have little problem.  Present the analysis that
you want, and mention (for any subtle critics) that you also did it the
other way and the results were the same.  If it does matter, then *I*
would want to check on the face-value of the linearity of the DV  with
the other variables in the model.

--
Rich Ulrich




 

[Bruce Weaver]

I'll add one comment to what Rich has said:  The assumption of normality for OLS models applies to the errors (which are estimated with the residuals--see link below), and normality of the errors is considerably less important than independence of the errors.  What this suggests is that you fit the model and then examine the residuals via residual plots etc.

   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics

HTH.

Rich Ulrich-2 wrote

- see below -

Date: Fri, 19 Aug 2011 17:07:52 +1000
From: [hidden email]
Subject: Transforming variables
To: [hidden email]

> Hello everybody,
> I would like to get people's general opinion on transforming variables. Is it a bad thing?, does it complicate interpretation?
Yes, transformation can complicate both presentation and interpretation.

So, if you are dealing with a single use of a particular variable, and it
does *not* make any important difference in the modeling or tests,
than using a transformation can be a bad thing because it distracts
from the discussion.

However, if you are using the variable often, or the untransformed
version is damaging to the model or the tests, then you really ought
to use a transformation.  (But what is "damaging"?)

> I ask because I am performing a MANOVA for some jury comprehension research. I have 3 DV's, application,
> comprehension and recognition. One of my DV's is skewed and has heterogenous variance. When I apply a
> square root transformation, the problem is fixed. However, I'm worried that interpretation may be difficult
> with one transformed dv and two untransformed DV's. Any opinions would be much appreciated.

Skewness and heterogeneous variance are useful warning signs that
a natural measurement is not being taken in its most useful units, but
that is not sufficient reason to transform.  Is the transformation a
"natural" one for the variable?  - I eventually decided that looking at
the scatter of Pre-Post, or any two replications, was useful for deciding
whether the measurement itself has heterogeneous error, or if
heterogeneity was a feature of the sampling.  However, no matter what
the variable itself looks like, linearity with the immediate Dependent Variable
is more important than heterogeneity for constructing a good model, even
if that means that you need to do something special for the testing.

 - Since you describe these as DVs, not IVs, it could be that you *should*
be concerned with the sampling error itself.  When I read "comprehension
and recognition", it occurs to me to wonder if these are both min-max-
limited ranges, which might be properly scaled as 0-100% and transformed
to logit, if one wished to check out the "natural" scaling against each other,
or against selected predictors.  Again - returning to my initial comment -
that could distract from the discussion and modeling if it makes no difference
for the data in hand.

For your data on hand:  You do not mention the N.  Tests for skewness,
etc., have too much power for large N, considering the robustness of
least squares.  That is to say, your tests *might*  still be in fine shape
even without transforming.  So, how extreme is the situation?  Are the
numbers skewed enough that you feel uncomfortable using the mean(s) as
the description? (That's a pretty good sign that transformation is desirable.)

Your square-root transformation is not as strong as taking log or reciprocal,
though the effective strength also depends on the range being close to zero.  
My point here is that a weaker transform is less likely to matter to the tests
or model than a stronger one.  "Square root" is more likely to be ignorable
than "reciprocal".  

The ultimate test of whether a transformation matters for your data is
to try it both ways and see, for this analysis, "Does it matter?"  If it
doesn't matter, then you have little problem.  Present the analysis that
you want, and mention (for any subtle critics) that you also did it the
other way and the results were the same.  If it does matter, then *I*
would want to check on the face-value of the linearity of the DV  with
the other variables in the model.

--
Rich Ulrich

[Benjamin Spivak (Med)]

Thank you all so much for your help so far.

I'm considering performing a rank transformation and then running separate ANOVAs for each DV. Is there anything I ought to be aware of? Also, When I am transforming to rank, do I rank the entire variable or do I rank within each condition?

Thanks in advance,

Ben.

On 20 August 2011 22:35, Bruce Weaver <[hidden email]> wrote:

I'll add one comment to what Rich has said:  The assumption of normality for
OLS models applies to the errors (which are estimated with the
residuals--see link below), and normality of the errors is considerably less
important than independence of the errors.  What this suggests is that you
fit the model and then examine the residuals via residual plots etc.

  http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics

HTH.

Rich Ulrich-2 wrote:
>
> - see below -
>
> Date: Fri, 19 Aug 2011 17:07:52 +1000
> From: [hidden email]
> Subject: Transforming variables
> To: [hidden email]
>
>> Hello everybody,
>> I would like to get people's general opinion on transforming variables.
>> Is it a bad thing?, does it complicate interpretation?
> Yes, transformation can complicate both presentation and interpretation.
>
> So, if you are dealing with a single use of a particular variable, and it
> does *not* make any important difference in the modeling or tests,
> than using a transformation can be a bad thing because it distracts
> from the discussion.
>
> However, if you are using the variable often, or the untransformed
> version is damaging to the model or the tests, then you really ought
> to use a transformation.  (But what is "damaging"?)
>
>> I ask because I am performing a MANOVA for some jury comprehension
>> research. I have 3 DV's, application,
>> comprehension and recognition. One of my DV's is skewed and has
>> heterogenous variance. When I apply a
>> square root transformation, the problem is fixed. However, I'm worried
>> that interpretation may be difficult
>> with one transformed dv and two untransformed DV's. Any opinions would be
>> much appreciated.
>
> Skewness and heterogeneous variance are useful warning signs that
> a natural measurement is not being taken in its most useful units, but
> that is not sufficient reason to transform.  Is the transformation a
> "natural" one for the variable?  - I eventually decided that looking at
> the scatter of Pre-Post, or any two replications, was useful for deciding
> whether the measurement itself has heterogeneous error, or if
> heterogeneity was a feature of the sampling.  However, no matter what
> the variable itself looks like, linearity with the immediate Dependent
> Variable
> is more important than heterogeneity for constructing a good model, even
> if that means that you need to do something special for the testing.
>
>  - Since you describe these as DVs, not IVs, it could be that you *should*
> be concerned with the sampling error itself.  When I read "comprehension
> and recognition", it occurs to me to wonder if these are both min-max-
> limited ranges, which might be properly scaled as 0-100% and transformed
> to logit, if one wished to check out the "natural" scaling against each
> other,
> or against selected predictors.  Again - returning to my initial comment -
> that could distract from the discussion and modeling if it makes no
> difference
> for the data in hand.
>
> For your data on hand:  You do not mention the N.  Tests for skewness,
> etc., have too much power for large N, considering the robustness of
> least squares.  That is to say, your tests *might*  still be in fine shape
> even without transforming.  So, how extreme is the situation?  Are the
> numbers skewed enough that you feel uncomfortable using the mean(s) as
> the description? (That's a pretty good sign that transformation is
> desirable.)
>
> Your square-root transformation is not as strong as taking log or
> reciprocal,
> though the effective strength also depends on the range being close to
> zero.
> My point here is that a weaker transform is less likely to matter to the
> tests
> or model than a stronger one.  "Square root" is more likely to be
> ignorable
> than "reciprocal".
>
> The ultimate test of whether a transformation matters for your data is
> to try it both ways and see, for this analysis, "Does it matter?"  If it
> doesn't matter, then you have little problem.  Present the analysis that
> you want, and mention (for any subtle critics) that you also did it the
> other way and the results were the same.  If it does matter, then *I*
> would want to check on the face-value of the linearity of the DV  with
> the other variables in the model.
>
> --
> Rich Ulrich
>

-----
--
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/Transforming-variables-tp4714751p4718352.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

[Bruce Weaver]

Rank the entire variable.  Re things to be aware of, see this:

http://en.wikipedia.org/wiki/ANOVA_on_ranks#Failure_of_ranking_in_the_factorial_ANOVA_and_other_complex_layouts

Benjamin Spivak (Med) wrote

Thank you all so much for your help so far.

I'm considering performing a rank transformation and then running separate
ANOVAs for each DV. Is there anything I ought to be aware of? Also, When I
am transforming to rank, do I rank the entire variable or do I rank within
each condition?

Thanks in advance,

Ben.

On 20 August 2011 22:35, Bruce Weaver <[hidden email]> wrote:

> I'll add one comment to what Rich has said:  The assumption of normality
> for
> OLS models applies to the errors (which are estimated with the
> residuals--see link below), and normality of the errors is considerably
> less
> important than independence of the errors.  What this suggests is that you
> fit the model and then examine the residuals via residual plots etc.
>
>   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics
>
> HTH.
>
>
> Rich Ulrich-2 wrote:
> >
> > - see below -
> >
> > Date: Fri, 19 Aug 2011 17:07:52 +1000
> > From: [hidden email]
> > Subject: Transforming variables
> > To: [hidden email]
> >
> >> Hello everybody,
> >> I would like to get people's general opinion on transforming variables.
> >> Is it a bad thing?, does it complicate interpretation?
> > Yes, transformation can complicate both presentation and interpretation.
> >
> > So, if you are dealing with a single use of a particular variable, and it
> > does *not* make any important difference in the modeling or tests,
> > than using a transformation can be a bad thing because it distracts
> > from the discussion.
> >
> > However, if you are using the variable often, or the untransformed
> > version is damaging to the model or the tests, then you really ought
> > to use a transformation.  (But what is "damaging"?)
> >
> >> I ask because I am performing a MANOVA for some jury comprehension
> >> research. I have 3 DV's, application,
> >> comprehension and recognition. One of my DV's is skewed and has
> >> heterogenous variance. When I apply a
> >> square root transformation, the problem is fixed. However, I'm worried
> >> that interpretation may be difficult
> >> with one transformed dv and two untransformed DV's. Any opinions would
> be
> >> much appreciated.
> >
> > Skewness and heterogeneous variance are useful warning signs that
> > a natural measurement is not being taken in its most useful units, but
> > that is not sufficient reason to transform.  Is the transformation a
> > "natural" one for the variable?  - I eventually decided that looking at
> > the scatter of Pre-Post, or any two replications, was useful for deciding
> > whether the measurement itself has heterogeneous error, or if
> > heterogeneity was a feature of the sampling.  However, no matter what
> > the variable itself looks like, linearity with the immediate Dependent
> > Variable
> > is more important than heterogeneity for constructing a good model, even
> > if that means that you need to do something special for the testing.
> >
> >  - Since you describe these as DVs, not IVs, it could be that you
> *should*
> > be concerned with the sampling error itself.  When I read "comprehension
> > and recognition", it occurs to me to wonder if these are both min-max-
> > limited ranges, which might be properly scaled as 0-100% and transformed
> > to logit, if one wished to check out the "natural" scaling against each
> > other,
> > or against selected predictors.  Again - returning to my initial comment
> -
> > that could distract from the discussion and modeling if it makes no
> > difference
> > for the data in hand.
> >
> > For your data on hand:  You do not mention the N.  Tests for skewness,
> > etc., have too much power for large N, considering the robustness of
> > least squares.  That is to say, your tests *might*  still be in fine
> shape
> > even without transforming.  So, how extreme is the situation?  Are the
> > numbers skewed enough that you feel uncomfortable using the mean(s) as
> > the description? (That's a pretty good sign that transformation is
> > desirable.)
> >
> > Your square-root transformation is not as strong as taking log or
> > reciprocal,
> > though the effective strength also depends on the range being close to
> > zero.
> > My point here is that a weaker transform is less likely to matter to the
> > tests
> > or model than a stronger one.  "Square root" is more likely to be
> > ignorable
> > than "reciprocal".
> >
> > The ultimate test of whether a transformation matters for your data is
> > to try it both ways and see, for this analysis, "Does it matter?"  If it
> > doesn't matter, then you have little problem.  Present the analysis that
> > you want, and mention (for any subtle critics) that you also did it the
> > other way and the results were the same.  If it does matter, then *I*
> > would want to check on the face-value of the linearity of the DV  with
> > the other variables in the model.
> >
> > --
> > Rich Ulrich
> >
>
>
> -----
> --
> 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/Transforming-variables-tp4714751p4718352.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
>

[Benjamin Spivak (Med)]

Hi Bruce,

Does this mean that I can still perform ranking in my factorial ANOVA, but that I should ignore the interaction term?

Hi Diana,

I'm not familiar with ordinal regression, do you think it would be a good fit for my purposes? I am using an experimental design, looking at the effect of different instruction conditions on comprehension.

Many thanks,

Ben.

On 21 August 2011 07:04, Bruce Weaver <[hidden email]> wrote:

Rank the entire variable.  Re things to be aware of, see this:

http://en.wikipedia.org/wiki/ANOVA_on_ranks#Failure_of_ranking_in_the_factorial_ANOVA_and_other_complex_layouts

Benjamin Spivak (Med) wrote:
>
> Thank you all so much for your help so far.
>
> I'm considering performing a rank transformation and then running separate
> ANOVAs for each DV. Is there anything I ought to be aware of? Also, When I
> am transforming to rank, do I rank the entire variable or do I rank within
> each condition?
>
> Thanks in advance,
>
> Ben.
>
> On 20 August 2011 22:35, Bruce Weaver <[hidden email]>
> wrote:
>
>> I'll add one comment to what Rich has said:  The assumption of normality
>> for
>> OLS models applies to the errors (which are estimated with the
>> residuals--see link below), and normality of the errors is considerably
>> less
>> important than independence of the errors.  What this suggests is that
>> you
>> fit the model and then examine the residuals via residual plots etc.
>>
>>   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics
>>
>> HTH.
>>
>>
>> Rich Ulrich-2 wrote:
>> >
>> > - see below -
>> >
>> > Date: Fri, 19 Aug 2011 17:07:52 +1000
>> > From: [hidden email]
>> > Subject: Transforming variables
>> > To: [hidden email]
>> >
>> >> Hello everybody,
>> >> I would like to get people's general opinion on transforming
>> variables.
>> >> Is it a bad thing?, does it complicate interpretation?
>> > Yes, transformation can complicate both presentation and
>> interpretation.
>> >
>> > So, if you are dealing with a single use of a particular variable, and
>> it
>> > does *not* make any important difference in the modeling or tests,
>> > than using a transformation can be a bad thing because it distracts
>> > from the discussion.
>> >
>> > However, if you are using the variable often, or the untransformed
>> > version is damaging to the model or the tests, then you really ought
>> > to use a transformation.  (But what is "damaging"?)
>> >
>> >> I ask because I am performing a MANOVA for some jury comprehension
>> >> research. I have 3 DV's, application,
>> >> comprehension and recognition. One of my DV's is skewed and has
>> >> heterogenous variance. When I apply a
>> >> square root transformation, the problem is fixed. However, I'm worried
>> >> that interpretation may be difficult
>> >> with one transformed dv and two untransformed DV's. Any opinions would
>> be
>> >> much appreciated.
>> >
>> > Skewness and heterogeneous variance are useful warning signs that
>> > a natural measurement is not being taken in its most useful units, but
>> > that is not sufficient reason to transform.  Is the transformation a
>> > "natural" one for the variable?  - I eventually decided that looking at
>> > the scatter of Pre-Post, or any two replications, was useful for
>> deciding
>> > whether the measurement itself has heterogeneous error, or if
>> > heterogeneity was a feature of the sampling.  However, no matter what
>> > the variable itself looks like, linearity with the immediate Dependent
>> > Variable
>> > is more important than heterogeneity for constructing a good model,
>> even
>> > if that means that you need to do something special for the testing.
>> >
>> >  - Since you describe these as DVs, not IVs, it could be that you
>> *should*
>> > be concerned with the sampling error itself.  When I read
>> "comprehension
>> > and recognition", it occurs to me to wonder if these are both min-max-
>> > limited ranges, which might be properly scaled as 0-100% and
>> transformed
>> > to logit, if one wished to check out the "natural" scaling against each
>> > other,
>> > or against selected predictors.  Again - returning to my initial
>> comment
>> -
>> > that could distract from the discussion and modeling if it makes no
>> > difference
>> > for the data in hand.
>> >
>> > For your data on hand:  You do not mention the N.  Tests for skewness,
>> > etc., have too much power for large N, considering the robustness of
>> > least squares.  That is to say, your tests *might*  still be in fine
>> shape
>> > even without transforming.  So, how extreme is the situation?  Are the
>> > numbers skewed enough that you feel uncomfortable using the mean(s) as
>> > the description? (That's a pretty good sign that transformation is
>> > desirable.)
>> >
>> > Your square-root transformation is not as strong as taking log or
>> > reciprocal,
>> > though the effective strength also depends on the range being close to
>> > zero.
>> > My point here is that a weaker transform is less likely to matter to
>> the
>> > tests
>> > or model than a stronger one.  "Square root" is more likely to be
>> > ignorable
>> > than "reciprocal".
>> >
>> > The ultimate test of whether a transformation matters for your data is
>> > to try it both ways and see, for this analysis, "Does it matter?"  If
>> it
>> > doesn't matter, then you have little problem.  Present the analysis
>> that
>> > you want, and mention (for any subtle critics) that you also did it the
>> > other way and the results were the same.  If it does matter, then *I*
>> > would want to check on the face-value of the linearity of the DV  with
>> > the other variables in the model.
>> >
>> > --
>> > Rich Ulrich
>> >
>>
>>
>> -----
>> --
>> 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/Transforming-variables-tp4714751p4718352.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
>>
>


-----
--
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/Transforming-variables-tp4714751p4719200.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

[Benjamin Spivak (Med)]

Sorry to bother everyone yet again,

 

Given that my heterogeneity of variance seems to be an obstacle in conducting my rank-transformed two way ANOVA, would it be possible to use a brown-forsythe/welch correction to overcome this?

 

Many thanks,

 

Ben.

On 21 August 2011 10:51, Benjamin Spivak <[hidden email]> wrote:

Hi Diana,

I'm not familiar with ordinal regression, do you think it would be a good fit for my purposes? I am using an experimental design, looking at the effect of different instruction conditions on comprehension.

Many thanks,

Ben.

On 21 August 2011 07:04, Bruce Weaver <[hidden email]> wrote:

Rank the entire variable.  Re things to be aware of, see this:

http://en.wikipedia.org/wiki/ANOVA_on_ranks#Failure_of_ranking_in_the_factorial_ANOVA_and_other_complex_layouts

Benjamin Spivak (Med) wrote:
>
> Thank you all so much for your help so far.
>
> I'm considering performing a rank transformation and then running separate
> ANOVAs for each DV. Is there anything I ought to be aware of? Also, When I
> am transforming to rank, do I rank the entire variable or do I rank within
> each condition?
>
> Thanks in advance,
>
> Ben.
>
> On 20 August 2011 22:35, Bruce Weaver <[hidden email]>
> wrote:
>
>> I'll add one comment to what Rich has said:  The assumption of normality
>> for
>> OLS models applies to the errors (which are estimated with the
>> residuals--see link below), and normality of the errors is considerably
>> less
>> important than independence of the errors.  What this suggests is that
>> you
>> fit the model and then examine the residuals via residual plots etc.
>>
>>   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics
>>
>> HTH.
>>
>>
>> Rich Ulrich-2 wrote:
>> >
>> > - see below -
>> >
>> > Date: Fri, 19 Aug 2011 17:07:52 +1000
>> > From: [hidden email]
>> > Subject: Transforming variables
>> > To: [hidden email]
>> >
>> >> Hello everybody,
>> >> I would like to get people's general opinion on transforming
>> variables.
>> >> Is it a bad thing?, does it complicate interpretation?
>> > Yes, transformation can complicate both presentation and
>> interpretation.
>> >
>> > So, if you are dealing with a single use of a particular variable, and
>> it
>> > does *not* make any important difference in the modeling or tests,
>> > than using a transformation can be a bad thing because it distracts
>> > from the discussion.
>> >
>> > However, if you are using the variable often, or the untransformed
>> > version is damaging to the model or the tests, then you really ought
>> > to use a transformation.  (But what is "damaging"?)
>> >
>> >> I ask because I am performing a MANOVA for some jury comprehension
>> >> research. I have 3 DV's, application,
>> >> comprehension and recognition. One of my DV's is skewed and has
>> >> heterogenous variance. When I apply a
>> >> square root transformation, the problem is fixed. However, I'm worried
>> >> that interpretation may be difficult
>> >> with one transformed dv and two untransformed DV's. Any opinions would
>> be
>> >> much appreciated.
>> >
>> > Skewness and heterogeneous variance are useful warning signs that
>> > a natural measurement is not being taken in its most useful units, but
>> > that is not sufficient reason to transform.  Is the transformation a
>> > "natural" one for the variable?  - I eventually decided that looking at
>> > the scatter of Pre-Post, or any two replications, was useful for
>> deciding
>> > whether the measurement itself has heterogeneous error, or if
>> > heterogeneity was a feature of the sampling.  However, no matter what
>> > the variable itself looks like, linearity with the immediate Dependent
>> > Variable
>> > is more important than heterogeneity for constructing a good model,
>> even
>> > if that means that you need to do something special for the testing.
>> >
>> >  - Since you describe these as DVs, not IVs, it could be that you
>> *should*
>> > be concerned with the sampling error itself.  When I read
>> "comprehension
>> > and recognition", it occurs to me to wonder if these are both min-max-
>> > limited ranges, which might be properly scaled as 0-100% and
>> transformed
>> > to logit, if one wished to check out the "natural" scaling against each
>> > other,
>> > or against selected predictors.  Again - returning to my initial
>> comment
>> -
>> > that could distract from the discussion and modeling if it makes no
>> > difference
>> > for the data in hand.
>> >
>> > For your data on hand:  You do not mention the N.  Tests for skewness,
>> > etc., have too much power for large N, considering the robustness of
>> > least squares.  That is to say, your tests *might*  still be in fine
>> shape
>> > even without transforming.  So, how extreme is the situation?  Are the
>> > numbers skewed enough that you feel uncomfortable using the mean(s) as
>> > the description? (That's a pretty good sign that transformation is
>> > desirable.)
>> >
>> > Your square-root transformation is not as strong as taking log or
>> > reciprocal,
>> > though the effective strength also depends on the range being close to
>> > zero.
>> > My point here is that a weaker transform is less likely to matter to
>> the
>> > tests
>> > or model than a stronger one.  "Square root" is more likely to be
>> > ignorable
>> > than "reciprocal".
>> >
>> > The ultimate test of whether a transformation matters for your data is
>> > to try it both ways and see, for this analysis, "Does it matter?"  If
>> it
>> > doesn't matter, then you have little problem.  Present the analysis
>> that
>> > you want, and mention (for any subtle critics) that you also did it the
>> > other way and the results were the same.  If it does matter, then *I*
>> > would want to check on the face-value of the linearity of the DV  with
>> > the other variables in the model.
>> >
>> > --
>> > Rich Ulrich
>> >
>>
>>
>> -----
>> --
>> 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/Transforming-variables-tp4714751p4718352.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
>>
>


-----
--
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/Transforming-variables-tp4714751p4719200.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

[Ryan]

Consider the MIXED procedure:

*Generate data.
set seed 65923454.
new file.
inp pro.
  loop subject= 1 to 1000.
  compute group = rv.bernoulli(0.50).
  compute error = rv.normal(0,1).
  compute y = 1.5*(group=1) + 1.0*(group=0) +
sqrt(.50)*error*(group=1) + error*(group=0).
 end case.
 end loop.
end file.
end inp pro.
exe.

*Compute descriptives.
EXAMINE VARIABLES=y BY group
  /PLOT HISTOGRAM
  /STATISTICS DESCRIPTIVES
  /CINTERVAL 95
  /MISSING LISTWISE
  /NOTOTAL.

*Fit model assuming equal variances.
MIXED y BY group
 /FIXED=group | SSTYPE(3)
 /METHOD=REML
 /PRINT=R SOLUTION.

*Fit model allowing for heterogeneous variances.
MIXED y BY group
 /FIXED=group | SSTYPE(3)
 /METHOD=REML
 /PRINT=R SOLUTION
 /REPEATED=group | SUBJECT(subject) COVTYPE(DIAG).

Ryan

On Sun, Aug 21, 2011 at 12:07 AM, Benjamin Spivak
<[hidden email]> wrote:

> Sorry to bother everyone yet again,
>
> Given that my heterogeneity of variance seems to be an obstacle in
> conducting my rank-transformed two way ANOVA, would it be possible to use a
> brown-forsythe/welch correction to overcome this?
>
> Many thanks,
>
> Ben.
>
> On 21 August 2011 10:51, Benjamin Spivak <[hidden email]> wrote:
>>
>>>
>>> Hi Diana,
>>> I'm not familiar with ordinal regression, do you think it would be a good
>>> fit for my purposes? I am using an experimental design, looking at the
>>> effect of different instruction conditions on comprehension.
>>> Many thanks,
>>> Ben.
>>>
>>> On 21 August 2011 07:04, Bruce Weaver <[hidden email]> wrote:
>>>>
>>>> Rank the entire variable.  Re things to be aware of, see this:
>>>>
>>>>
>>>> http://en.wikipedia.org/wiki/ANOVA_on_ranks#Failure_of_ranking_in_the_factorial_ANOVA_and_other_complex_layouts
>>>>
>>>>
>>>>
>>>> Benjamin Spivak (Med) wrote:
>>>> >
>>>> > Thank you all so much for your help so far.
>>>> >
>>>> > I'm considering performing a rank transformation and then running
>>>> > separate
>>>> > ANOVAs for each DV. Is there anything I ought to be aware of? Also,
>>>> > When I
>>>> > am transforming to rank, do I rank the entire variable or do I rank
>>>> > within
>>>> > each condition?
>>>> >
>>>> > Thanks in advance,
>>>> >
>>>> > Ben.
>>>> >
>>>> > On 20 August 2011 22:35, Bruce Weaver <[hidden email]>
>>>> > wrote:
>>>> >
>>>> >> I'll add one comment to what Rich has said:  The assumption of
>>>> >> normality
>>>> >> for
>>>> >> OLS models applies to the errors (which are estimated with the
>>>> >> residuals--see link below), and normality of the errors is
>>>> >> considerably
>>>> >> less
>>>> >> important than independence of the errors.  What this suggests is
>>>> >> that
>>>> >> you
>>>> >> fit the model and then examine the residuals via residual plots etc.
>>>> >>
>>>> >>   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics
>>>> >>
>>>> >> HTH.
>>>> >>
>>>> >>
>>>> >> Rich Ulrich-2 wrote:
>>>> >> >
>>>> >> > - see below -
>>>> >> >
>>>> >> > Date: Fri, 19 Aug 2011 17:07:52 +1000
>>>> >> > From: [hidden email]
>>>> >> > Subject: Transforming variables
>>>> >> > To: [hidden email]
>>>> >> >
>>>> >> >> Hello everybody,
>>>> >> >> I would like to get people's general opinion on transforming
>>>> >> variables.
>>>> >> >> Is it a bad thing?, does it complicate interpretation?
>>>> >> > Yes, transformation can complicate both presentation and
>>>> >> interpretation.
>>>> >> >
>>>> >> > So, if you are dealing with a single use of a particular variable,
>>>> >> > and
>>>> >> it
>>>> >> > does *not* make any important difference in the modeling or tests,
>>>> >> > than using a transformation can be a bad thing because it distracts
>>>> >> > from the discussion.
>>>> >> >
>>>> >> > However, if you are using the variable often, or the untransformed
>>>> >> > version is damaging to the model or the tests, then you really
>>>> >> > ought
>>>> >> > to use a transformation.  (But what is "damaging"?)
>>>> >> >
>>>> >> >> I ask because I am performing a MANOVA for some jury comprehension
>>>> >> >> research. I have 3 DV's, application,
>>>> >> >> comprehension and recognition. One of my DV's is skewed and has
>>>> >> >> heterogenous variance. When I apply a
>>>> >> >> square root transformation, the problem is fixed. However, I'm
>>>> >> >> worried
>>>> >> >> that interpretation may be difficult
>>>> >> >> with one transformed dv and two untransformed DV's. Any opinions
>>>> >> >> would
>>>> >> be
>>>> >> >> much appreciated.
>>>> >> >
>>>> >> > Skewness and heterogeneous variance are useful warning signs that
>>>> >> > a natural measurement is not being taken in its most useful units,
>>>> >> > but
>>>> >> > that is not sufficient reason to transform.  Is the transformation
>>>> >> > a
>>>> >> > "natural" one for the variable?  - I eventually decided that
>>>> >> > looking at
>>>> >> > the scatter of Pre-Post, or any two replications, was useful for
>>>> >> deciding
>>>> >> > whether the measurement itself has heterogeneous error, or if
>>>> >> > heterogeneity was a feature of the sampling.  However, no matter
>>>> >> > what
>>>> >> > the variable itself looks like, linearity with the immediate
>>>> >> > Dependent
>>>> >> > Variable
>>>> >> > is more important than heterogeneity for constructing a good model,
>>>> >> even
>>>> >> > if that means that you need to do something special for the
>>>> >> > testing.
>>>> >> >
>>>> >> >  - Since you describe these as DVs, not IVs, it could be that you
>>>> >> *should*
>>>> >> > be concerned with the sampling error itself.  When I read
>>>> >> "comprehension
>>>> >> > and recognition", it occurs to me to wonder if these are both
>>>> >> > min-max-
>>>> >> > limited ranges, which might be properly scaled as 0-100% and
>>>> >> transformed
>>>> >> > to logit, if one wished to check out the "natural" scaling against
>>>> >> > each
>>>> >> > other,
>>>> >> > or against selected predictors.  Again - returning to my initial
>>>> >> comment
>>>> >> -
>>>> >> > that could distract from the discussion and modeling if it makes no
>>>> >> > difference
>>>> >> > for the data in hand.
>>>> >> >
>>>> >> > For your data on hand:  You do not mention the N.  Tests for
>>>> >> > skewness,
>>>> >> > etc., have too much power for large N, considering the robustness
>>>> >> > of
>>>> >> > least squares.  That is to say, your tests *might*  still be in
>>>> >> > fine
>>>> >> shape
>>>> >> > even without transforming.  So, how extreme is the situation?  Are
>>>> >> > the
>>>> >> > numbers skewed enough that you feel uncomfortable using the mean(s)
>>>> >> > as
>>>> >> > the description? (That's a pretty good sign that transformation is
>>>> >> > desirable.)
>>>> >> >
>>>> >> > Your square-root transformation is not as strong as taking log or
>>>> >> > reciprocal,
>>>> >> > though the effective strength also depends on the range being close
>>>> >> > to
>>>> >> > zero.
>>>> >> > My point here is that a weaker transform is less likely to matter
>>>> >> > to
>>>> >> the
>>>> >> > tests
>>>> >> > or model than a stronger one.  "Square root" is more likely to be
>>>> >> > ignorable
>>>> >> > than "reciprocal".
>>>> >> >
>>>> >> > The ultimate test of whether a transformation matters for your data
>>>> >> > is
>>>> >> > to try it both ways and see, for this analysis, "Does it matter?"
>>>> >> >  If
>>>> >> it
>>>> >> > doesn't matter, then you have little problem.  Present the analysis
>>>> >> that
>>>> >> > you want, and mention (for any subtle critics) that you also did it
>>>> >> > the
>>>> >> > other way and the results were the same.  If it does matter, then
>>>> >> > *I*
>>>> >> > would want to check on the face-value of the linearity of the DV
>>>> >> >  with
>>>> >> > the other variables in the model.
>>>> >> >
>>>> >> > --
>>>> >> > Rich Ulrich
>>>> >> >
>>>> >>
>>>> >>
>>>> >> -----
>>>> >> --
>>>> >> 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/Transforming-variables-tp4714751p4718352.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
>>>> >>
>>>> >
>>>>
>>>>
>>>> -----
>>>> --
>>>> 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/Transforming-variables-tp4714751p4719200.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

[Ryan]

Run the code and take a look at the output for the Residual Covariance
(R) Matrix. The off-diagonal elements indicate that the residual
co-variance between groups is zero. However, the group-specific
residual variances are estimated separately. Moreover, I did not
parameterize the model to account for any correlation among
observations in the same group (i.e. random intercept for group).

Really, all I did was fit a model for which the assumption of equal
variances is relaxed.

Ryan

On Sun, Aug 21, 2011 at 12:40 AM, Benjamin Spivak
<[hidden email]> wrote:

> Hi Ryan,
>
> Thanks very much for your syntax.
>
> Will the mixed procedure still make theoretical sense?, considering that
> there is no sig. intraclass correlation b/w my groups (juries).
>
> Cheers,
>
> Ben.
> On 21 August 2011 14:23, R B <[hidden email]> wrote:
>>
>> Consider the MIXED procedure:
>>
>> *Generate data.
>> set seed 65923454.
>> new file.
>> inp pro.
>>  loop subject= 1 to 1000.
>>  compute group = rv.bernoulli(0.50).
>>  compute error = rv.normal(0,1).
>>  compute y = 1.5*(group=1) + 1.0*(group=0) +
>> sqrt(.50)*error*(group=1) + error*(group=0).
>>  end case.
>>  end loop.
>> end file.
>> end inp pro.
>> exe.
>>
>> *Compute descriptives.
>> EXAMINE VARIABLES=y BY group
>>  /PLOT HISTOGRAM
>>  /STATISTICS DESCRIPTIVES
>>  /CINTERVAL 95
>>  /MISSING LISTWISE
>>  /NOTOTAL.
>>
>> *Fit model assuming equal variances.
>> MIXED y BY group
>>  /FIXED=group | SSTYPE(3)
>>  /METHOD=REML
>>  /PRINT=R SOLUTION.
>>
>> *Fit model allowing for heterogeneous variances.
>> MIXED y BY group
>>  /FIXED=group | SSTYPE(3)
>>  /METHOD=REML
>>  /PRINT=R SOLUTION
>>  /REPEATED=group | SUBJECT(subject) COVTYPE(DIAG).
>>
>> Ryan
>>
>> On Sun, Aug 21, 2011 at 12:07 AM, Benjamin Spivak
>> <[hidden email]> wrote:
>> > Sorry to bother everyone yet again,
>> >
>> > Given that my heterogeneity of variance seems to be an obstacle in
>> > conducting my rank-transformed two way ANOVA, would it be possible to
>> > use a
>> > brown-forsythe/welch correction to overcome this?
>> >
>> > Many thanks,
>> >
>> > Ben.
>> >
>> > On 21 August 2011 10:51, Benjamin Spivak <[hidden email]>
>> > wrote:
>> >>
>> >>>
>> >>> Hi Diana,
>> >>> I'm not familiar with ordinal regression, do you think it would be a
>> >>> good
>> >>> fit for my purposes? I am using an experimental design, looking at the
>> >>> effect of different instruction conditions on comprehension.
>> >>> Many thanks,
>> >>> Ben.
>> >>>
>> >>> On 21 August 2011 07:04, Bruce Weaver <[hidden email]>
>> >>> wrote:
>> >>>>
>> >>>> Rank the entire variable.  Re things to be aware of, see this:
>> >>>>
>> >>>>
>> >>>>
>> >>>> http://en.wikipedia.org/wiki/ANOVA_on_ranks#Failure_of_ranking_in_the_factorial_ANOVA_and_other_complex_layouts
>> >>>>
>> >>>>
>> >>>>
>> >>>> Benjamin Spivak (Med) wrote:
>> >>>> >
>> >>>> > Thank you all so much for your help so far.
>> >>>> >
>> >>>> > I'm considering performing a rank transformation and then running
>> >>>> > separate
>> >>>> > ANOVAs for each DV. Is there anything I ought to be aware of? Also,
>> >>>> > When I
>> >>>> > am transforming to rank, do I rank the entire variable or do I rank
>> >>>> > within
>> >>>> > each condition?
>> >>>> >
>> >>>> > Thanks in advance,
>> >>>> >
>> >>>> > Ben.
>> >>>> >
>> >>>> > On 20 August 2011 22:35, Bruce Weaver <[hidden email]>
>> >>>> > wrote:
>> >>>> >
>> >>>> >> I'll add one comment to what Rich has said:  The assumption of
>> >>>> >> normality
>> >>>> >> for
>> >>>> >> OLS models applies to the errors (which are estimated with the
>> >>>> >> residuals--see link below), and normality of the errors is
>> >>>> >> considerably
>> >>>> >> less
>> >>>> >> important than independence of the errors.  What this suggests is
>> >>>> >> that
>> >>>> >> you
>> >>>> >> fit the model and then examine the residuals via residual plots
>> >>>> >> etc.
>> >>>> >>
>> >>>> >>   http://en.wikipedia.org/wiki/Errors_and_residuals_in_statistics
>> >>>> >>
>> >>>> >> HTH.
>> >>>> >>
>> >>>> >>
>> >>>> >> Rich Ulrich-2 wrote:
>> >>>> >> >
>> >>>> >> > - see below -
>> >>>> >> >
>> >>>> >> > Date: Fri, 19 Aug 2011 17:07:52 +1000
>> >>>> >> > From: [hidden email]
>> >>>> >> > Subject: Transforming variables
>> >>>> >> > To: [hidden email]
>> >>>> >> >
>> >>>> >> >> Hello everybody,
>> >>>> >> >> I would like to get people's general opinion on transforming
>> >>>> >> variables.
>> >>>> >> >> Is it a bad thing?, does it complicate interpretation?
>> >>>> >> > Yes, transformation can complicate both presentation and
>> >>>> >> interpretation.
>> >>>> >> >
>> >>>> >> > So, if you are dealing with a single use of a particular
>> >>>> >> > variable,
>> >>>> >> > and
>> >>>> >> it
>> >>>> >> > does *not* make any important difference in the modeling or
>> >>>> >> > tests,
>> >>>> >> > than using a transformation can be a bad thing because it
>> >>>> >> > distracts
>> >>>> >> > from the discussion.
>> >>>> >> >
>> >>>> >> > However, if you are using the variable often, or the
>> >>>> >> > untransformed
>> >>>> >> > version is damaging to the model or the tests, then you really
>> >>>> >> > ought
>> >>>> >> > to use a transformation.  (But what is "damaging"?)
>> >>>> >> >
>> >>>> >> >> I ask because I am performing a MANOVA for some jury
>> >>>> >> >> comprehension
>> >>>> >> >> research. I have 3 DV's, application,
>> >>>> >> >> comprehension and recognition. One of my DV's is skewed and has
>> >>>> >> >> heterogenous variance. When I apply a
>> >>>> >> >> square root transformation, the problem is fixed. However, I'm
>> >>>> >> >> worried
>> >>>> >> >> that interpretation may be difficult
>> >>>> >> >> with one transformed dv and two untransformed DV's. Any
>> >>>> >> >> opinions
>> >>>> >> >> would
>> >>>> >> be
>> >>>> >> >> much appreciated.
>> >>>> >> >
>> >>>> >> > Skewness and heterogeneous variance are useful warning signs
>> >>>> >> > that
>> >>>> >> > a natural measurement is not being taken in its most useful
>> >>>> >> > units,
>> >>>> >> > but
>> >>>> >> > that is not sufficient reason to transform.  Is the
>> >>>> >> > transformation
>> >>>> >> > a
>> >>>> >> > "natural" one for the variable?  - I eventually decided that
>> >>>> >> > looking at
>> >>>> >> > the scatter of Pre-Post, or any two replications, was useful for
>> >>>> >> deciding
>> >>>> >> > whether the measurement itself has heterogeneous error, or if
>> >>>> >> > heterogeneity was a feature of the sampling.  However, no matter
>> >>>> >> > what
>> >>>> >> > the variable itself looks like, linearity with the immediate
>> >>>> >> > Dependent
>> >>>> >> > Variable
>> >>>> >> > is more important than heterogeneity for constructing a good
>> >>>> >> > model,
>> >>>> >> even
>> >>>> >> > if that means that you need to do something special for the
>> >>>> >> > testing.
>> >>>> >> >
>> >>>> >> >  - Since you describe these as DVs, not IVs, it could be that
>> >>>> >> > you
>> >>>> >> *should*
>> >>>> >> > be concerned with the sampling error itself.  When I read
>> >>>> >> "comprehension
>> >>>> >> > and recognition", it occurs to me to wonder if these are both
>> >>>> >> > min-max-
>> >>>> >> > limited ranges, which might be properly scaled as 0-100% and
>> >>>> >> transformed
>> >>>> >> > to logit, if one wished to check out the "natural" scaling
>> >>>> >> > against
>> >>>> >> > each
>> >>>> >> > other,
>> >>>> >> > or against selected predictors.  Again - returning to my initial
>> >>>> >> comment
>> >>>> >> -
>> >>>> >> > that could distract from the discussion and modeling if it makes
>> >>>> >> > no
>> >>>> >> > difference
>> >>>> >> > for the data in hand.
>> >>>> >> >
>> >>>> >> > For your data on hand:  You do not mention the N.  Tests for
>> >>>> >> > skewness,
>> >>>> >> > etc., have too much power for large N, considering the
>> >>>> >> > robustness
>> >>>> >> > of
>> >>>> >> > least squares.  That is to say, your tests *might*  still be in
>> >>>> >> > fine
>> >>>> >> shape
>> >>>> >> > even without transforming.  So, how extreme is the situation?
>> >>>> >> >  Are
>> >>>> >> > the
>> >>>> >> > numbers skewed enough that you feel uncomfortable using the
>> >>>> >> > mean(s)
>> >>>> >> > as
>> >>>> >> > the description? (That's a pretty good sign that transformation
>> >>>> >> > is
>> >>>> >> > desirable.)
>> >>>> >> >
>> >>>> >> > Your square-root transformation is not as strong as taking log
>> >>>> >> > or
>> >>>> >> > reciprocal,
>> >>>> >> > though the effective strength also depends on the range being
>> >>>> >> > close
>> >>>> >> > to
>> >>>> >> > zero.
>> >>>> >> > My point here is that a weaker transform is less likely to
>> >>>> >> > matter
>> >>>> >> > to
>> >>>> >> the
>> >>>> >> > tests
>> >>>> >> > or model than a stronger one.  "Square root" is more likely to
>> >>>> >> > be
>> >>>> >> > ignorable
>> >>>> >> > than "reciprocal".
>> >>>> >> >
>> >>>> >> > The ultimate test of whether a transformation matters for your
>> >>>> >> > data
>> >>>> >> > is
>> >>>> >> > to try it both ways and see, for this analysis, "Does it
>> >>>> >> > matter?"
>> >>>> >> >  If
>> >>>> >> it
>> >>>> >> > doesn't matter, then you have little problem.  Present the
>> >>>> >> > analysis
>> >>>> >> that
>> >>>> >> > you want, and mention (for any subtle critics) that you also did
>> >>>> >> > it
>> >>>> >> > the
>> >>>> >> > other way and the results were the same.  If it does matter,
>> >>>> >> > then
>> >>>> >> > *I*
>> >>>> >> > would want to check on the face-value of the linearity of the DV
>> >>>> >> >  with
>> >>>> >> > the other variables in the model.
>> >>>> >> >
>> >>>> >> > --
>> >>>> >> > Rich Ulrich
>> >>>> >> >
>> >>>> >>
>> >>>> >>
>> >>>> >> -----
>> >>>> >> --
>> >>>> >> 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/Transforming-variables-tp4714751p4718352.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
>> >>>> >>
>> >>>> >
>> >>>>
>> >>>>
>> >>>> -----
>> >>>> --
>> >>>> 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/Transforming-variables-tp4714751p4719200.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
>
>

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