--- title: "Choosing link, variance and covariance functions" author: "Prof. Wagner Hugo Bonat" date: "`r paste('mcglm', packageVersion('mcglm'), Sys.Date())`" vignette: > %\VignetteIndexEntry{Choosing link, variance and covariance functions} %\VignetteEngine{knitr::rmarkdown} %\VignetteEncoding{UTF-8} --- ```{r setup, include=FALSE} ##---------------------------------------------------------------------- library(knitr) ``` **** To install the stable version of [`mcglm`][], use `devtools::install_git()`. For more information, visit [mcglm/README]. ```{r, eval=FALSE} library(devtools) install_git("bonatwagner/mcglm") ``` ```{r, eval=FALSE, error=FALSE, message=FALSE, warning=FALSE} library(mcglm) packageVersion("mcglm") ``` ##### Abstract The `mcglm` package implements the multivariate covariance generalized linear models (McGLMs) proposed by Bonat and J$\o$rgensen (2016). The core fit function `mcglm` is employed for fitting a set of models. In this introductory vignette we restrict ourselves to model independent data, although a simple model for longitudinal data analysis in the Gaussian case is also presented. We present models to deal with continuous, binomial/bounded and count univariate response variables. We explore the specification of different link, variance and covariance functions. **** ## Regression models for continuous data Consider a simple regression model, for univariate and independent Gaussian data: $$Y \sim N(X \beta, \tau_0 Z_0).$$ ```{r, warning = FALSE, message = FALSE} # Loading extra packages require(mcglm) require(Matrix) require(mvtnorm) require(tweedie) # Setting the seed set.seed(2503) # Fixed component x1 <- seq(-1,1, l = 100) X <- model.matrix(~ x1) mu1 <- mcglm::mc_link_function(beta = c(1,0.8), X = X, offset = NULL, link = "identity") # Random component y1 <- rnorm(100, mu1$mu, sd = 0.5) # Data structure data <- data.frame("y1" = y1, "x1" = x1) # Matrix linear predictor Z0 <- mc_id(data) # Fit fit1.id <- mcglm(linear_pred = c(y1 ~ x1), matrix_pred = list(Z0), data = data) ``` The `mcglm` package offers the following set of `S3-methods` for model summarize. ```{r} print(methods(class = "mcglm")) ``` The traditional `summary` for a fitted model can be obtained by ```{r} summary(fit1.id) ``` The function `summary.mcglm` was designed to be similar to the native `summary` functions for the classes `lm` and `glm`. Extra features were included to describe the dispersion structure. The mean formula call, along with the the `link`, `variance` and `covariance` functions are presented. The parameter estimates are presented in two blocks, the first presents the regression estimates while the second presents the dispersion estimates. In both cases the parameter estimates are summarized by point estimates, standard errors, Z-values and p-values associated with the Wald test whose null hypothesis is defined as $\beta = 0$ and $\tau = 0$ for the mean and dispersion structures, respectively. Finally, the selected algorithm, if the correction term is employed or not and the number of iterations is printed. The same linear regresion model can be fitted by using a different covariance link function, for example the inverse covariance link function. ```{r} # Fit using inverse covariance link function fit1.inv <- mcglm(linear_pred = c(y1 ~ x1), matrix_pred = list(Z0), covariance = "inverse", data = data) ``` Furthermore, we can use a less conventional covariance link function, as the exponential-matrix. ```{r, message=FALSE, warning=FALSE} # Fit using expm covariance link function fit1.expm <- mcglm(linear_pred = c(y1 ~ x1), matrix_pred = list(Z0), covariance = "expm", data = data) ``` The function `mcglm` returns an object of mcglm class, for which we can use the method `coef` to extract the parameters estimates. ```{r} # Comparing estimates using different covariance link functions cbind(coef(fit1.id)$Estimates, coef(fit1.inv)$Estimates, coef(fit1.expm)$Estimates) # Applying the inverse transformation c(coef(fit1.id)$Estimates[3], 1/coef(fit1.inv)$Estimates[3], exp(coef(fit1.expm)$Estimates[3])) ``` Consider an extension of the linear regression models to deal with heteroscedasticity: $$ Y \sim N(X \beta, \tau_0 Z_0 + \tau_1 Z_1),$$ where $Z_0$ is a identity matrix and $Z_1$ is a diagonal matrix whose elements are given by the values of a known covariate. Such a model, can be fitted easily using the `mcglm` package. ```{r} # Mean model set.seed(1811) x1 <- seq(-1,1, l = 100) X <- model.matrix(~ x1) mu1 <- mcglm::mc_link_function(beta = c(1,0.8), X = X, offset = NULL, link = "identity") z1 <- rnorm(100, mean = 0, sd = 0.25) data <- data.frame("id" = 1, "x1" = x1, "z1" = z1) # Matrix linear predictor Z <- mc_dglm(~ z1, id = 'id', data = data) # Covariance model Sigma <- mcglm::mc_matrix_linear_predictor(tau = c(0.2, 0.15), Z = Z) # Simulating the response variable y1 <- rnorm(100, mu1$mu, sd = sqrt(diag(Sigma))) data$y <- y1 # Fitting fit2.id <- mcglm(linear_pred = c(y1 ~ x1), matrix_pred = list(Z), data = data) ``` We can also extend the linear regression model to deal with longitudinal data analysis. The code below presents an example of such a model. ```{r} # Mean model x1 <- seq(-1,1, l = 100) X <- model.matrix(~ x1) mu1 <- mcglm::mc_link_function(beta = c(1,0.8), X = X, offset = NULL, link = "identity") # Data structure data <- data.frame("id" = as.factor(rep(1:10, each = 10)), "x1" = x1) # Covariance model Z0 <- mc_id(data) Z1 <- mc_mixed(~ 0 + id, data = data) Sigma <- mcglm::mc_matrix_linear_predictor(tau = c(0.2, 0.15), Z = c(Z0,Z1)) # Simulating the Response variable y1 <- as.numeric(rmvnorm(1, mean = mu1$mu, sigma = as.matrix(Sigma))) data <- data.frame("y1" = y1, "x1" = x1) # Fit fit3.id <- mcglm(linear_pred = c(y1 ~ x1), matrix_pred = list("resp1" = c(Z0,Z1)), data = data) ``` The model summary ```{r} summary(fit3.id) ``` Note that, the dispersion structure now has two parameters. In that case, the parameter $\tau_1$ represents the longitudinal structure for which we are assuming a compound symmetry model. This model is an equivalent to a random intercept model in the context of Linear Mixed Models (LMMs). ## Regression models for binomial and bounded data The `mcglm` package offers a rich set of models to deal with binomial and bounded response variables. The `logit`, `probit`, `cauchit`, `cloglog`, and `loglog` link functions along with the extended binomial variance function combined with the linear covariance structure, provide a flexible class of models for handling binomial and bounded response variables. The extended binomial variance function is given by $\mu^p (1- \mu)^q$ where the two extra power parameters offer more flexibility to model the relationship between mean and variance. Consider the following simulated dataset. ```{r} # Mean model x1 <- seq(-1,1, l = 500) X <- model.matrix(~ x1) mu1 <- mcglm::mc_link_function(beta = c(1,0.8), X = X, offset = NULL, link = "logit") # Data structure data <- data.frame("x1" = x1) # Covariance model Z0 <- mc_id(data) # Simulating the response variable set.seed(123) data$y <- rbinom(500, prob = mu1$mu, size = 10)/10 ``` The most traditional regression model to deal with binomial data is the logistic regression model that can be fitted using the `mcglm` package using the following code: ```{r} # Fit fit4.logit <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "logit", variance = "binomialP", power_fixed = TRUE, Ntrial = list(rep(10,500)), data = data) ``` It is important to highlight that the response variable collumn should be between $0$ and $1$ and in the case of more than one trial the argument `Ntrial` should be used for fitting the model. The argument `link` specifies the link function whereas the argument `variance` specifies the variance function in that case `binomialP`. The variance function `binomialP` represents a simplification of the extended binomial variance function given by $\mu^p (1- \mu)^p$. Note that, in this example the argument 'power_fixed = TRUE' specifies that the power parameter $p$ will not be estimated, but fixed at the initial value $p = 1$ corresponding to the orthodox binomial variance function. We can easily fit the model using a different link function, for example the `cauchit`. ```{r} fit4.cauchit <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "cauchit", variance = "binomialP", Ntrial = list(rep(10,250)), data = data) ``` We can also estimate the extra power parameter $p$. ```{r} fit4.logitP <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "logit", variance = "binomialP", power_fixed = FALSE, Ntrial = list(rep(10,500)), data = data) ``` Furthermore, we can estimate the two extra power parameters involved in the extended binomial variance function. ```{r} fit4.logitPQ <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "logit", variance = "binomialPQ", power_fixed = FALSE, Ntrial = list(rep(10,500)), control_algorithm = list(tuning = 0.5, max_iter = 100), data = data) ``` The estimation of the extra power parameters involved in the extended binomial variance function is challenging mainly for small data sets. Note that, in this simulated example, we have to control the step-length of the `chaser` algorithm to avoid unrealistic values for the parameters involved in the dispersion structure. To do that, we used the extra argument `control_algorithm` that should be a named list. For a detailed description of the arguments that can be passed to the `control_algorithm` function see `?fit_mcglm`. ## Regression models for count data The analysis of count data in the `mcglm` package relies on the structure of the Poisson-Tweedie distribution. Such a distribution is characterized by the following dispersion function: $$ \nu(\mu, p) = \mu + \tau_0 \mu^p. $$ The power parameter is an index that identify different distributions, examples include the Hermite ($p = 0$), Neyman-Type A ($p = 1$) and the negative binomial ($p = 2$). The orthodox Poisson model can be fitted using the `mcglm` package using the `Tweedie` variance function $\nu(\mu, p ) = \mu^p$ where the power parameter $p$ is fixed at $1$. For example, ```{r} # Mean model x1 <- seq(-2,2, l = 200) X <- model.matrix(~ x1) mu <- mcglm::mc_link_function(beta = c(1,0.8), X = X, offset = NULL, link = "log") # Data structure data <- data.frame("x1" = x1) # Covariance model Z0 <- mc_id(data) # Data structure data$y <- rpois(200, lambda = mu$mu) # Fit fit.poisson <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "log", variance = "tweedie", power_fixed = TRUE, data = data) ``` Another very useful model for count data is the negative binomial. ```{r} # Simulating negative binomial models set.seed(1811) x <- rtweedie(200, mu = mu$mu, power = 2, phi = 0.5) y <- rpois(200, lambda = x) data <- data.frame("y1" = y, "x1" = x1) fit.pt <- mcglm(linear_pred = c(y ~ x1), matrix_pred = list(Z0), link = "log", variance = "poisson_tweedie", power_fixed = FALSE, data = data) summary(fit.pt) ``` Note that, we estimate the power parameter rather than fix it at $p = 2$. [`mcglm`]: https://github.com/bonatwagner/mcglm [mcglm/README]: https://github.com/bonatwagner/mcglm/blob/main/README.md