If you don't have a "natural zero" then you should not be
considering the log as an early choice for transformation.
(If you do have a natural zero, *why* do your scores run
from minus 8 to 10?) The best guide to the appropriate
transformation, in my experience, is to consider how the
data have been generated.
If you do have a large N, then there is a fair chance that
you don't have any need for a transformation, anyway.
What makes "testing for normality" an optional sort of
procedure is that it *matters most* when the N is small,
and that is when the power of the test for normality is (also)
small. When the N is large enough, a test for normality will
usually "reject" because there are few distributions that are
truly normal. - Since you characterize your data as having
integer limits, we might assume that it is not Normal for
either of two reasons -- (a) Normal distributions have infinite
range, and (b) normal distributions are continuous (not integers).
Finally, what matters for test purposes is that the *residual*
or errors are sufficiently Normal so that the test statistic is
accurate. What throws off the test statistic, usually, is either
too many outliers or too many zeroes.
How were your data generated? If adding a constant does
not shift the scores to reflect a natural zero, then there's not
much to recommend it. Statisticians, as opposed to stat-pack
amateurs, are more apt to *replace* a zero with something
appropriate than to use an add-on. (I'm thinking, in particular,
of cases where measured drug or enzyme levels are reported
as zero owing to measurement insufficiency... and the proper
replacement score might be one-half the lowest detected level.)
--
Rich Ulrich
> Date: Sat, 29 Dec 2012 10:41:26 -0800
> From:
[hidden email]> Subject: data transformation for negative values to get normality
> To:
[hidden email]>
> Hi all,
>
> I am working with negative and positive values for a variable (range data
> beetwen -8.00 - 10.00), lamentably this data is nor normal according
> Kolmogorov smirnov test...
>
> I am trying to transform with log10 but the tipical transformation log10
> (x+1) is not possible because my lowest of negative values is -8.00, could i
> use the transformation: log10 (x+9) to dont get negative values??
>
...
Locked
