Marcello Pagano
> > As I have said before, I would very much like, with Allan
> Reese, an
> > option for equi-probability histograms. I think that this
> > is especially
> > useful when thinking of the histogram as an estimator of an
> > underlying
> > density function of a continuous variable. I do not see a strong
> > argument, other than tradition (laziness?), for being
> > constrained to
> > histograms with equi-spaced bins.
Nick Cox
> An interesting problem.
> I see one key question arising.
>
> If I understand this properly, we just need to find so many
> quantiles
> x_(i) equally spaced on the probability scale and calculate bar
> heights proportional to 1 / (x_(i+1) - x_(i)). In practice, it is
> not that uncommon to get ties, implying infinite bar heights. We
> can fudge this by interpolating linearly between different
> quantiles.
>
> What lies below is a hack. The number of bins is wired in at no
> more than 20. My experiments indicate that you need
> a lot of data to avoid somewhat bizarre details. Essentially
> you are bound to see the kind of instability inherent in crude
> numerical derivatives.
>
> Don't point out that density estimates are nicer! I know that.
> This is an attempt at what Marcello wants.
< code omitted >
Having slept on this,
1. I realised that the method of adjusting for tied
quantiles was arbitrary, biased, implemented incorrectly
even in its own terms and in any case not applicable
if quantiles tied with the maximum. In short, a bad
idea. After thinking about various alternatives,
I decided that the most honest thing to do
was to refuse to draw the graph if quantiles are tied.
2. I generalised to allow frequency and analytic
weights.
3. Most interestingly, Vince Wiggins kindly told me
of an undocumented option -bartype(spanning)-
which removes the need for the convoluted
trickery used to draw bars. Thanks! (It
appears to have one quirk, which is worked
around in my code.)
If you copied and pasted the code posted yesterday,
please junk it and use this instead. I will send a regular
version to SSC shortly.
Nick
[email protected]
program eqprhistogram, sortpreserve
*! for Marcello, eppur si muove!
*! NJC 2.0.0 31 May 2003
version 8
syntax varname(numeric) [if] [in] [aweight fweight] ///
[ , bin(numlist int >1 <=20) * ]
// #bins defaults to 20
if "`bin'" == "" local bin = 20
local binp1 = `bin' + 1
local binp2 = `bin' + 2
local binm1 = `bin' - 1
// enough data?
marksample touse
qui count if `touse'
if r(N) < `binp2' {
di as err "insufficient observations"
error 2000
}
tempvar quantile qnum density
qui {
// get quantiles
su `varlist' [`weight' `exp'] if `touse', meanonly
generate `quantile' = r(min) in 1
replace `quantile' = r(max) in `binp1'
_pctile `varlist' [`weight' `exp'] if `touse', nq(`bin')
forval i = 1/`binm1' {
replace `quantile' = r(r`i') in `= `i' + 1'
}
// check for tied quantiles
bysort `quantile' : ///
gen byte `qnum' = _N * (`quantile' < .)
su `qnum', meanonly
if r(max) > 1 {
noi di as txt ///
"{p}`varlist' has tied quantiles: " ///
"try fewer bins? graph inappropriate?{p_end}"
exit 0
}
// prepare graph
gen `density' = ///
1 / (`bin' * (`quantile'[_n+1] - `quantile'))
_crcslbl `quantile' `varlist'
label var `density' "Density"
// work-around spanning quirk
replace `quantile' = `quantile'[1] in `binp2'
replace `density' = 0 in `binp2'
sort `quantile' `density'
}
twoway bar `density' `quantile', ///
bartype(spanning) bstyle(histogram) `options'
end
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/
