I would like to get the confidence interval (CI) for the predicted mean of a Linear Mixed Effect Model on a large dataset (~40k rows), which is itself a subset of an even larger dataset. This CI is then used for estimating the uncertainty of another calculation that uses the mean and its related CI as input data.
I managed to create a prediction estimate and interval for the full dataset, but a Prediction Interval is not the same and much larger than a CI. Beside bootstrapping (which takes way too much time with this much data), I cannot find a method that would allow me to estimate a CI – either because it is throwing errors or because it only offers to calculate Prediction intervals.
I quite recently moved into LME and I might therefore have overseen some obvious method.
Here is what I did so far in more detail:
The input data is confidential and I can therefore unfortunately not share any extract.
But in general, we have one dependent variable (y) representing the probability of a event and 2 categorical (c1 and c2) and two continuous variables (x1 and x2) with some weighting factor (w1). Some values in the dataset are missing. An extract of the first rows of the data could look like the example below:
c1
c2
x1
x2
w1
y
London
small
1
10
NA
NA
London
small
1
20
NA
NA
London
large
2
10
0.2
0.1
Paris
small
1
10
0.2
0.23
Paris
large
2
10
0.3
0.3
Based on this input data, I am then fitting a LMER model in the following form:
lmer1 <- lme4::lmer( y ~ x1 * poly(x2, 5) + ((x1 * poly(x2 ,5)) | c1),
data = df,
weights = w1,
control = lme4::lmerControl(check.conv.singular = lme4::.makeCC(action = "ignore", tol = 1e-3)))
This runs for some minutes and returns several warnings:
Warning messages: 1: In optwrap(optimizer, devfun, getStart(start,
rho$pp), lower = rho$lower, : convergence code 5 from nloptwrap:
NLOPT_MAXEVAL_REACHED: Optimization stopped because maxeval (above)
was reached.
2: In checkConv(attr(opt, “derivs”), opt$par, ctrl =
control$checkConv, : unable to evaluate scaled gradient
3: In checkConv(attr(opt, “derivs”), opt$par, ctrl =
control$checkConv, : Model failed to converge: degenerate Hessian with
11 negative eigenvalues
I increased the MAXEVAL parameter but this still did not help to get rid of the warnings and I found that despite these warnings, the model is still fitted. I therefore started to apply different methods to get a prediction of the mean for the whole dataset and the related CI for the mean.
predictInterval
I started with creating a Prediction Interval for the full dataset:
predictions <- merTools::predictInterval(lmer1,
newdata = df,
which = "full",
n.sims = 1000,
include.resid.var = FALSE,
level=0.95,
stat="mean")
However, as stated above, the Prediction Interval is not the same as the CI (see also https://datascienceplus.com/prediction-interval-the-wider-sister-of-confidence-interval/).
I found that the general predict function has the option to set interval to either “prediction” or “confidence”, but this option does not exist with the prediction from a LMER object. And I could not find another possibility to switch from Prediction Interval to CI – even though I would believe that the data drawn should be sufficient to do this.
confint
I then saw that there is a function called “confint”, but when running this function I get the following error:
predicition_ci = lme4::confint.merMod(lmer1)
Computing profile confidence intervals ...
Error in zeta(shiftpar, start = opt[seqpar1][-w]) : profiling
detected new, lower deviance
In addition: Warning messages:
1: In commonArgs(par, fn, control, environment()) : maxfun < 10 *
length(par)^2 is not recommended.
2: In optwrap(optimizer, devfun, x#theta, lower = x#lower, calc.derivs
= TRUE, : convergence code 1 from bobyqa: bobyqa -- maximum number of function evaluations exceeded
I found this thread (Error when estimating CI for GLMM using confint()), which said that I need to reduce the “devtol” parameter by setting a different profile. But doing so results in the same error:
lmer1_devtol = profile(lmer1, devtol = 1e-7)
Error in zeta(shiftpar, start = opt[seqpar1][-w]) : profiling
detected new, lower deviance
In addition: Warning messages:
1: In commonArgs(par, fn, control, environment()) : maxfun < 10 *
length(par)^2 is not recommended.
2: In optwrap(optimizer, devfun, x#theta, lower = x#lower, calc.derivs
= TRUE, : convergence code 1 from bobyqa: bobyqa -- maximum number of function evaluations exceeded
add_ci
I found the function “add_ci” but this again resulted in another error:
predictions_ci = ciTools::add_ci(df, lmer1,
alpha = 0.05)
Error in levelfun(r, n, allow.new.levels = allow.new.levels) : new
levels detected in newdata
I then set the new “allow.new.levels” parameter to TRUE like in the description of the prediction function, but this parameter seems not to be carried through:
predictions_ci = ciTools::add_ci(df, lmer1,
alpha = 0.05,
allow.new.levels = TRUE)
Error in levelfun(r, n, allow.new.levels = allow.new.levels) : new
levels detected in newdata
Diag
I found a method to calculate CI intervals for the sleepstudy data, which uses a matrix conversion with diag.
Designmat <- model.matrix(as.formula("y ~ x1 * poly(x2, 5)")[-2], df)
predvar <- diag(Designmat %*% vcov(lmer1) %*% t(Designmat))
#With new data
newdat = df
newdat$pred <- predict(lmer1, newdat, allow.new.levels = TRUE)
Designmat <- model.matrix(formula(lmer1)[-2], newdat)
But the diag method does not work for such large datasets.
bootMer
As said earlier, the boostrapping of the confidence interval with bootMer is taking too much time for this subset of data (I started it 1 day ago and it is still running). I tried to use some parallel processing with the sleepstudy sample data but this could not increase the speed dramatically, so I would assume it will have the same effect on my large dataset.
merBoot <- bootMer(lmer1, predict, nsim = 1000, re.form = NA)
Others
I have read through all these post (and more), but none of them could help me to get the CI in reasonable time for my case. But maybe I have overseen something.
https://stats.stackexchange.com/questions/344012/confidence-intervals-from-bootmer-in-r-and-pros-cons-of-different-interval-type
https://stats.stackexchange.com/questions/117641/how-trustworthy-are-the-confidence-intervals-for-lmer-objects-through-effects-pa
How to get coefficients and their confidence intervals in mixed effects models?
Error when estimating CI for GLMM using confint()
https://stats.stackexchange.com/questions/235018/r-extract-and-plot-confidence-intervals-from-a-lmer-object-using-ggplot
How to get confidence intervals for lmer object?
Confidence intervals for the predicted probabilities from glmer object, error with bootMer
https://rdrr.io/cran/ciTools/man/add_ci.lmerMod.html
Error when estimating Confidence interval in lme4
https://fromthebottomoftheheap.net/2018/12/10/confidence-intervals-for-glms/
https://cran.r-project.org/web/packages/merTools/vignettes/Using_predictInterval.html
https://drewtyre.rbind.io/classes/nres803/week_12/lab_12/
Unsurprising to me but unfortunate for you, nonconvergence of mixed model estimation and difficulty in generating confidence intervals results from the misuse of a linear model for data with a limited dependent variable. "Despite these warnings, the model is still fitted" is a dangerous practice, as iterations are not to be used from predictions if not converged. As you described, the dependent variable (y) represents the probability of an event, which is a continuous variable between zero and one. Using a linear model to predict probability constitutes a linear probability regression, which requires censoring predicted outcomes (e.g. forcing all predicted values greater than .99 to be .99 while forcing all predicted values smaller than .01 to be .01) and adjusting for heterogenous variances using weighted least squares (see https://bookdown.org/ccolonescu/RPoE4/heteroskedasticity.html). Having continuous variables produce both fixed and random effects also burden the convergence, while some or all the random effects of continuous variables may not be necessary. The use of weights can be also problematic.
Instead of a linear probability regression, beta regression works best for dependent variables which are proportions and probabilities. Beta regression without random effects is done in betareg::betareg(). glmmTMB::glmmTMB() handles beta regression with random effects. Start from a simple setting where only the intercept has random effects such as
glmmTMB(y ~ 1 + x1 * poly(x2, 5) + c2 + (1 | c1), family = list(family = "beta", link ="logit"), data = df)
You may compare the result with glmer() and lmer()
glmer(y ~ 1 + x1 * poly(x2, 5) + c2 + (1 | c1), family = gaussian(link = "logit"), data = df)
lmer(log(y/(1-y)) ~ 1 + x1 * poly(x2, 5) + c2 + (1 | c1), data = df)
glmer() and lmer() with the above specifications are equivalent, and both assume that predicting log(y/(1-y)) has normal residuals, while glmmTMB() assumes that y follows a gamma distribution. lmer() results are easier to explain and receive wider support from other packages, since they are linear models. On the other hand, glmmTMB() may fit better according to AIC, BIC, and log likelihood. Note that all three requires y strictly in (0, 1) noninclusive. To include occasional zeros and ones, manipulate observations at both boundaries by introducing a small tolerance usually equal to half of the smaller distance from a boundary to its closest observed value (see https://stats.stackexchange.com/questions/109702 and https://graphworkflow.com/eda/bounded01/). For probabilities with either or both of many zeros and ones, zero-, one-, and zero-one–inflated beta regression is fitted via gamlss::gamlss(). See Korosteleva, O. (2019). Advanced regression models with SAS and R. CRC Press.
Add random effects of slopes if necessary according to likelihood ratio tests. Make sure there are enough levels in c1 (e.g. more than 10 different cities) to necessitate mixed effect models. The {glmmTMB} package extends glm() and glmer(). Its alternative {brms} package is built for Bayesian approach. Note that the weights = argument in glmmTMB() as in glm() specifies that values in weights are inversely proportional to the dispersions and are not automatically scaled to sum to one unless integer values which specifies number of observation units. Therefore, you need to investigate what w1 stands for and evaluate how to use it in modeling.
merTools::predictInterval() generates many kinds of intervals for mixed models, some comparable to confidence intervals and prediction intervals in linear models without random effects. However, it supports lmer() model objects only. See https://cran.r-project.org/web/packages/merTools/vignettes/merToolsIntro.html and https://cran.r-project.org/web/packages/merTools/vignettes/Using_predictInterval.html.
predictInterval(lmer(), include.resid.var = F) includes uncertainty from both fixed and random effects of all coefficients including the intercept but excludes variation from multiple measurements of the same group or individual. This can be considered similar to prediction intervals of linear models without random effects. predictInterval(lmer(), include.resid.var = F, fix.intercept.variance = T) generates shorter CI than above by accounting for covariance between the fixed and random effects of the intercept. predictInterval(lmer(), include.resid.var = F, ignore.fixed.terms = "(Intercept)") also shortens CI by removing uncertainty from the fixed effect of the intercept. If there are no random slopes other than random intercept, the last two methods are comparable to confidence intervals of of linear models without random effects. confint(lmear()) and confint(profile(lmear())) generates confidence intervals of modal parameters such as a slope, so they do not produce confidence intervals of predicted outcomes.
You may also find the following functions and packages useful for generating CIs of mixed effect models.
ggeffect() {ggeffects} predictions() {marginaleffects} and margins() prediction() {margins} {predictions}
They can produce predictions averaged over observed distribution of covariates, instead of making predictions by holding some predictors at specific values such as means or modes which can be misleading and not useful.
Related
I am currently trying to generate a general additive model in R using a response variable and three predictor variables. One of the predictors is linear, and the dataset consists of 298 observations.
I have run the following code to generate a basic GAM:
GAM <- gam(response~ linearpredictor+ s(predictor2) + s(predictor3), data = data[2:5])
This produces a model with 18 degrees of freedom and seems to substantially overfit the data. I'm wondering how I might generate a GAM that maximizes smoothness and predictive error. I realize that each of these features is going to come at the expense of the other, but is there good a way to find the optimal model that doesn't overfit?
Additionally, I need to perform leave one out cross validation (LOOCV), and I am not sure how to make sure that gam() does this in the MGCV package. Any help on either of these problems uld be greatly appreciated. Thank you.
I've run this to generate a GAM, but it overfits the data.
GAM <- gam(response~ linearpredictor+ s(predictor2) + s(predictor3), data = data[2:5])
I have also generated 1,000,000 GAMs with varying combinations of smoothing parameters and ranged the maximum degrees of freedom allowed from 10 (as shown in the code below) to 19. The variable "combinations2" is a list of all 1,000,000 combinations of smoothers I selected. This code is designed to try and balance degrees of freedom and AIC score. It does function, but I'm not sure that I'm actually going to be able to find the optimal model from this. I also cannot tell how to make sure that it uses LOOCV.
BestGAM <- gam(response~ linearpredictor+ predictor2+ predictor3, data = data[2:5])
for(i in 1:100000){
PotentialGAM <- gam(response~ linearpredictor+ s(predictor2) + s(predictor3), data = data[2:5], sp=c(combinations2[i,]$Var1,combinations2[i,]$Var2))
if (AIC(PotentialGAM,BestGAM)$df[1] <= 10 & AIC(PotentialGAM,BestGAM)$AIC[1] < AIC(PotentialGAM,BestGAM)$AIC[2]){
BestGAM <<- PotentialGAM
listNumber <- i
}
}
You are fitting your GAM using generalised cross validation (GCV) smoothness selection. GCV is a way to get around the invariance problem of ordinary cross validation (OCV; what you also call LOOCV) when estimating GAMs. Note that GCV is the same as OCV on a rotated version of the fitting problem (rotating y - Xβ by Q, any orthogonal matrix), and while when fitting with GCV {mgcv} doesn't actually need to do the rotation and the expected GCV score isn't affected by the rotation, GCV is just OCV (wood 2017, p. 260)
It has been shown that GCV can undersmooth (resulting in more wiggly models) as the objective function (GCV profile) can become flat around the optimum. Instead it is preferred to estimate GAMs (with penalized smooths) using REML or ML smoothness selection; add method = "REML" (or "ML") to your gam() call.
If the REML or ML fit is as wiggly as the GCV one with your data, then I'd be likely to presume gam() is not overfitting, but that there is something about your response data that hasn't been explained here (are the data ordered in time, for example?)
As to your question
how I might generate a GAM that maximizes smoothness and [minimize?] predictive error,
you are already doing that using GCV smoothness selection and for a particular definition of "smoothness" (in this case it is squared second derivatives of the estimated smooths, integrated over the range of the covariates, and summed over smooths).
If you want GCV but smoother models, you can increase the gamma argument above 1; gamma 1.4 is often used for example, which means that each EDF costs 40% more in the GCV criterion.
FWIW, you can get the LOOCV (OCV) score for your model without actually fitting 288 GAMs through the use of the influence matrix A. Here's a reproducible example using my {gratia} package:
library("gratia")
library("mgcv")
df <- data_sim("eg1", seed = 1)
m <- gam(y ~ s(x0) + s(x1) + s(x2) + s(x3), data = df, method = "REML")
A <- influence(m)
r <- residuals(m, type = "response")
ocv_score <- mean(r^2 / (1 - A))
I'm trying to tune hyperparameters epsilon and cost using the tune function in e1071, but I keep getting this error whenever I try to expand the ranges of values that I want to test:
"Error in predict.svm(ret, xhold, decision.values = TRUE) :
Model is empty!"
I'm dealing with the regression application, not a classification one, and the data I'm using is for density profiles, where "x" describes the position alongside a board and "y corresponds to the value of the density measured. This is the code I'm using:
model <- tune(svm, y~x, data = profiles, ranges = list(cost = 2^(0:10), epsilon = 10^(-10:0), tunecontrol = tune.control(cross = 5))
The data is all numeric (doubles) and the problem seems to occur only when I try to test such a large range of values. Has anybody experienced a similar issue?
It may be the range of your cost and epsilon values. I ran into the same problem, i.e. svm regression with all numeric data. I was tuning using a range of epsilon values from .1 to 10 and I was getting the model empty error. I then reduced the epsilon range from .1 to 1 and it was able to converge with no errors. There is probably some interaction between cost and epsilon that generates unstable predictions, i.e. high cost and high epsilon are not kosher.
I am using "glmnet" package (in R) mostly to perform regularized linear regression.
However I am wondering if it can perform LASSO-type regressions with non-negative (integer) continuous (dependent) outcome variable.
I can use family = poisson, but the outcome variable is not specifically "count" variable. It is just a continuous variable with lower limit 0.
I aware of "lower.limits" function, but I guess it is for covariates (independent variables). (Please correct me if my understanding of this function not right.)
I look forward to hearing from you all! Thanks :-)
You are right that setting lower limit in glmnet is meant for covariates. Poisson will set a lower limit to zero because you exponentiate to get back the "counts".
Going along those lines, most likely it will work if you transform your response variable. One quick way is to take the log of your response variable, do the fit and transform it back, this will ensure that it's always positive. you have to deal with zeros
An alternative is a power transformation. There's a lot to think about and I can only try a two parameter box-cox with a dataset since you did not provide yours:
library(glmnet)
library(mlbench)
library(geoR)
data(BostonHousing)
data = BostonHousing
data$chas=as.numeric(data$chas)
# change it to min 0 and max 1
data$medv = (data$medv-min(data$medv))/diff(range(data$medv))
Then here I use a quick approximation via pca (without fitting all the variables) to get the suitable lambda1 and lambda2 :
bcfit = boxcoxfit(object = data[,14],
xmat = prcomp(data[,-14],scale=TRUE,center=TRUE)$x[,1:2],
lambda2=TRUE)
bcfit
Fitted parameters:
lambda lambda2 beta0 beta1 beta2 sigmasq
0.42696313 0.00001000 -0.83074178 -0.09876102 0.08970137 0.05655903
Convergence code returned by optim: 0
Check the lambda2, it is the one thats critical for deciding whether you get a negative value.. It should be rather small.
Create the functions to power transform:
bct = function(y,l1,l2){((y+l2)^l1 -1)/l1}
bctinverse = function(y,l1,l2){(y*l1+1)^(1/l1) -l2}
Now we transform the response:
data$medv_trans = bct(data$medv,bcfit$lambda[1],bcfit$lambda[2])
And fit glmnet:
fit = glmnet(x=as.matrix(data[,1:13]),y=data$medv_trans,nlambda=500)
Get predictions over all lambdas, and you can see there's no negative predictions once you transform back:
pred = predict(fit,as.matrix(data[,1:13]))
range(bctinverse(pred,bcfit$lambda[1],bcfit$lambda[2]))
[1] 0.006690685 0.918473356
And let's say we do a fit with cv:
fit = cv.glmnet(x=as.matrix(data[,1:13]),y=data$medv_trans)
pred = predict(fit,as.matrix(data[,1:13]))
pred_transformed = bctinverse(pred,bcfit$lambda[1],bcfit$lambda[2]
plot(data$medv,pred_transformed,xlab="orig response",ylab="predictions")
I would like to estimate the standard error and variance of predictions from a linear mixed model. I'm using merTools::predictInterval to estimate prediction intervals because I want to include some of the uncertainty in the random effects (in addition to the uncertainty in the fixed effects). Is it acceptable to use the simulations from merTools::predictInterval to estimate the se and variance of predictions? If so, how should I calculate them? I can think of 2 ways:
To get variance corresponding to the prediction interval (i.e. including residual variance), I would first get the simulated predictions:
predictions <- merTools::predictInterval(...,
include.resid.var = TRUE,
returnSims = TRUE)
1. Then I could estimate variance using the normal approximation (calculate the distance between the fit and the upper/lower interval and then divide that by 1.96):
var1 <- ((predictions$upr - predictions$lwr)/2/1.96)^2
2. Or I could just take the variance of the simulated values:
var2 <- apply(X = attr(x = predictions, which = 'sim.results'), MARGIN = 1, FUN = var)
The SE would then be the square root of the variance. To get the SE and/or variance relating to the confidence interval, I could repeat this with include.resid.var = FALSE in the merTools::predictInterval call.
Is either of these methods acceptable? Is one preferable to the other?
I'm trying to understand the function lmer. I've found plenty of information about how to use the command, but not much about what it's actually doing (save for some cryptic comments here: http://www.bioconductor.org/help/course-materials/2008/PHSIntro/lme4Intro-handout-6.pdf). I'm playing with the following simple example:
library(data.table)
library(lme4)
options(digits=15)
n<-1000
m<-100
data<-data.table(id=sample(1:m,n,replace=T),key="id")
b<-rnorm(m)
data$y<-rand[data$id]+rnorm(n)*0.1
fitted<-lmer(b~(1|id),data=data,verbose=T)
fitted
I understand that lmer is fitting a model of the form Y_{ij} = beta + B_i + epsilon_{ij}, where epsilon_{ij} and B_i are independent normals with variances sigma^2 and tau^2 respectively. If theta = tau/sigma is fixed, I computed the estimate for beta with the correct mean and minimum variance to be
c = sum_{i,j} alpha_i y_{ij}
where
alpha_i = lambda/(1 + theta^2 n_i)
lambda = 1/[\sum_i n_i/(1+theta^2 n_i)]
n_i = number of observations from group i
I also computed the following unbiased estimate for sigma^2:
s^2 = \sum_{i,j} alpha_i (y_{ij} - c)^2 / (1 + theta^2 - lambda)
These estimates seem to agree with what lmer produces. However, I can't figure out how log likelihood is defined in this context. I calculated the probability density to be
pd(Y_{ij}=y_{ij}) = \prod_{i,j}[f_sigma(y_{ij}-ybar_i)]
* prod_i[f_{sqrt(sigma^2/n_i+tau^2)}(ybar_i-beta) sigma sqrt(2 pi/n_i)]
where
ybar_i = \sum_j y_{ij}/n_i (the mean of observations in group i)
f_sigma(x) = 1/(sqrt{2 pi}sigma) exp(-x^2/(2 sigma)) (normal density with sd sigma)
But log of the above is not what lmer produces. How is log likelihood computed in this case (and for bonus marks, why)?
Edit: Changed notation for consistency, striked out incorrect formula for standard deviation estimate.
The links in the comments contained the answer. Below I've put what the formulae simplify to in this simple example, since the results are somewhat intuitive.
lmer fits a model of the form , where and are independent normals with variances and respectively. The joint probability distribution of and is therefore
where
.
The likelihood is obtained by integrating this with respect to (which isn't observed) to give
where is the number of observations from group , and is the mean of observations from group . This is somewhat intuitive since the first term captures spread within each group, which should have variance , and the second captures the spread between groups. Note that is the variance of .
However, by default (REML=T) lmer maximises not the likelihood but the "REML criterion", obtained by additionally integrating this with respect to to give
where is given below.
Maximising likelihood (REML=F)
If is fixed, we can explicitly find the and which maximise likelihood. They turn out to be
Note has two terms for variation within and between groups, and is somewhere between the mean of and the mean of depending on the value of .
Substituting these into likelihood, we can express the log likelihood in terms of only:
lmer iterates to find the value of which minimises this. In the output, and are shown in the fields "deviance" and "logLik" (if REML=F) respectively.
Maximising restricted likelihood (REML=T)
Since the REML criterion doesn't depend on , we use the same estimate for as above. We estimate to maximise the REML criterion:
The restricted log likelihood is given by
In the output of lmer, and are shown in the fields "REMLdev" and "logLik" (if REML=T) respectively.