SVY linear regression

[Katarina Vlcakova]

Hello everybody, I have a quite well working regression model (used STATA) with using sample weights (prefix svy). Could anyone help me with the postestimation, verification of the model (E.g. heteroskedasticity test, autocorrelation test, normality test). I am new to stata and survey analysis, so I would be glad for any help with this. Please respect, that I use stata only a ew months... thank you

[Steve Samuels]

You haven't used "STATA", Katarina, but you have used "Stata". See the FAQ, Section 18.


The assumptions of homoskedasticity, normality, and independence (e.g. no auto-correlation) are all assumptions needed for ordinary least squares. They are not needed for svy: regress, which bases its inference on the sample design (stratification and variation between primary sampling units) and is robust to violations of those assumptions. In other words, there's no need to test for those violations.

What you do want to test is that your model is correct. The basic test of fit is linktest, but you can also augment your model with non-linear ( (e.g. square) terms and interactions, then test that the added terms are all zero.

[Marcos Almeida]

Hello Katarina,

I wonder if you already know this, but maybe not, since you said you are "new to Stata".

So, apart from the useful tips Steve Samuels gave you, I suggest you take a look on some alternatives of postestimation after - svy:regress -

http://www.stata.com/manuals13/svysv...postestimation

Also, you can get this information directly from your Stata software just by typing - help svy postestimation -

Best,

Marcos

[Katarina Vlcakova]

Thank you Steve Samuels. I found in forums, something similar, that you wrote... Could you please reccoment some materials or papers to backup my research? Thank you

[Richard Williams]

UCLA (lower third of page) has some FAQs on SVY analysis that may be helpful:

http://www.ats.ucla.edu/stat/stata/faq/

Stata's own FAQ is at

http://www.stata.com/support/faqs/statistics/#survey

I like Steve's points. I guess I more or less already knew that but he states it very well. I wish the manual or a FAQ would elaborate a bit more on issues like this, e.g. diagnostic tests. I may just quote Steve in my next set of notes on this.

With some diagnostics, I also think it is not so bad just to run non-svy commands, e.g. do so to identify outliers. You are often just trying to identify problems and figure out how to solve them -- it isn't like your statistics have to be accurate to 3 decimal places.

[Steve Samuels]

See Rich Valliant's talk: “Regression Diagnostics for Complex Survey Data”, Stata Users’ Conference. Washington DC, 2009.
http://www.stata.com/meeting/dcconf09/dc09_valliant.pdf.

For identifying outliers, it is important to start with a method that is resistant to outliers (slide 24). Unfortunately least squares is the method least resistant. In Stata, qreg will do median regression and will accept pweights in Stata 13.

I

[Katarina Vlcakova]

Thank you so much for your help

[Katarina Vlcakova]

Could I have just one quick question? In the linktest, _hatsq was not significant (P-value 0,474), but is a p-value 0,810 in case of _hat OK, Is there no misspecification? Ovtest shows no omitted vars (0.4578). Thank you

[Steve Samuels]

Katarina, the p-value for _hat is not important in judging the fit of the model.

[Dean Smith]

"The assumptions of homoskedasticity, normality, and independence (e.g. no auto-correlation) are all assumptions needed for ordinary least squares. They are notneeded for svy: regress, which bases its inference on the sample design (stratification and variation between primary sampling units) and is robust to violations of those assumptions. In other words, there's no need to test for those violations."

Could you provide the reference for this: That svy: regress (or svy: glm which is what I'm using) is robust to homoskedasticity, normality, and autocorrelation ? Thanks!