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

RE: st: estimating an asymmetrical relationship

From   "Nick Cox" <>
To   <>
Subject   RE: st: estimating an asymmetrical relationship
Date   Wed, 16 Dec 2009 12:19:07 -0000

I was asking whether you are using time-series ideas such that present response is a function of past and present values of predictor only or whether you were lumping together data regardless of time. It sounds like the latter, but in that case I wouldn't call it a time series model. 

As no other variables are mentioned, and no theory is being tested, the next step would seem to be to use your favourite scatter plot smooth on all the data and on the data separated by increasing or decreasing to suggest a specification. You'd still need to check for autocorrelation of error terms as usual, among other things. 


Gisella Young

Thank you very much Nick and David for your thoughtful and helpful replies.

On Nick's point about whether I need a model that treats
my variables as linked time series or whether I am lumping together data
without regard to time, I am not sure whether I understand this correctly and maybe I am exposing my ignorance here, but I am estimating a normal (?) time series model, so the observations have dates but I am not doing any explicit linking...    One can look at this relationship in terms of economic theory and there are explanations, but what I need to do here is empirically test whether there is an asymmetry for a particular country over a particular time period.  I haven't found literature that econometrically tests for this type of asymmetrical relationship in other contexts, which is why I am struggling about the correct way of doing it. 

On David's suggestion of dummies for the different periods (increasing and decreasing of thee explanatory variable), actually this was something I considered but was told my an econometrician colleague that this was wrong (not sure why) and that only a nonlinear estimation would be appropriate. I would be interested in hearing the views of other list members on this?

Or alternatively, a suitable way of specifying and testing the relationship in STATA, whether using OLS or nonlinear specification?

--- On Tue, 12/15/09, David Jacobs <> wrote:

> From: David Jacobs <>
> Subject: Re: st: estimating an asymmetrical relationship
> To:
> Date: Tuesday, December 15, 2009, 6:18 PM
> Gisella, how about this technique
> (and let's see what others on the list say about it, as it's
> been a while since I thought about this issue)?
> Construct a dummy coded 1 if there's growth in the
> explanatory variable of interest and construct another if
> there's negative change.  Multiply each dummy with the
> explanatory variable of interest thereby creating two
> interacted explanatory variables.  Enter both product
> terms in a model.  If there's a statistically
> significant difference between the coefficients on the two
> interacted terms, that should provide evidence that you
> indeed have asymmetric effects.
> At 11:55 AM 12/15/2009, you wrote:
> > This question is partly econometric and partly STATA,
> and I would greatly appreciate any advice. I need to
> estimate a relationship which I believe to be asymmetric -
> between growth as the explanatory variable of interest and
> unemployment as the dependent variable. Asymmetric in the
> sense that, while overall I expect a negative relationship,
> I expect the magnitude of the relationship to be higher when
> growth is going down than when it is going up (for reasons
> not necessary to go into here). That is, I expect the
> coefficient on the relationship to be higher during periods
> of high or increasing growth than during periods of low or
> decreasing growth (I realise that high and increasing is not
> the same thing but I will later figure out which). So I
> believe that OLS is not appropriate as it will not pick up
> these differences, and what I am really interested in
> investigating is whether there is asymmetry and how strong
> it is. Ideally I would like to get
> >  different coefficients for the 2 types of
> periods (note that this is not one long period and then
> another long period, but ups and downs). I believe that I
> need a nonlinear estimation method to do this. But beyond
> that I am not sure how to proceed and would be terribly
> grateful for any suggestions.

*   For searches and help try:

© Copyright 1996–2017 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index