[R-sig-ME] Latent variable regression in lme4 as in HLM

[Ben Bolker]

On 5/26/20 10:12 PM, Simon Harmel wrote:
> Dear James,
>
> Thanks for your response. I should research this a bit more. I have 
> the feeling that HLM software under the /Latent Variable Regression 
> /tab might be doing something different. But I highly appreciate you 
> insightful response.
>
> Dear Ben, lmerControl(check.nobs.vs.nRE="ignore") didn't work.


    Really?  (Whenever you say "didn't work" you should give more 
detail; as I show below, it worked on my platform.  That it didn't work 
for you could mean a lot of different things [e.g. it really didn't work 
on your computer due to differences in lme4 version or OS or R version 
or ... ; you misinterpreted the results; ... Knowing what didn't work 
gives us a big head-start in helping you resolve the problem ...)

    I think it did what it was intended to do, which is bypass the check 
of numbers of obs vs number of random effect levels.

   I get:

    Warning message:
In checkConv(attr(opt, "derivs"), opt$par, ctrl = control$checkConv,  :
   Model failed to converge with max|grad| = 0.0028078 (tol = 0.002, 
component 1)

   However, the results look sensible.  Further exploration:

    aa <- allFit(m1)

 > aa
original model:
y ~ year + (year | stid)
data:  dat
optimizers (7): bobyqa, Nelder_Mead, nlminbwrap, nmkbw, 
optimx.L-BFGS-B,nloptwrap.NLOPT_LN_N...
differences in negative log-likelihoods:
max= 4.25e-06 ; std dev= 1.58e-06


summary(aa)

   Differences among parameter estimates, log-likelihoods, etc. are all 
very small, so the convergence warning is a false positive.

   cheers

     Ben Bolker


>
> library(lme4)
> dat <- read.csv('https://raw.githubusercontent.com/hkil/m/master/z.csv')
> m1 <- lmer(y~ year + (year|stid), data = dat,
>  control=lmerControl(check.nobs.vs.nRE="ignore"))
>
> On Tue, May 26, 2020 at 7:38 PM Ben Bolker <bbolker using gmail.com 
> <mailto:bbolker using gmail.com>> wrote:
>
>      For what it's worth, *if* you're sufficiently sure that your model
>     is identifiable, you can override the checks that test the relative
>     numbers of observations/levels/etc.; see the "check.*" options in
>     ?lme4::lmerControl, and set the relevant ones to "ignore"
>
>     On Tue, May 26, 2020 at 7:12 PM Uanhoro, James
>     <uanhoro.1 using buckeyemail.osu.edu
>     <mailto:uanhoro.1 using buckeyemail.osu.edu>> wrote:
>     >
>     > Hello Simon,
>     >
>     > I'm not sure what HLM does. However: if your question is about using
>     > the random intercepts (individuals' starting points) to predict the
>     > random slopes (their linear growth rate), then the model you
>     need is:
>     >
>     > summary(m2 <- lmer(y ~ year + (1 + year | stid), dat))
>     >
>     > whcih returns the random intercept and a random slope on time.
>     >
>     > The correlation between both random effects is the regression
>     > coefficient from regressing the slope on the intercept (or
>     vice-versa)
>     > when both variables are standardized.
>     >
>     > More generally, you can always obtain regression coefficients from a
>     > correlation/covariance matrix of random effects. With a two-by-two
>     > correlation matrix, the single correlation is the coefficient
>     (in both
>     > directions). In a larger matrix of random effects, you can use the
>     > solve() function in R to obtain coefficients from the matrix.
>     See here:
>     >
>     https://stackoverflow.com/questions/40762865/how-do-i-get-regression-
>     > coefficients-from-a-variance-covariance-matrix-in-r
>     >
>     > I tried your exact example, and m2 above will not fit because
>     some of
>     > your participants have under 2 time points while the maximum
>     number of
>     > time points is 3, resulting in a situation where the software is
>     > attempting to compute more random effect values than there are
>     rows in
>     > the data - the software complains. Also, it is a good idea to
>     rescale
>     > that y variable prior to data analysis. I was able to get the
>     model to
>     > run by limiting the data to cases with more than 1 recorded time
>     point:
>     >
>     > summary(m3 <- lmer(y.s ~ year + (1 + year | stid), data = t.dat,
>     subset
>     > = n > 1))
>     >
>     > I arrived at a correlation/coefficient of -0.06.
>     >
>     > Hope this helps, -James.
>     >
>     > Sent from Outlook Mobile
>     >
>     > From: R-sig-mixed-models
>     <r-sig-mixed-models-bounces using r-project.org
>     <mailto:r-sig-mixed-models-bounces using r-project.org>> on
>     > behalf of Simon Harmel <sim.harmel using gmail.com
>     <mailto:sim.harmel using gmail.com>>
>     > Sent: Tuesday, May 26, 2020, 18:27
>     > To: r-sig-mixed-models
>     > Subject: [R-sig-ME] Latent variable regression in lme4 as in HLM
>     >
>     > Dear All,
>     >
>     > I know that in the HLM software, it is possible to use "intercept"
>     > (e.g.,
>     > initial place of students at year "0") as the *predictor *of "slope"
>     > (e.g.,
>     > fixed rate of change in years) under the *Latent Variable Regression
>     > *tab.
>     >
>     > I was wondering if this is also possible in "lme4" or any other
>     > mixed-modeling packages in R?  *Thanks, Simon*
>     >
>     > *## Here is an example dataset for demonstration:*
>     > library(lme4)
>     > dat <- read.csv('
>     >
>     https://urldefense.com/v3/__https://raw.githubusercontent.com/hkil/m/master/z.csv__;!!KGKeukY!gjIgidLro6PaJJUHZOY1gk9IW8FfrzGWzo9IEaCRgFwkorvpE1tkLXqn3ujcsmy6OvxsB2xlUIM$
>     >  ')
>     > m1 <- lmer(y ~ year + (1|stid), data = dat)      #### 'stid' =
>     student
>     > id
>     >
>     >         [[alternative HTML version deleted]]
>     >
>     > _______________________________________________
>     > R-sig-mixed-models using r-project.org
>     <mailto:R-sig-mixed-models using r-project.org> mailing list
>     >
>     >
>     https://urldefense.com/v3/__https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models__;!!KGKeukY!gjIgidLro6PaJJUHZOY1gk9IW8FfrzGWzo9IEaCRgFwkorvpE1tkLXqn3ujcsmy6Ovxs517Hn00$
>     >
>     > _______________________________________________
>     > R-sig-mixed-models using r-project.org
>     <mailto:R-sig-mixed-models using r-project.org> mailing list
>     > https://stat.ethz.ch/mailman/listinfo/r-sig-mixed-models
>

	[[alternative HTML version deleted]]