Issue: re-normalize fixed-effects after regressions. Q: Any way to speed up the "contrast" command?

[Lihan Liu]

Dear Stata users:

I use “contrast” to normalize fixed-effects to grand mean and compute corresponding standard errors after “regress” command. After 13 days, no results come up yet. Would really appreciate if you can help me out. Thank you in advance for your help. Below are more details,

Task: normalize fixed-effects to grand mean and compute corresponding standard errors
Simplified Model: Y=aX+bW+u, where W represents a set of dummies
Data size: 205 MB
Commands:

1. stage 1: obtain fixed-effects
   1. codes: regress y X i.W [aw=weights]
   2. it takes about 2 hours
   3. there are about 7000 dummies (W)
2. stage 2: normalize fixed-effects to grand mean
   1. codes: parmby "contrast gw.W, post"
   2. Not results after 13 days

Questions
1. Is there any way to speed up the "contrast" command?
2. felsdvregdm might be faster than contrast. But my understanding is it does not allow for weights. So it will not work here.
3. Should I consider any other commands besides "contrast" or " felsdvregdm"?
4. I have dropped unused data, unnecessary variables, compress the data set, and reduce maximum matrix size and variable size as much as possible. In general, is there any other way to speed up computation?

Version of stata: StataMP 14 (64-bit)

Any thought would be highly appreciated!

Lihan

[Carlo Lazzaro]

Lihan:
have you already taken a look at -areg-?

[Maarten Buis]

Here is one way without standard errors :

// open some example data
sysuse nlsw88, clear

// this is what you did
reg wage hours i.occupation
contrast gw.occupation

// alternative
xtreg  wage hours, i(occupation) fe
predict u, u
table occupation, c(mean u)

[Lihan Liu]

Lihan:
have you already taken a look at -areg-?

Dear Carlo, thank you for your reply. I did considered "areg". However, contrast does not work after "areg".

For example:

// open some example data
sysuse nlsw88, clear

// use areg
areg wage hours, absorb(occupation)
contrast gw.occupation
stata erros: variable occupation not found

[Lihan Liu]

Here is one way without standard errors :

// open some example data
sysuse nlsw88, clear

// this is what you did
reg wage hours i.occupation
contrast gw.occupation

// alternative
xtreg wage hours, i(occupation) fe
predict u, u
table occupation, c(mean u)

Dear Maarten, thank you very much for your suggestion. It is great to know! I do need standard errors. Plus xtreg does not work in my case because I have weights which vary within group.

[Lihan Liu]

Update: I wrote my own codes to compute point estimates and standard errors. It is much faster than "contrast" command. Seems the "contrast" command computes a lot more things, which I don't need. In one experiment, my codes take 7min. However, the contrast command takes 1Hour and 20 min.

The problem is solved. Thank everyone for your help!

[Lihan Liu]

FYI, specify "nowald" and "noeffect" did speed up "contrast". But still it is slow.

[paulvonhippel]

Lihan Liu : Can you share your code for calculating grand weighted mean contrasts with standard errors?
I agree that contrast gw. is very slow.