If a model is "singular", this means that some dimensions of the
variance-covariance matrix have been estimated as exactly zero. This
often occurs for mixed models with complex random effects structures.
“While singular models are statistically well defined (it is theoretically
sensible for the true maximum likelihood estimate to correspond to a
singular fit), there are real concerns that (1) singular fits correspond
to overfitted models that may have poor power; (2) chances of numerical
problems and mis-convergence are higher for singular models (e.g. it
may be computationally difficult to compute profile confidence intervals
for such models); (3) standard inferential procedures such as Wald
statistics and likelihood ratio tests may be inappropriate.”
(lme4 Reference Manual)
There is no gold-standard about how to deal with singularity and which
random-effects specification to choose. Beside using fully Bayesian methods
(with informative priors), proposals in a frequentist framework are:
avoid fitting overly complex models, such that the variance-covariance matrices can be estimated precisely enough (Matuschek et al. 2017)
use some form of model selection to choose a model that balances predictive accuracy and overfitting/type I error (Bates et al. 2015, Matuschek et al. 2017)
“keep it maximal”, i.e. fit the most complex model consistent with the experimental design, removing only terms required to allow a non-singular fit (Barr et al. 2013)