st: Re: mvsumm and missing obs

[Kit Baum <baum@bc.edu>]

I agree; I'm not thrilled about the notion of adding this flexibility 
to mvsumm. One would then have to compute, for each statistic and each 
point in time, the number of data points that went into it, and would 
have to be able to specify what a minimum amount of information might 
be to calculate anything at all. Neither of those are technically 
daunting, but the concept of a moving statistic based upon a variable 
amount of information should be treated with extreme caution (e.g. say 
that I am calculating 24-month averages, but for some periods only my 
minimum of 4 data points are available. The standard errors of those 
means would vary radically). Users are free to create 'mvsumm2', 
'mvsumm3', etc. but I am not keen on this flexibility for mvsumm or 
mvcorr.

A solution not involving programming would be to generate, e.g., mvsumm 
for 5 obs. as one variable, mvsumm for 4 as another, and mvsumm for 3 
as a third, and then use simple logic to define your result as that 
based on the largest sample for which the statistic is non-missing 
(taking 3 as the minimum). Imagine that stat5, stat4, stat3 in the 
table are mvsumm results for 5, 4 and 3 obs respectively. Then

. g mystat = cond(stat5<.,stat5,cond(stat4<.,stat4,stat3))

. l

     +--------------------------------+
     | stat3   stat4   stat5   mystat |
     |--------------------------------|
  1. |    10      16       .       16 |
  2. |    11       .       .       11 |
  3. |    12      13      14       14 |
     +--------------------------------+

Kit

On Aug 10, 2004, at 12:18 PM, Nick Cox wrote:


> I think you need to clone -mvsumm- and
> get your clone, called something else,
> to do what you want. It's not difficult,
> but speaking as a co-author of -mvsumm-
> you don't have my support in this aim.
> Insisting on complete windows is, I suggest,
> to be regarded as a small feature
> in -mvsumm-. You can nobble the code if you wish,
> but that is your responsibility.
>
> Kit Baum may or may not agree with me on this.
>
> Nick
> n.j.cox@durham.ac.uk
>
> jansoko@umich.edu
> > I have a similar question to Adrian's about -mvsumm-.
> > Unfortunately, the
> > proposed solution (using -tssmooth-) doesn't work as I am
> > interested in moving
> > variances, not moving averages.
> >
> > How can I get create a variable that contains the variance of
> > the five preceding
> > observations in a panel data set, even if one or two of the
> > those preceding
> > observations are missing?  And how can I specify the minimum number of
> > non-missing observations within the window?
*
*   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/