Thanks for your detailed reply.
Based on your following comment:-
> Hmmm. Well, PROC MIXED *might* be the right tool for this. It might not.
> You should keep an open mind. As should your boss.
and this comment :-
> Don't go off on an HLM tangent. It *might* be relevant. It quite likely
> is not. HLM is often more appropriate when you have data collected at
> multiple levels.
Based on above 2. I have a question. With respect to this project my
Boss recently bought HB-Reg (Hierarchical Bayes regression software)
from Sawtooth and the logic for using it was that data is sparse and
this software would help in getting better estimates by pooling
information from a marketetc. Now I understand that I should not go off
towards HLM tangent, but is there a relation between hierarchical in
HLM and hierchical in HB-reg?
>This is really saying something about the universe
> you wish to model, rather than something about random variability. If you
> have 50 values of MARKET and these are all the values that MARKET can have (or
> they are all the values of MARKET you are going to want to work with and you
> will not extrapolate to other markets) then you have a fixed effect
Thanks for this. I think am able to appreciate why a random statement
might be used. One SAS'er mentioned about "statistical space" and that
too made lot of sense. I also went through Fixed Vs random effects
section within varcomp and it cleared up.
> So tell me. Based on this, do you think that MARKET is a random
You have got me in to a .... by asking this question(!!!). Frankly
speaking I dont know whether the goal is to estimate for Markets other
than what I already have in my data (Information not forthcoming from
boss!!). But yes, while using Proc Mixed I was asked to put store
merchandizing conditions within random statement (Boolean variables for
whether a product was on Display or not, simialarly whether it was on
feature or not and lastly whether product was on temporary price
reduction or not). Nowhere did we venture anywhere near putting the
Market variable in random statement. On the other hand while ffeding
this same data to HB-reg software the ID was generated at market level.
and what I could understand was that differentParameter estimates were
generated for each variable within each market for a particular
product. Does this mean that in Proc Mixed also we should have probably
tried putting Market in Random Statement? or within proc Mixed what
vraiable (or combination of variables) should have been in order to fit
the same model as in HB-Reg?
> If this is advertising dollars or advertising effort, then I would say that
> data *should* be zeroes instead of missing values. What are your
> "Advertising related variables" ? Does it make sense for these values to
> be set to zero at these spots?
> Replacing missing values is a variable-specific issue that depends on the
> meaning of the data, not on statistical stuff.
The above has been quite helpful to me. I learn now that since the
variables represent effort (or returns like TV GRP's etc) so it makes
sense to probably replace with zero rather than missing.
PS: Im going to post another thread in this topic which is probably one
more fundamental block missing from my understanding of Mixed (I agree
that this is a "mixed" up statement!).