{smcl}
{* 08jun2003}{...}
{hline}
help for {hi:wgttest}
{hline}

{title:Test the impact of sampling weights in regression analysis}

{p 8 15 2}
{cmd:wgttest} {it:depvar} [{it:varlist}] [{cmd:if} {it:exp}] [{cmd:in} {it:range}]
{cmd:,} {cmd:wgt(}{it:wgtvar}{cmd:)} [ {cmd:cmd(}{it:estimation_command}{cmd:)}
{cmdab:pre:fix(}string{cmd:)} {cmd:testopt(}{it:test_options}{cmd:)}
{cmd:nonoise} {it:estimation_options} ]

{p 4 4 2}
{cmd:by} {it:...} {cmd::} may be used with {cmd:wgttest}; see help {help by}.


{title:Description}

{p 4 4 2}
Whether to use sampling weights (pweights) in regression analysis should be carefully
evaluated. Often, the weights do not have great influence on the parameter estimates (see
e.g. Winship and Radbill, 1994, to learn when and why). In such cases, unweighted
estimates are preferable because they are more efficient than the weighted estimates.

{p 4 4 2}
{cmd:wgttest} performs a test proposed by DuMouchel and Duncan (1983) to evaluate the
significance of the impact of sampling weights on estimation results. First, a regression model
of {it:depvar} on {it:varlist} including {it:wgtvar} and the first order interactions between
{it:varlist} and {it:wgtvar} as additional covariates will be estimated. Second, the coefficients
of these covariates (i.e. {it:wgtvar} and the interactions) are tested against zero using a
standard F test (as provided by the post-estimation command {cmd:test}). "If the F test is not
significant, then the weighted and unweighted estimates are not significantly different and the
analyst can proceed by using unweighted OLS. Weighted and unweighted estimates are significantly
different if the F test is significant" (Winship and Radbill, 1994: 248). In the later case, the
unweighted estimators are probably biased by sample selection and the weighted estimators are
preferable. Be aware, however, that significant differences between weighted and unweighted
estimates may also be due to model misspecification.


{title:Options}

{p 4 8 2}
{cmd:cmd(}{it:estimation_command}{cmd:)} allows users to choose a command
other than {cmd:regress} for model estimation. Technically, {cmd:wgttest} will work with most 
estimation commands (if not all). This, however, does not mean that the test is always valid
(DuMouchel and Duncan, 1983, who proposed the test, discuss it solely within the framework of 
linear regression).

{p 4 8 2}
{cmd:nonoise} suppresses the estimation results.

{p 4 8 2}
{cmd:prefix(}{it:string}{cmd:)} allows users to choose a prefix other than {hi:_I} for
the interaction variables. The length of the prefix is restricted to 4 characters. Note that the 
interaction variables will only be created temporarily.

{p 4 8 2} {cmd:testopt(}{it:test_options}{cmd:)} may be used to pass options thru to the
post-estimation command {cmd:test}.

{p 4 8 2} {cmd:wgt(}{it:wgtvar}{cmd:)} specifies the sampling weights (mandatory). 

{p 4 8 2} {it:estimation_options} are passed thru to the estimation command.


{title:Examples}

{p 4 4 2} Test the impact of the sampling weights (variable {it:pwt}) for a linear 
regression model of wages on education and work experience:

{p 8 8 2} {cmd:. wgttest wage education experience, wgt(pwt)}


{title:References}

{p 4 8 2} DuMouchel, W. H. and G. J. Duncan (1983). Using Sample Survey Weights in Multiple 
Regression Analyses of Stratified Samples. Journal of the American Statistical 
Association 78: 535-543.
{p_end}
{p 4 8 2} Winship, C. and L. Radbill (1994). Sampling Weights and Regression 
Analysis. Sociological Methods and Research 23: 230-257.


{title:Author}

{p 4 4 2}
Ben Jann, ETH Zurich, jann@soz.gess.ethz.ch


{title:Also see}

{p 4 13 2}
Manual:  {hi:[U] 23 Estimation and post-estimation commands},{break}
 {hi:[U] 29 Overview of Stata estimation commands},{break}
 {hi:[R] test},{break}
 {hi:[R] regress}

{p 4 13 2}
Online:  help for {help weights}, {help test}, {help regress}