[R-sig-ME] About warning
bbo|ker @end|ng |rom gm@||@com
Sun May 16 23:28:38 CEST 2021
There are two issues here.
- fixed-effect model matrix is rank deficient: this is not necessarily
a problem, but you should try to understand what's going on. There is a
detailed StackOverflow post here:
You can look at your results to figure out which coefficients are
missing, or you can use attr(getME(m,"X"), "col.dropped") to see which
columns of the original model frame were dropped. (It would be useful
if I or someone would write a more detailed diagnostic function to
report which columns of the original model matrix were collinear ...)
- singular fit.
This is a little harder. In some sense this means that your model is
overfitted (there is sufficiently little data that the best estimate of
the inter-group variance is zero). Conclusions on the fixed effects are
most likely OK. *If* the singular component corresponds to the variance
of an entire random-effect term being exactly zero (e.g. you have a
random intercept model), then you'll get exactly the same results for
the fixed effects, at least to the level of point estimates and Wald
confidence intervals/p-values, if you drop the RE term.
There is a lot more discussion in ?lme4::isSingular, for a start ...
On 5/16/21 6:10 AM, Atsuko Nakagawa wrote:
> Thank you very much for your help.
> When we ran Package lme4 to our longitudinal data, we got following
> warnings repeatedly.
>> fixed-effect model matrix is rank deficient so dropping 2 columns /
>> boundary (singular) fit: see ?isSingular
> Nevertheless, we got some results. Are these valid?
> When I look for the Q&A in some Statistic Forums, sometimes I found the
> previous answer that we do not have to care these warnings.
> In actuality, how do we deal with these warnings?
> Thank you for your time.
> Atsuko Nakagawa
> Graduate School of Humanities and Social Sciences, Nagoya City
> University, JAPAN
> [[alternative HTML version deleted]]
> R-sig-mixed-models using r-project.org mailing list
More information about the R-sig-mixed-models