Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.

# Re: st: ladder question for right-skewed variable

 From Nick Cox To "statalist@hsphsun2.harvard.edu" Subject Re: st: ladder question for right-skewed variable Date Fri, 26 Apr 2013 19:45:46 +0100

```Three assertions based on a mix of experience and prejudice:

1. The best way to check for normality is with -qnorm-. Even if
normality is not your reference case, asymmetry will show up clearly
on a -qnorm- graph.

2. 90% of the time, choosing transformations boils down to whether
three possible transformations are any use, root, logarithm or
reciprocal.

3. So, do-it-yourself is easy:

gen rtmyvar = sqrt(myvar)
gen logmyvar = log(myvar)
gen recmyvar = 1/myvar

qnorm myvar, name(a)
qnorm rtmyvar, name(b)
qnorm logmyvar, name(c)
qnorm recmyvar, name(d)

Not universally known fact: Giving a name to a graph means that it
sticks around until _you_ close it. So, you have four graphs on your
monitor. Arrange them with your mouse so you can compare. Usually it's
easy to pick what works best, without any formal machinery.

Nick
njcoxstata@gmail.com

On 26 April 2013 19:20, Nick Cox <njcoxstata@gmail.com> wrote:
> Just to underline that kurtosis in your variable was calculated by
> -summarize- 108. That's BIG. No wonder -sktest- can't cope.
> Nick
> njcoxstata@gmail.com
>
>
> On 26 April 2013 19:17, Nick Cox <njcoxstata@gmail.com> wrote:
>> That's not quite "no transformations appeared in the output" as
>> -ladder- is signalling P-values for some cases.
>>
>> But I readily agree that -ladder- is not doing a good job here at all.
>>
>> In fact, I am now reminded of evident -ladder- problems shown in a
>> http://www.stata.com/statalist/archive/2013-02/msg00862.html
>>
>> I can't find a public email, even though I thought I posted on this,
>> but my impression from looking at the code is that -ladder- is
>> essentially fragile. The real problem here is within -sktest-. It can
>> break down, it seems, for large sample sizes and/or large deviations
>> from Gaussianity. Then it bounces back missings.
>>
>> I think you just need to abandon -ladder-. It's not essential. You
>> don't need _any_ test to tell you that some transformation will help
>> if the goal is to reduce asymmetry, and there are only a few credible
>> alternatives.
>>
>> As David and I pointed out, log transformation should work quite well
>>
>> but but but: (my suggestion; David may not agree) why transform at
>> -nbreg-).
>>
>> BTW, -ladder- is a command, not a function, and in Stata ne'er the
>> twain shall meet.
>>
>> Nick
>> njcoxstata@gmail.com
>>
>>
>> On 26 April 2013 18:55, Gabriel Nelson <lgabrielnelson@gmail.com> wrote:
>>> Thanks Nick, yes exactly, my question is why the ladder function fails
>>> to provide any chi-square values here. I'll attach the Stata output
>>> here:
>>>
>>>
>>> Transformation         formula               chi2(2)       P(chi2)
>>> ------------------------------------------------------------------
>>> cubic                  dis~2000^3                 .            .
>>> square                 dis~2000^2                 .            .
>>> identity               dis~2000                   .            .
>>> square root            sqrt(dis~2000)             .        0.000
>>> log                    log(dis~2000)              .        0.000
>>> 1/(square root)        1/sqrt(dis~2000)           .        0.000
>>> inverse                1/dis~2000                 .        0.000
>>> 1/square               1/(dis~2000^2)             .        0.000
>>> 1/cubic                1/(dis~2000^3)             .        0.000
>>>
>>> . sum disp_2000, detail
>>>
>>>       Number displaced 2000 (if data unavailable go up
>>>                            to 2003
>>> -------------------------------------------------------------
>>>       Percentiles      Smallest
>>>  1%            1              1
>>>  5%            2              1
>>> 10%            3              1       Obs                1010
>>> 25%            6              1       Sum of Wgt.        1010
>>>
>>> 50%         15.5                      Mean           281.5297
>>>                         Largest       Std. Dev.      1217.168
>>> 75%           82           9421
>>> 90%        436.5           9505       Variance        1481497
>>> 95%         1251          16255       Skewness       9.012044
>>> 99%         5953          19569       Kurtosis       108.8061
>>>
>>> On Fri, Apr 26, 2013 at 10:47 AM, Nick Cox <njcoxstata@gmail.com> wrote: