RE: st: RE: A modest proposal - missing data doesn't count

["Nick Cox" <n.j.cox@durham.ac.uk>]

Paradoxical though it may seem, Allan has got to produce a stronger
argument before it can judged exactly how good his proposal is. To that
extent I agree with David. Otherwise I remain sceptical, while
acknowledging fully that sceptical attitude differs from logical
debunking. 

Although it is difficult to demonstrate here, it should be underlined,
from e.g. discussions at users' meetings, that StataCorp have thought
long and hard about better solutions to this over many years. The
absence of a solution that satisfies all is not indicative of absence of
effort to find such a solution. 

Nick 
n.j.cox@durham.ac.uk 

David Bell

I am in principle supportive of this new -value- expression.  I get bit
periodically and the various workarounds are sometimes inelegant.  But
some of the counterarguments seem potentially fatal.  However,
assertions that a working solution _cannot_ exist have not yet been
established.

In the spirit of seeing whether there is a resolution to this discussion
of a potential -value- expression, I see two steps that are brought up
by the discussion:

(1) what is a proposed description of what -value()- should do, and
(2) what are the various special situations and combinations of special
situations, and how would the proposed deal with them?

Could there be a way that this new expression could correspond to
people's intuitive understandings in a way that, while having perhaps
its own issues, would bite less than the - >10 - problem bites?

On Jun 15, 2010, at 12:40 PM, Austin Nichols wrote:

> Martin--
> I agree.  In particular the treatment of p&q and p|q in the presence
> of one missing value should worry anyone who actually uses data; this
> is something that requires careful thought in any specific instance,
> depending on what one is trying to estimate.
> 
> Allan--
> Both x<10 and x>10 can be affected by presence of missings in x; think
of
> . gen dummy=(x<10)
> for example.  No solution is perfect, and the treatment in functions
> like max() and sum(), or egen functions like rowtotal() etc., logical
> expressions, algebra on variables, and so forth may be hard for some
> users to internalize, but so would another "solution" create
> difficulties, not always immediately obvious, and bound to trip up
> those of us who are used to the current behavior (but still get
> tripped up by the current behavior when we aren't careful).
> 
> On Tue, Jun 15, 2010 at 6:53 AM, Martin Weiss <martin.weiss1@gmx.de>
wrote:
> <snip>
>> I found the alternative approaches to missings presented in
>> http://www.stata.com/meeting/uk08/KIMacD.presentation.ppt wholly
>> unconvincing and worse than the current state.
> <snip>
>> -----Allan Reese-----
>> It's so easy to forget that x<10 will not be affected by missing
values
>> but x>10 will be.  It would be good to have software that guards
against
>> pitfalls rather than invites you in.
> <snip>

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/