This page provides answers to some of the questions that get asked most often about mgcv
How can I compare gamm models? In the identity link normal errors case, then AIC and hypotheis
testing based methods are fine. Otherwise it is best to work out a strategy based on the summary.gam
Alternatively, simple random effects can be fitted with gam
, which makes comparison straightforward.
Package gamm4
is an alternative, which allows AIC type model selection for generalized models.
How do I get the equation of an estimated smooth? This slightly misses the point
of semi-parametric modelling: the idea is that we estimate the form of the function from data
without assuming that it has a particular simple functional form. Of course for practical computation
the functions do have underlying mathematical representations, but they are not very helpful, when written
down. If you do need the functional forms then see chapter 4 of Wood (2006). However for most purposes it
is better to use predict.gam
to evaluate the function for whatever argument values you need.
If derivatives are required then the simplest approach is to use finite differencing (which also allows
SEs etc to be calculated).
Some of my smooths are estimated to be straight lines and their confidence intervals vanish at some point in the middle. What is wrong? Nothing. Smooths are subject to sum-to-zero identifiability constraints. If a smooth is estimated to be a straight line then it consequently has one degree of freedom, and there is no choice about where it passes through zero — so the CI must vanish at that point.
How do I test whether a smooth is significantly different from a straight line. See tprs
and the example therein.
An example from an mgcv helpfile gives an error - is this a bug? It might be, but first please check that the version of mgcv you have loaded into R corresponds to the version from which the helpfile came. Many such problems are caused by trying to run code only supported in a later mgcv version in an earlier version. Another possibility is that you have an object loaded whose name clashes with an mgcv function (for example you are trying to use the mgcv multinom
function, but have another object called multinom
loaded.)
Some code from Wood (2006) causes an error: why? The book was written using mgcv version 1.3. To allow for
REML estimation of smoothing parameters in versions 1.5, some changes had to be made to the syntax. In particular the
function gam.method
no longer exists. The smoothness selection method (GCV, REML etc) is now controlled by the
method
argument to gam
while the optimizer is selected using the optimizer
argument.
See gam
and http://www.maths.bris.ac.uk/~sw15190/igam/index.html for details.
Why is a model object saved under a previous mgcv version not usable with the current mgcv version? I'm sorry about this issue, I know it's really annoying. Here's my defence. Each mgcv version is run through an extensive test suite before release, to ensure that it gives the same results as before, unless there are good statistical reasons why not (e.g. improvements to p-value approximation, fixing of an error). However it is sometimes necessary to modify the internal structure of model objects in a way that makes an old style object unusable with a newer version. For example, bug fixes or new R features sometimes require changes in the way that things are computed which in turn require modification of the object structure. Similarly improvements, such as the ability to compute smoothing parameters by RE/ML require object level changes. The only fix to this problem is to access the old object using the original mgcv version (available on CRAN), or to recompute the fit using the current mgcv version.
When using gamm
or gamm4
, the reported AIC is different for the gam
object and
the lme
or lmer
object. Why is this? There are several reasons for this. The most important is that the
models being used are actually different in the two representations. When treating the GAM as a mixed model, you are
implicitly assuming that if you gathered a replicate dataset, the smooths in your model would look completely
different to the smooths from the original model, except for having the same degree of smoothness. Technically you would expect
the smooths to be drawn afresh from their distribution under the random effects model. When viewing the gam from the
usual penalized regression perspective, you would expect smooths to look broadly similar under replication of the data.
i.e. you are really using Bayesian model for the smooths, rather than a random effects model (it's just that the frequentist random
effects and Bayesian computations happen to coincide for computing the estimates). As a result of the different assumptions
about the data generating process, AIC model comparisons can give rather different answers depending on the model adopted.
Which you use should depend on which model you really think is appropriate. In addition the computations of the AICs are
different. The mixed model AIC uses the marginal liklihood and the corresponding number of model parameters. The gam model
uses the penalized likelihood and the effective degrees of freedom.
What does 'mgcv' stand for? 'Mixed GAM Computation Vehicle', is my current best effort (let me know if you can do better). Originally it stood for ‘Multiple GCV’, which has long since ceased to be usefully descriptive, (and I can't really change 'mgcv' now without causing disruption). On a bad inbox day 'Mad GAM Computing Vulture'.
My new method is failing to beat mgcv, what can I do? If speed is the problem, then make sure that you use the
slowest basis possible ("tp"
) with a large sample size, and experiment with different optimizers to find one that is
slow for your problem. For prediction error/MSE, then leaving the smoothing basis dimensions at their
arbitrary defaults, when these are inappropriate for the problem setting, is a good way of reducing performance.
Similarly, using p-splines in place of derivative penalty based splines will often shave a little more
from the performance here. Unlike REML/ML, prediction error based smoothness selection criteria such as Mallows Cp and GCV
often produce a small proportion of severe overfits, so careful choise of smoothness selection method can help further.
In particular GCV etc. usually result in worse confidence interval and p-value performance than ML or REML. If all this
fails, try using a really odd simulation setup for which mgcv is clearly not suited: for example poor performance is almost
guaranteed for small noisy datasets with large numbers of predictors.
Simon N. Wood simon.wood@r-project.org
Wood S.N. (2006) Generalized Additive Models: An Introduction with R. Chapman and Hall/CRC Press.
Questions? Problems? Suggestions? Tweet to @rdrrHQ or email at ian@mutexlabs.com.
All documentation is copyright its authors; we didn't write any of that.