Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: instrumental variable technique to address large measurement error?


From   Stas Kolenikov <skolenik@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: instrumental variable technique to address large measurement error?
Date   Mon, 18 Jul 2011 10:57:06 -0500

I would build an SEM instead. With measurement error of this
magnitude, you might have bias of 50% or so in your parameter
estimates -- i.e., they will be no good for anything.

Without you explaining your instrumental variables strategy, there is
no way anybody could give any advice on how well it might work in
controlling the measurement error. It looks like you've seen very
dated advice on instrumental variables from somebody stuck in the
1960s. If you have followed Bollen's approach with model-implied
instrumental variables
(http://www.citeulike.org/user/ctacmo/article/553229 and other stuff
in http://www.citeulike.org/user/ctacmo/tag/miiv), the procedure would
be as follows:
1. identify the most reliable indicator for each trait (using my
-confa- package, for instance)
2. use it as a scaling indicator for this latent (so that its loading is 1)
3. use all other indicators of the latent as instruments
4. build a regression of the outcome of interest on the scaling
indicators, putting all other indicators as instruments
5. use an overidentification test (-estat overid- following
-ivregress-) as an overall test that's powerful against cross-loadings
and correlated measurement errors involving the scaling indicators.
This test will not care much for correlated measurement errors among
the variables that were used as instruments. You can view this as a
lack of sensitivity issue, or as a robustness advantage if you really
want to estimate your regression and are willing to tolerate a
marginal measurement model.

On Mon, Jul 18, 2011 at 10:30 AM, xueliansharon <xuelianstata@gmail.com> wrote:
> Dear all,
>
> I got quite low Cronbach's alpha coefficients for the five personality
> traits in my sample (extraversion, 0.54; agreeableness, 0.55;
> conscientiousness, 0.60; emotional stability, 0.62; intellect, 0.55.), and
> all these five personality traits are included in the model as the RHS
> independent variables.
>
> I know that such low alpha coefficients indicate large measurement errors,
> and I intend to use instrumental variable techniques to address for
> measurement error problem, i.e. I will employ 5 instruments for the five
> personality traits, and try to see what the effects of five personality
> traits are on the LHS dependent variable.
>
> I just wonder whether such IV technique is enough to correct for measurement
> errors.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   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/


© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index