Panel Data - Variance decomposition

[Valeria Frigerio]

Suppose I have a panel dataset, in which I have a variable (ex: wage) of which I want to analyze the determinants and decompose the contribution of each in explaining the total variance.
My regressors are of two basic kinds: some vary only over time and/or groups of individuals, others are specific to individuals.
For example, I may want to explain wage depending on country of living (or, some index of labour market "elasticity"), sector of the company, macroeconomic conditions in the country, and then on individual characteristics like education, job experience, age, and so on.
In terms of regression, I would estimate it with Blundell-Bond model, as I want the model to be dynamic (dependent variable today function of its value in previous period).
I would like to add a variance decomposition analysis to this, and looking at Stata Journal I think I could use either Anova or Mixed effects. However, what if my individual characteristics interact with group variables? Is it still appropriate?
Thank you for help

Valeria

[Carlo Lazzaro]

Valeria,
welcome to the Forum.
As per FAQ, please provide full reference about Blundell-Bond model, as I'm not (and possibly Other Statalisters who are much more active than me in replying are not) familiar with all the literature that is specific for each different research fields we come from.
As far as your query is concerned, you may be interested in taking a look at - help xtdpdsys - (Arellano-Bover/Blundell-Bond linear dynamic panel-data estimation).
Kind Regards,
Carlo

[Valeria Frigerio]

Carlo,
thanks for your kind reply. I am already familiar with xtabond and xtdpdsys, what I was asking is something different.
I have a panel, where I have multiple observations for firms' data, and variables that measure the business cycle and the quality of national institutions.
What I want to achieve is a variance decomposition of how a firm characteristic changes (for example the CEO pay) as a function of other individual firm characteristics (profits, sales growth, and so on) and as a function of macro variables (business cycle, institutions).
I want to understand how much of the total variance is explained by each of the independent variables (firm specific and not firm specific), and then also how much of the within variance, and how much of the between variance, so that I can understand which factors are most relevant across companies and/or over time.
Kind regards
Valeria

[Roman Mostazir]

Not sure if this is what you are asking for. In simple language, aren't you essentially looking for effect sizes for your independent variables in a mixed model setting i.e. amount of variance explained by each IVs ? If so, what about Cohen's f² ? You can easily get it done in Stata by using global macros. This UCLA page has an example http://www.ats.ucla.edu/stat/stata/f...ffect_size.htm .

[Carlo Lazzaro]

Valeria,
probably also the following thread can be interesting for you
http://www.statalist.org/forums/foru...ance-explained.

Kind regards,
Carlo

[stata3]

Hi


I have come across this paper which does a variance decomposition.
https://www.ilr.cornell.edu/director...ions_mar04.pdf

The paper refers to the stata routine which I am not sure how to implement: am neither a Stata expert nor an econometrics expert
http://www.komkon.org/~tacik/stata/gfields.ado

If anybody could help me, I would really appreciate. I also need this variance decomposition for my fixed effects model.

Thank you!

[Stephen Jenkins]

"Stata3": Gary Fields' regression approach to inequality decomposition is just one approach. For a recent review of the field placing this in context, I recommend that you read Cowell and Fiorio's 2010 paper on inequality decomposition at http://www.uva-aias.net/uploaded_fil...ell,Fiorio.pdf . You could also have used Stata's resources to find relevant material, e.g. search regression decomposition throws up many hits. One of them is ineqrbd: ssc describe ineqrbd. (This can implement a Fields-type decomposition.)
You are new to Statalist -- welcome. But, please, read the Forum FAQ (hit the black bar at the top of page) and follow recommended protocol. Amongst these is a very strong recommendation that you use your real name (firstname lastname). So, please re-register by hitting the "Contact Us" link in the blue bar at the bottom of the page, and make your request. It's easy and fast. (Cutesy tags like "Stata3" do not make you look serious, in my opinion, and are likely to diminish the chances of you getting responses in future.)

[Christian Agethen]

Dear Valeria,

your panel setting sounds like a nice dataset for using multilevel regressions, especially for decomposing variances. In your setting the data would be nested over several levels (at least two levels where companies are nested within countries). I'd recommend looking up Rabe-Hesketh's "Multilevel and Longitudinal Modeling Using Stata", where you'll find great examples for interactions between levels, in your surrounding the covariation of macro- and company-level variables.

Best,
Christian

PS: Welcome to the forum!

[Valeria Frigerio]

Thank you all for your kind replies, which have been very helpful.