Message-ID: <0e42048d-d240-74cc-61b5-c12582ca1f71@gmail.com>
Date: 2021-10-04T13:12:03Z
From: Ben Bolker
Subject: Issue with boundary (singular) fit: see ?isSingular
In-Reply-To: <CAF08B3iWYHzsiVSHUyTCSgg-nbG2dX6nv0D3QW7PPAm+yGvAnA@mail.gmail.com>
On 10/4/21 10:05 AM, Sasha Vasconcelos wrote:
> Hi,
>
> I'm running a piecewise SEM with 3 component models:
>
> lmer(response variable1 ~ predictors + (1|Point) + (1|Year), input_table)
>
> glmer(response variable2 ~ predictors + (1| Point) + (1|Year), family =
> "binomial", input_table)
>
> glmer(response variable3 ~ predictors + (1| Point) + (1|Year), family =
> "binomial", input_table)
>
> Because sampling involved visiting 18 points in spring of 2018 and again in
> spring of 2019, I specified samping point and year as random effects.
If there are only two years, it's not surprising that you'll get
estimates of zero variance for (1|Year). I would probably make Year a
fixed effect.
>
> When I run the model, this warning message appears:
> Check model convergence: log-likelihood estimates lead to negative
> Chi-squared!
I can't find this warning message anywhere, even in the development
branch of piecewiseSEM:
https://github.com/jslefche/piecewiseSEM/search?q=convergence
??
>
> This message also appears:
> boundary (singular) fit: see ?isSingular
>
> From what I've read about the second message, it could be due to random
> effect variance estimates of zero. I checked and this happens in the 1st
> and 3rd component models. In the 1st model "Point" has zero variance, and
> in the 3rd model "Year" has zero variance.
>
> My first question is (and I apologize in advance if this is silly to ask)
> whether this means that there's not really an effect coming from Point in
> component model 1 and from Year in component model 2? If so, would it be
> possible to remove those random effects to end up with:
>
> lmer(Response variable1 ~ Predictors + (1|Year), input_table)
>
> glmer(Response variable2 ~Predictors + (1| Point) + (1|Year), family =
> "binomial", input_table)
>
> glmer(Response variable3 ~ Predictors + (1| Point), family = "binomial",
> input_table)
Seems reasonable.
>
> My second question is whether the warning "Check model convergence:
> log-likelihood estimates lead to negative Chi-squared!" is related to these
> singularity issues?
>
> Oh and I am using the development version of the piecewise SEM package
> installed using devtools. This is because this version provides additional
> standardized coefficients for GLMM.
>
>
> Thanks!
>
>
--
Dr. Benjamin Bolker
Professor, Mathematics & Statistics and Biology, McMaster University
Director, School of Computational Science and Engineering
Graduate chair, Mathematics & Statistics