[R-sig-ME] lme4, failure to converge with a range of optimisers, trust the fitted model anyway?

[Ben Bolker]

-----BEGIN PGP SIGNED MESSAGE-----
Hash: SHA1

On 15-04-04 02:15 PM, Hans Ekbrand wrote:
> On Sat, Apr 04, 2015 at 09:10:35PM +1100, Ken Beath wrote:
>> One of the problems is that you have a relatively high random
>> effects variance. A standard deviation of the intercept of 3 is a
>> huge amount, it means that there is massive variation in the
>> random effect value needed to model each cluster, to the point
>> that some clusters will be all zeros and some will be all ones.
>> In this situation the assumption of approximate normality of the
>> likelihood around the nodes which is required for using Laplace's
>> method is very far from met.
> 
> Thanks for your advice, I really appreciate it!
> 
> I tried nAGQ=5, but met with:
> 
> ## Error: nAGQ > 1 is only available for models with a single,
> scalar random-effects term
> 
> As you point out, some clusters will be all zeros (and some will
> be all ones). While my data is on the individual level,
> 
> a) the variable I'm mainly interested in, KilledPerMillion5Log,
> varies only at the country level, &
> 
> b) I currently have no variables in the model that vary at the 
> individual level
> 
> So, perhaps I could aggregate the individual level data to the
> cluster level, and do without the random term for cluster? I mean,
> calculate the proportions of yes in each cluster and use that as
> the dependent variable.
> 
> This would, I assume, require that each cluster was given a weight 
> that corresponded to the number of individuals in it - or I would
> not be able to say anything about probabilities at the individual
> level, right?

 I would say pretty much any time you can aggregate without losing
information, you should, for ease of computation (this accords with
the Murtaugh 2007 "Simplicity and complexity" paper that I cite here
on a regular basis).  If all of your covariates are at the cluster
level, then this reduces to a binomial GLM (you can account for
overdispersion either by using a quasi-binomial model in glm() or by
staying with glmer() and keeping the random effect of cluster (now an
observation-level random effect).

  Ben Bolker


-----BEGIN PGP SIGNATURE-----
Version: GnuPG v1.4.11 (GNU/Linux)

iQEcBAEBAgAGBQJVIHVjAAoJEOCV5YRblxUHUYsH/243gzJN9hWayUICMmQop6iZ
+p+RIxh3lpSmB1lnRX+VXpghNpB4a5s98m+q5HZJB2eUyBCJ7pwXbHuTaZZ2t2D0
v6V/Ih2VeoqnbVqaSrPDNgDufNO1RhHlc4h6L0bIGf/etmOHgtHaPnly5JUXUmq4
QZdt5Sxotqshdbqf+4FvMNFVa1cNWJNLBY0XB2GecSLa8MI7vF1UHvlqdtDWIO9e
WalpfJq6C4oqIyz2kCsZWRijSHroFpPJ6eMlKmbLFLxfY6rP5dQZK11ktsANQ7lu
bnPHoF1Xzwnl87XDl/v6K9P1ge2x7bvihDUVodjBoK31CHc8UAYvh/5Ud4AGKoY=
=9xK3
-----END PGP SIGNATURE-----