### sas >> Log and other transforms (was How to back transform

XXXX@XXXXX.COM replied:
>
> >>> David L Cassell < XXXX@XXXXX.COM > 03/28/07 12:56 AM >>> wrote
><<<
>In my experience, log-transforming to correct for skewness was usually
>followed by disgruntled grumbling, as the log transform corrected too
>much or not enough, or ruined some other aspect of the underlying
>assumptions. I haven't had a real data set which called out for a log
>transform since the 1980's.
> >>>>
>
>
>I agree with David that blindly transforming things because 'that's what
>people do' is silly and likely to be
>counter-productive.
>
>I think a better approach is to figure out whether the transformation
>makes any substantive sense. If the scale is arbitrary and not
>well-known, and you are lucky enough that some transformation creates
>fewer problems than it solves, then it may be sensible to use the BoxCox
>transformation, even if the transformation is bizarre. OTOH, if the
>scale is known, or inherently sensible, then transformation is less
>likely to be the ideal solution, and, if a transformation is necessary,
>it should be one that preserves the sense of the variable.
>
>So......a colleague wanted to log transform a ratio. WAIT, I said.
>Does it make any substantive sense to change a ratio into a difference?
>(It did not). OTOH, if called for, an inverse transformation of a ratio
>can make sense (changing miles per gallon to gallons per mile)
>
>As always, I think the substantive question should drive the statistics,
>not vice versa
>
>
>HTH
>
>Peter

I agree.

And also, watch out for those Box-Cox lambdas. I had a project once
where some people wanted to use Box-Cox for transformations to
normality, and they were getting typical values like 2.3836 and 0.183733
and they weren't even thinking that the first was like a square and
the second was like a log transform. (Not that any of the transforms
were actually appropriate for what they wanted to do with the data, but
that's another issue.)

David
_________________________________________________________________
Interest Rates near 39yr lows! \$430,000 Mortgage for \$1,399/mo - Calculate
new payment
http://www.lowermybills.com/lre/index.jsp?sourceid=lmb-9632-18466&moid=7581

```>>> David L Cassell < XXXX@XXXXX.COM > 03/28/07 12:56 AM >>> wrote
<<<
In my experience, log-transforming to correct for skewness was usually
followed by disgruntled grumbling, as the log transform corrected too
much or not enough, or ruined some other aspect of the underlying
assumptions.  I haven't had a real data set which called out for a log
transform since the 1980's.
>>>>

I agree with David that blindly transforming things because 'that's what
people do' is silly and likely to be
counter-productive.

I think a better approach is to figure out whether the transformation
makes any substantive sense.  If the scale is arbitrary and not
well-known, and you are lucky enough that some transformation creates
fewer problems than it solves, then it may be sensible to use the BoxCox
transformation, even if the transformation is bizarre.   OTOH, if the
scale is known, or inherently sensible, then transformation is less
likely to be the ideal solution, and, if a transformation is necessary,
it should be one that preserves the sense of the variable.

So......a colleague wanted to log transform a ratio.  WAIT, I said.
Does it make any substantive sense to change a ratio into a difference?
(It did not).  OTOH, if called for, an inverse transformation of a ratio
can make sense (changing miles per gallon to gallons per mile)

As always, I think the substantive question should drive the statistics,
not vice versa

HTH

Peter
```

```Hi all,

I was wondering if anyone knows of a way to back transform a factor loading
matrix (with say, varimax rotation) to a correlation or covariance matrix.
Alternatively, can one input the pattern matrix into any of the SAS Procs to
output a corr or cov matrix?

I recently read two papers on a survey instrument that I know were
incorrectly analyzed. Unfortunately, the authors only published the factor
loading matrices. I would like to input these matrices into Proc Factor and
Proc Calis to reanalyze their conclusions. Does anyone know of a way to do this?

thanks,
Cristian
```

```Hi! I have some binomial disease data and have analysed it with probit
and logit link functions. I have been able to back transform the
logits but not the probits. Has anyone got code to back transform the
probits? That would be great if you do. Thanks.
```

```option1 implicate inference on transform world where your dependent
variable is not normally distributed but poisson distributed.

In case of option2 your dependent variable is continous but its scale
is adjusted & it is normally distributed.

Option1 follows the assumption of the genralized linear model whereas
option2 follows the assumption of  general linear model.

On 2/27/09, nuria < XXXX@XXXXX.COM > wrote:
> What is the difference between:
>
> option 1
> proc genmod;
> class treat;
> model count = treat / dist=poisson link=log;
> run;
>
> option2
> data dataset;
> set dataset;
> logcount= log (count);
>
> proc mixed;
> class treat;
> model logcount = treat;
> run;
>
> Thanks!
>
```