Re: st: about residuals and coefficients

[David Hoaglin <dchoaglin@gmail.com>]

Ronan,

The sorts of factors that I had in mind are ones that would be
classified as fixed in analysis of variance.  Experimental units (such
as the rats that you mention) would have combinations of fixed factors
assigned to them by the design, and the analysis would allow for
random variation among them (but not treat them as identical).

David Hoaglin

On Sat, Sep 7, 2013 at 7:22 AM, Ronan Conroy <rconroy@rcsi.ie> wrote:
> On 2013 MFómh 7, at 01:42, David Hoaglin wrote:
>
>> Most experiments in social science cannot collect data that allows all
>> variables to be held constant.  A good design, however, may include
>> all combinations of two or more factors, so that one can study the
>> effect of one factor without changing the other factors.  Usually,
>> many additonal variables can only be observed.  Those are analyzed as
>> covariates (and adjusted for).
>
> I don't tend to believe in this. It implies that the same experimental units are studied under constant conditions, while in real life we study experimental units treated as identical (Wistar rats, for example) under conditions are are not really constant but in which random variation is tolerated (exact site of injection, nearness of rats' cage to radio playing some awful music all day…)
>
> Experiments in the prediction of lifespan of genetically identical roundworms observed in a controlled environment have consistently failed to predict lifespan based on any observable data such as nutrient intake, energy expenditure etc.
>
> See the brilliantly-argued paper by George Davey Smith -
> Davey Smith, G. Epidemiology, epigenetics and the “Gloomy Prospect”: embracing randomness in population health research and practice. Int J Epidemiol. 2011 Jun;40(3):537–62.
>
> Ronán Conroy

*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/faqs/resources/statalist-faq/
*   http://www.ats.ucla.edu/stat/stata/