Re: [R] Modelling heteroskedasticity in a multilevel model

From: Spencer Graves <spencer.graves_at_pdf.com>
Date: Tue 02 May 2006 - 12:11:10 EST

          I generally prefer to start with the distributional characteristics of the data, then consider "fixed effets" / explanatory variables, then dependence structure in the deviations from the model.

  1. What is (are) your response variable(s)? If the numbers are financial numbers of volume of sales in whatever units and can never be negative, I would automatically take logarithms before I did anything else -- possibly after replacing any 0's by some small positive number to prevent sending part of the data to (-Inf). Have you made normal probability plots and time series plots by company? If it's something like stock prices, I might focus on first differences of something like log(price per share). (And where are you getting your data? For certain financial research questions, you need include dividends appropriately when computing "total increase in shareholder value". I believe the Center for Research in Securities P rices at the University of Chicago does a fairly good job of this, and it's probably difficult to match what they do.)
  2. If you've appropriately transformed your data, what have you done to explore the relationships between your response variables and potential explanatory variables? Plots, preceeded or followed by "lm" fits are wise.
  3. Then I want to consider residuals, by industry and correlated over time -- and correlations in absolute values [which might suggest the need for something like "ARCH / GARCH" (autoregressive conditional heteroscedasticity)? I mention this because you talk about "'unexpected' shocks at especific points in time". If these "shocks" can be explained by some other variable, then you don't need ARCH / GARCH. However, if the sources can not be reasonably modeled from other variables, then you might need ARCH / GARCH.]
	  hope this helps.
	  spencer graves

Antonio Revilla wrote:

> Dear list members,
>
> I am facing a 3-level model, for which my research hypotheses suggest that
> the variance of both level-1 and level-2 residuals may be a function of a
> level-3 variable.
>
> To be a bit more clear: I am fitting a longitudinal model for a panel of
> companies grouped in industries. I suggest that some industry variables may
> create 'unexpected' shocks at especific points in time; such shocks are not
> accounted for by the explanatory variables in the model, so that they will
> presumably increase variance of level-1 residuals. On the other hand,
> industry-level attributes may also affect the relative relative size of
> firm-level permanent effects (represented by level-2 residuals)
>
> Do you know how could I model such a residual structure in R? I have been
> looking at the varfunc command in the nlme package, but I am not sure if
> such a function can perform the kind of analysis I actually need.
>
> Thank you very much in advance,
>
> Antonio
>
> _________________________________________________________________
> ¿Estás pensando en cambiar de coche? Todas los modelos de serie y extras en
> MSN Motor. http://motor.msn.es/researchcentre/
>
> ______________________________________________
> R-help@stat.math.ethz.ch mailing list
> https://stat.ethz.ch/mailman/listinfo/r-help
> PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html



R-help@stat.math.ethz.ch mailing list
https://stat.ethz.ch/mailman/listinfo/r-help PLEASE do read the posting guide! http://www.R-project.org/posting-guide.html Received on Tue May 02 15:38:43 2006

Archive maintained by Robert King, hosted by the discipline of statistics at the University of Newcastle, Australia.
Archive generated by hypermail 2.1.8, at Tue 02 May 2006 - 18:09:52 EST.

Mailing list information is available at https://stat.ethz.ch/mailman/listinfo/r-help. Please read the posting guide before posting to the list.