Skip to content
Prev 3724 / 20628 Next

Random vs. fixed effects

I am surprised at suggestions that the problem of random effects with very few levels can be overcome by changing modelling framework or estimation algorithms.

It seems to me that the problem is a basic one of lack of information. If you want to estimate the variance of something you really need more than 3 observations. No amount of fiddling with the model or algorithms will change that. The only thing that might change it is to use Bayesian methods with informative priors, but then you are changing the problem by taking some information from elsewhere.

I reach the same conclusion if I think about the 'scope of inference'. I cannot expect to say much about a population I have just 3 observations from.

Statistical finesse cannot compensate for a basic lack of information. 


Richard Coe
Principal Scientist ? Research Methods
World Agroforestry Centre (ICRAF), Nairobi

and

Statistical Services Centre, U Reading.


-----Original Message-----
From: r-sig-mixed-models-bounces at r-project.org [mailto:r-sig-mixed-models-bounces at r-project.org] On Behalf Of Emmanuel Charpentier
Sent: Saturday, April 24, 2010 12:57 AM
To: r-sig-mixed-models at r-project.org
Subject: Re: [R-sig-ME] Random vs. fixed effects

Two cents from an humble practitioner :

Le vendredi 23 avril 2010 ? 14:11 -0400, Ben Bolker a ?crit :
Schmilosophically speaking, option #3 has a set of interesting
features :

- It entails the creation and fitting of a full joint probability model.
Internal consistency is guranteed.

- Bayesian modeling tools (especially BUGS) offer enough flexibility to
*require* explicit *choices* for model assumption. For exmple, the
"standard" choice of a normal distribution for level II effects has to
be *explcitely writtern by the modeler, which gives him/her an
opportunity to consider and/or justify his/her choice. This is also true
of the choice of the modelization of nuisance parameters.

- There is no "formal" distinction between fixed and random effects ;
the latter are given a distribution (to be fitted), whereas the formere
are not.

However : 

- getting such a model to converge can be hard (especilly with BUGS).
Current tools need a real (maybe too much) understanding of the
numerical analysis problems to choose an *efficient* model expression,
and some "tricks" used to get convergence (such as rendering a model
deliberately unidentifible...) are difficult to justify rom a modeling
standpoint. Again, Gelman & Hill comparison of the current Bayesian
tools to early-generation regression software seems quite apt.

- The choices made about shape of distributions, relationships, etc ...
should be submitted to a sensitivity analysis, which raises again the
statistician's workload.

In short, option #3 is the *expensive* way of (maybe) getting the most
precise honest answer to your problem... if such thing exists and is
reachble with current numerical tools.

Murphy's law ensures that option #2 will "break" (and warnngs will be
ignored) especilly when warnings mean something important.

But option #1 is *also* an acceptable one : using it entails modeling
*conditionally on your experimental setup*. You won't be able to state
(directly) an estimation of the possible "population" effects of the
formerly-random effects, and the validity of your inferences about the
really-fixed effects will be formerly limited to the particular
population defined by these formerly-random effects. But that might be
okay if you have other information allowing you (and your reader) to
independently assess the possible value of an extrapolation from this
particulr subpopulation to an hypothetical population.

In short, option #1 is a cheap wy of getting a ossibly ccetable
approximate solution to your problem, whose value has to be assessed by
other means.

Of course, you won't benefit from artial pooling, and serious inter-unit
interactions will mess your inferences ...

HTH,

					Emmanuel Charpentier
_______________________________________________
R-sig-mixed-models at r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
 

__________ Information from ESET NOD32 Antivirus, version of virus signature database 5055 (20100424) __________

The message was checked by ESET NOD32 Antivirus.

http://www.eset.com
 
 

__________ Information from ESET NOD32 Antivirus, version of virus signature database 5055 (20100424) __________

The message was checked by ESET NOD32 Antivirus.

http://www.eset.com