Nothing
specification_control <- function(){
.mmes_control_build(task=list(objective="ml"),
convergence=list(maxIters=200L,likelihoodTol=1e-8,parameterTol=1e-8),
diagnostics=list(verbose=FALSE,dateWarning=FALSE))
}
specification_data <- function(){
set.seed(913)
data <- expand.grid(x=seq(-1.8,1.8,length.out=7),
Name=factor(letters[1:3]),Env=factor(c("A","B")))
data$y <- 1+.3*(data$Env=="B")+1.7*tanh(.9*data$x)+
rnorm(nrow(data),sd=.18)
data
}
test_that("parameter declarations and formula markers are strict",{
expect_s3_class(parameter(fixed=~Env,random=~vsm(ism(Name)),link="log"),
"sommer_parameter")
expect_error(parameter(fixed=y~x),"one-sided")
expect_error(parameter(link=c("log","identity")),"link")
expect_error(parameter(link="logit"),"link")
expect_error(curve_term("curve"),"formula marker")
expect_error(curve_term(1),"curve name")
})
test_that("ordinary fixed-only models match lm and glm exactly",{
data <- specification_data()
control <- specification_control()
fit <- .mmes_specification_fit(y~Env+x,data=data,control=control)
oracle <- lm(y~Env+x,data)
expect_equal(coef(fit),coef(oracle))
expect_equal(fitted(fit),fitted(oracle))
expect_equal(as.numeric(logLik(fit)),as.numeric(logLik(oracle)))
expect_equal(vcov(fit),vcov(oracle)*(nrow(data)-3)/nrow(data))
expect_equal(predict(fit,data[1:3,]),predict(oracle,data[1:3,]))
expect_equal(residuals(fit),residuals(oracle))
expect_equal(nobs(fit),nrow(data))
expect_s3_class(summary(fit),"summary.mmes.specification")
control$task$objective <- "reml"
reml <- .mmes_specification_fit(y~Env+x,data=data,control=control)
expect_equal(as.numeric(logLik(reml)),as.numeric(logLik(oracle,REML=TRUE)))
expect_equal(vcov(reml),vcov(oracle))
set.seed(711)
data$z <- rbinom(nrow(data),1,plogis(.2+.8*data$x))
fit <- .mmes_specification_fit(z~Env+x,data=data,family=binomial(),control=control)
oracle <- glm(z~Env+x,data=data,family=binomial())
expect_equal(coef(fit),coef(oracle))
expect_equal(fitted(fit),fitted(oracle))
expect_equal(predict(fit,data,type="link"),predict(oracle,data,type="link"))
expect_equal(as.numeric(logLik(fit)),as.numeric(logLik(oracle)))
expect_equal(vcov(fit),vcov(oracle))
expect_equal(residuals(fit,type="pearson"),residuals(oracle,type="pearson"))
expect_equal(dim(simulate(fit,nsim=2,seed=91)),c(nrow(data),2L))
})
test_that("nonlinear parameter regressions agree with an independent nls oracle",{
data <- specification_data()
control <- specification_control()
declarations <- list(a=parameter(),k=parameter(fixed=~Env,link="log"))
fit <- .mmes_specification_fit(y~Env+a*tanh(k*x),data=data,
parameters=declarations,start=c(a=1.5,k=1),control=control)
oracle <- nls(y~b0+b1*(Env=="B")+a*tanh(exp(k0+k1*(Env=="B"))*x),
data=data,start=list(b0=1,b1=.3,a=1.5,k0=0,k1=0))
expect_true(fit$convergence)
expect_equal(unname(coef(fit)),unname(coef(oracle)),tolerance=2e-4)
expect_equal(fitted(fit),as.numeric(fitted(oracle)),tolerance=2e-5)
expect_equal(as.numeric(logLik(fit)),as.numeric(logLik(oracle)),tolerance=1e-7)
expect_equal(predict(fit,data[1:4,]),fitted(fit)[1:4],tolerance=1e-10)
expect_equal(predict(fit,target="zero"),fitted(fit))
expect_match(summary(fit)$derivativeMethod,"numerical")
expect_true(all(is.finite(vcov(fit))))
explicit <- .mmes_specification_fit(y~Env+a*tanh(k*x),data=data,
parameters=declarations,start=list(coefficients=coef(fit)),control=control)
expect_equal(as.numeric(logLik(explicit)),as.numeric(logLik(fit)),tolerance=1e-7)
})
test_that("exact additive Gaussian integration preserves relationship DG covariance",{
data <- specification_data()
A <- matrix(c(1,.2,.1,.2,1,.3,.1,.3,1),3,3,
dimnames=list(levels(data$Name),levels(data$Name)))
Ai <- Matrix::Matrix(solve(A),sparse=TRUE)
attr(Ai,"inverse") <- TRUE
control <- specification_control()
fit <- .mmes_specification_fit(y~Env+a*tanh(k*x),
random=~vsm(dsm(Env),ism(Name),Gu=Ai,sigma2=.03,fixedSigma2=TRUE),
data=data,parameters=list(a=parameter(),k=parameter(link="log")),
start=c(a=1.5,k=1),control=control)
# The independent oracle uses the retained common covariance descriptors,
# but independently constructs and factors the marginal observation law.
z <- fit$randomRecords[[1]]$z
g <- fit$randomCovariances[[1]]
zero <- predict(fit,target="zero",type="link")
v <- z %*% g %*% t(z)+diag(fit$dispersion,nrow(data))
r <- data$y-zero
oracle <- -.5*(nrow(data)*log(2*pi)+as.numeric(determinant(v,TRUE)$modulus)+
sum(r*solve(v,r)))
expect_equal(as.numeric(logLik(fit)),oracle,tolerance=1e-8)
expect_equal(g,kronecker(diag(diag(g)[c(1,4)]),A),tolerance=1e-9)
expect_match(fit$likelihoodMethod,"exact")
expect_equal(predict(fit,data[1:4,]),fitted(fit)[1:4],tolerance=1e-8)
expect_equal(predict(fit,target="population"),zero,tolerance=1e-8)
})
test_that("log-rate random effects use genuine conditional-mode Laplace integration",{
skip_on_cran()
data <- specification_data()
control <- specification_control()
fit <- .mmes_specification_fit(y~Env+a*tanh(k*x),data=data,
parameters=list(a=parameter(),k=parameter(link="log",
random=~vsm(ism(Name),sigma2=.025,fixedSigma2=TRUE))),
start=c(a=1.5,k=1),control=control)
b <- coef(fit)
conditional <- function(u){
eta <- b[1]+b[2]*(data$Env=="B")+
b[3]*tanh(exp(b[4]+u[data$Name])*data$x)
-sum(dnorm(data$y,eta,sqrt(fit$dispersion),log=TRUE))+
sum(u^2)/(.025*2)
}
oracle <- optim(rep(0,3),conditional,method="BFGS",
control=list(reltol=1e-12,maxit=1000))
h <- optimHess(oracle$par,conditional)
likelihood <- -oracle$value-3/2*log(.025)-.5*as.numeric(determinant(h,TRUE)$modulus)
expect_equal(fit$randomMode[[1]],oracle$par,tolerance=2e-5)
expect_equal(as.numeric(logLik(fit)),unname(likelihood),tolerance=2e-5)
expect_match(fit$likelihoodMethod,"Laplace")
expect_equal(predict(fit,data[1:5,]),fitted(fit)[1:5],tolerance=1e-8)
population <- predict(fit,target="population")
expect_true(all(is.finite(population)))
expect_gt(max(abs(population-predict(fit,target="zero"))),1e-5)
unseen <- data[1:2,]
unseen$Name <- "unseen"
expect_error(predict(fit,unseen),"Unseen")
expect_true(all(is.finite(predict(fit,unseen,target="zero"))))
expect_error(.mmes_specification_fit(y~Env+a*tanh(k*x),data=data,
parameters=list(a=parameter(),k=parameter(link="log",
random=~vsm(ism(Name),sigma2=.025,fixedSigma2=TRUE))),
start=c(a=1.5,k=1),control=control,
specificationMethod=list(integration="exact",approximation="likelihood")),
"Exact integration")
})
test_that("nonlinear Bernoulli and Poisson likelihoods match independent GLMs",{
data <- specification_data()
control <- specification_control()
set.seed(512)
data$binary <- rbinom(nrow(data),1,plogis(.1+.6*data$x))
data$count <- rpois(nrow(data),exp(.2+.3*data$x))
for(family in list(binomial(),poisson())){
response <- if(family$family=="binomial") "binary" else "count"
formula <- as.formula(paste(response,"~a*x"))
fit <- .mmes_specification_fit(formula,data=data,family=family,
parameters=list(a=parameter()),start=c(a=.4),control=control)
oracle <- glm(as.formula(paste(response,"~x")),data=data,family=family)
expect_equal(unname(coef(fit)),unname(coef(oracle)),tolerance=2e-5)
expect_equal(as.numeric(logLik(fit)),as.numeric(logLik(oracle)),tolerance=1e-8)
}
})
test_that("nonGaussian additive random effects use Laplace, not a working fit",{
skip_on_cran()
data <- specification_data()
set.seed(761)
data$binary <- rbinom(nrow(data),1,plogis(.1+.7*data$x))
fit <- .mmes_specification_fit(binary~a*x,data=data,family=binomial(),
random=~vsm(ism(Name),sigma2=.2,fixedSigma2=TRUE),
parameters=list(a=parameter()),start=c(a=.5),
control=specification_control())
conditional <- function(u){
eta <- coef(fit)[1]+coef(fit)[2]*data$x+u[data$Name]
-sum(dbinom(data$binary,1,plogis(eta),log=TRUE))+sum(u^2)/.4
}
oracle <- optim(rep(0,3),conditional,method="BFGS",
control=list(reltol=1e-12,maxit=1000))
h <- optimHess(oracle$par,conditional)
likelihood <- -oracle$value-1.5*log(.2)-.5*as.numeric(determinant(h,TRUE)$modulus)
expect_equal(as.numeric(logLik(fit)),unname(likelihood),tolerance=2e-5)
expect_match(fit$likelihoodMethod,"Laplace")
})
test_that("Gamma log likelihood and fitted dispersion match independent ML",{
skip_on_cran()
data <- specification_data()
set.seed(365)
data$positive <- rgamma(nrow(data),shape=5,
scale=exp(.2+.3*data$x)/5)
control <- specification_control()
fit <- .mmes_specification_fit(positive~a*x,data=data,family=Gamma("log"),
parameters=list(a=parameter()),start=c(a=.3),control=control,rcov=~units)
oracle <- glm(positive~x,data=data,family=Gamma("log"),
control=glm.control(epsilon=1e-12,maxit=100L))
phi <- optimize(function(logPhi)
-sum(dgamma(data$positive,shape=exp(-logPhi),
scale=fitted(oracle)*exp(logPhi),log=TRUE)),
interval=c(-10,5),tol=1e-10)
expect_lt(max(abs(unname(coef(fit))-unname(coef(oracle)))),5e-5)
expect_equal(fit$dispersion,exp(phi$minimum),tolerance=2e-5)
expect_equal(as.numeric(logLik(fit)),-phi$objective,tolerance=1e-7)
ordinary <- .mmes_specification_fit(positive~x,data=data,family=Gamma("log"),
control=control,rcov=~units)
expect_equal(ordinary$dispersion,exp(phi$minimum),tolerance=2e-5)
expect_equal(as.numeric(logLik(ordinary)),-phi$objective,tolerance=1e-7)
expect_error(.mmes_specification_fit(positive~a*x,data=data,
family=Gamma("identity"),parameters=list(a=parameter()),
start=c(a=.3),control=control),"Unsupported native Laplace family/link")
mixed <- .mmes_specification_fit(positive~a*x,data=data,family=Gamma("log"),
random=~vsm(ism(Name),sigma2=.025,fixedSigma2=TRUE),
parameters=list(a=parameter()),start=c(a=.3),control=control,rcov=~units)
conditional <- function(u){
mu <- exp(coef(mixed)[1]+coef(mixed)[2]*data$x+u[data$Name])
-sum(dgamma(data$positive,shape=1/mixed$dispersion,
scale=mu*mixed$dispersion,log=TRUE))+sum(u^2)/.05
}
reference <- optim(rep(0,3),conditional,method="BFGS",
control=list(reltol=1e-12,maxit=1000L))
h <- optimHess(reference$par,conditional)
expect_equal(as.numeric(logLik(mixed)),
-reference$value-1.5*log(.025)-.5*as.numeric(determinant(h,TRUE)$modulus),
tolerance=2e-5,ignore_attr=TRUE)
})
test_that("uncertainty is post-fit and its evaluator survives serialization",{
data <- specification_data()
fit <- .mmes_specification_fit(y~a*tanh(k*x),data=data,
parameters=list(a=parameter(),k=parameter(link="log")),
start=c(a=1.5,k=1),control=specification_control())
expect_null(fit$parameterCovariance)
expect_null(fit$hessian)
restored <- unserialize(serialize(fit,NULL))
expect_equal(vcov(restored),vcov(fit),tolerance=1e-9)
expect_equal(.spec_evaluate(restored$optimizer$par,restored$likelihoodContext),
-as.numeric(logLik(fit)),tolerance=1e-9)
})
test_that("one shared observation mask respects explicit NA policies",{
skip_on_cran()
data <- specification_data()
data$y[1] <- NA
data$x[2] <- NA
data$Name[3] <- NA
control <- specification_control()
args <- list(fixed=y~a*tanh(k*x),data=data,
parameters=list(a=parameter(),k=parameter(link="log",
random=~vsm(ism(Name),sigma2=.025,fixedSigma2=TRUE))),
start=c(a=1.5,k=1),control=control)
fit <- do.call(.mmes_specification_fit,args)
expect_equal(fit$obsInfo$excluded,1:3)
expect_equal(nobs(fit),nrow(data)-3L)
expect_equal(length(fitted(fit)),nrow(data)-3L)
expect_error(do.call(.mmes_specification_fit,c(args,list(naMethodX="fail"))),"naMethodX")
expect_error(do.call(.mmes_specification_fit,c(args,list(naMethodY="fail"))),"naMethodY")
expect_error(do.call(.mmes_specification_fit,c(args,list(naMethodRandom="fail"))),"naMethodRandom")
expect_error(do.call(.mmes_specification_fit,c(args,list(naMethodR="impute"))),"only")
expect_error(do.call(.mmes_specification_fit,c(args,list(familyParameters=list(phi=.2)))),
"familyParameters")
})
test_that("conditional scores certify requested mode tolerance and controls are owned",{
skip_on_cran()
data <- specification_data()
control <- specification_control()
control$task$nonlinear$mode$scoreTol <- 2e-6
control$task$nonlinear$mode$stepTol <- 3e-10
fit <- .mmes_specification_fit(y~a*tanh(k*x),data=data,
parameters=list(a=parameter(),k=parameter(link="log",
random=~vsm(ism(Name),sigma2=.025,fixedSigma2=TRUE))),
start=c(a=1.5,k=1),control=control,rcov=~units)
expect_lte(fit$modeScore,2e-6)
expect_equal(fit$likelihoodContext$control$task$nonlinear$mode$stepTol,3e-10)
expect_equal(fit$control$convergence$likelihoodTol,control$convergence$likelihoodTol)
})
test_that("public wrapper supports homogeneous residuals Gamma and numerical controls",{
data <- specification_data()
set.seed(365)
data$positive <- rgamma(nrow(data),shape=5,
scale=exp(.2+.3*data$x)/5)
control <- mmes_control(optimizer=list(gradient="central"),
convergence=list(maxIters=100L,likelihoodTol=1e-8,parameterTol=1e-8),
mode=list(tol=2e-6,stepTol=3e-10))
fit <- mmes(positive~a*x,data=data,family=Gamma("log"),rcov=~units,
parameters=list(a=parameter()),start=c(a=.3),
method=mmes_method(objective="ml"),control=control)
expect_s3_class(fit,"mmes.specification")
expect_true(is.finite(logLik(fit)))
expect_equal(predict(fit,data[1:3,]),fitted(fit)[1:3],tolerance=1e-9)
bad <- mmes_control(optimizer=list(gradient="analytic"))
expect_error(mmes(positive~a*x,data=data,family=Gamma("log"),rcov=~units,
parameters=list(a=parameter()),start=c(a=.3),
method=mmes_method(objective="ml"),control=bad),"analytic")
})
test_that("invalid options, starts, expressions, and prediction levels fail explicitly",{
data <- specification_data()
control <- specification_control()
fit <- function(...,start=c(a=1.5,k=1),formula=y~a*tanh(k*x))
.mmes_specification_fit(formula,data=data,control=control,
parameters=list(a=parameter(),k=parameter(link="log")),start=start,...)
expect_error(fit(unknown=1),"Unknown")
expect_error(fit(start=c(a=1,k=-1)),"positive")
expect_error(fit(start=c(a=1)),"exactly")
expect_error(fit(formula=y~a*get("x")+k),"Unsupported")
expect_error(fit(formula=y~a*tanh(k*missing)),"Unsupported")
expect_error(fit(rcov=~Env),"independent residuals")
control$task$objective <- "reml"
expect_error(fit(),"require task")
control$task$objective <- "ml"
data$x[1] <- NA
expect_error(fit(naMethodX="fail"),"Missing values")
})
Any scripts or data that you put into this service are public.
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.