Introduction

The unifed distribution is the Exponential Dispersion Family (EDF) generated by the uniform distribution (see Jørgensen 1997, chap. 2 and 3 to see how an EDF can be generated from a distribution). It has support on the interval (0,1).

The unifed R package focuses on the case where the dispersion parameter is equal to one. This is due to some numerical problems encountered when trying to implement the general case (see Quijano Xacur 2019 for details). The package also includes code for using this distribution with stan.

The unifed can be used as the response distribution of a GLM. Note that although the functions in this package implement the case with dispersion 1, it is still possible to have classes with weights different than one in a unifed GLM. This is because for a GLM we only need to minimize the deviance and disregard the rest of the density (which is the part that gives numerical stability).

We start with a section about Exponential Dispersion Families and their properties. Then we introduce the unifed density, it’s cumulant generator, and unit deviance function. This is followed by an example, where we use the package for fitting a frequentist and a Bayesian unifed GLM to a publicly available dataset. We close with a brief comparison between the unifed GLM and the beta regression.

Exponential Dispersion Families

An Exponential Dispersion Family (EDF) is a set of distributions whose densities are given by

θ and Θ are called the canonical parameter and canonical space, respectively and ϕ is known as the dispersion parameter. The parametrization above is know as the canonical parametrization. Forθ ∈ int(Θ) (here int stands for interior), $$ \mathbb{E}[Y] = \dot{\kappa}\left(\theta\right)\qquad \mbox{and}\qquad \mathbb{V}[Y]=\phi\ddot{\kappa}\left(\theta\right),$$ where κ̇ = κ′ and $\ddot{\kappa}=\dot{\kappa}'$. This motivates the following definitions.

Definition Given an exponential dispersion family, the mean domain of the family is defined as Ω = {μ = κ̇(θ) : θ ∈ int(Θ)}.

Another important property is that the support of the distribution only depends on ϕ (and not on θ). For a given family, let C_ϕ be the convex support of any member of the family with dispersion parameter ϕ. We define the convex support of the family as C_Φ = ⋃_{ϕ ∈ Φ}C_ϕ.

Definition The unit deviance function of an exponential dispersion family is defined asd : C_Φ × Ω → [0, ∞) with d(y, μ) = 2[sup_{θ ∈ Θ}{θy − κ(θ)} − yκ̇⁻¹(μ) + κ(κ̇⁻¹(μ))].

The unit deviance function allows to re-parametrize the densities of the family as for some function c. This is known as the mean–value parametrization.

The Unifed Family

The unifed can be characterized as the only EDF containing the uniform distribution. In fact, unifed = uniform + edf

Due to the numerical instability (see Quijano Xacur 2019) of the density when the dispersion parameter is different than 1, we focus here in the case where the dispersion parameter is equal to 1. In this case, the density is given by

$$f(x;\theta) = \left\{ \begin{array}{ll} \frac{\theta}{e^\theta - 1} e^{x \theta} & \mbox{if } \theta \neq 0\\ 1 & \mbox{if } \theta = 0 \end{array} \right. \quad \mbox{for } x \in (0,1),$$

where θ is called the canonical parameter. The unifed package provides the functions dunifed, punifed, qunfied and runifed for the density, distribution, quantile and random generation functions respectively. Here are some examples of use of these functions taken from the documentation

library(unifed)
dunifed( c(0.1,0.3,0.7), 10)

x <- c(0.1,0.4,0.7,1)
punifed(x,-5)

p <- 1:9/10
qunifed(p,5)

runifed(20,-3.3)

The mean of this distribution can be any value in (0,1). In the next a section explicit formula for the mean in terms of θ is given. A with any proper exponential family, there is a one-to-one correspondence between the mean μ and θ. The next graphs show the density of the unifed as the mean varies.

Cumulant generator

The cumulant generator function characterizes and exponential family. For the unifed it is given by

$$\kappa(\theta)=\left\{ \begin{array}{ll} \log\left(\frac{e^\theta-1}{\theta}\right)& \mbox{if }\theta\neq 0\\ 0 & \mbox{if }\theta=0 \end{array} \right..$$

One difficulty in implementing κ in a computer function is that e^θ above takes the value infinity even for relatively small θ. For instance exp(1000) in R gives Inf.

To avoid this problem the function that implements κ in the package, unifed.kappa, uses an approximation for values of θ that are greater than 50. This function is implemented as follows

$$\kappa_{mod}(\theta)=\left\{ \begin{array}{ll} \kappa(\theta) & \mbox{if }\theta \leq 50\\ \theta - \log(\theta) & \mbox{if }\theta > 50 \end{array} \right..$$

The justification for this approximation is that if y is defined as $$ y = \log\left(\frac{e^\theta - 1 }{\theta}\right),$$

then also,

y = θ − log (θ + e^−y).

Now, κ is an increasing function and y goes to infinity as θ goes to infinity. Thus for large values of θ the term e^−y is close to zero. We use 50 as the threshold for the approximation since evaluating the following code in R already gives zero

 log( ( exp(50) - 1) / 50  ) - ( 50 - log (50) )

The Variance and Unit Deviance Functions

From the properties of exponential families we know then that if X follows a unifed distribution distribution with canonical parameter θ, its mean and variance are given by

$$\begin{aligned} \mathbb{E}[X]&= \dot{\kappa}(\theta) = \left\{ \begin{array}{ll} \frac{(\theta-1)e^\theta + 1}{\theta(e^\theta -1)} & \mbox{if }\theta \neq 0\\ \frac{1}{2} & \mbox{if }\theta=0 \end{array} \right., \qquad \\ \mathbb{V}[X]&= \ddot{\kappa}(\theta)=\left\{ \begin{array}{ll} \left(\frac{ e^{2\theta} - (\theta+2)e^\theta + 1} {\theta^2 (e^\theta-1)^2}\right)& \mbox{if }\theta \neq 0 \\ \frac{1}{12} & \mbox{if }\theta=0. \end{array} \right. \end{aligned}$$

Where κ̇ and $\ddot{\kappa}$ are the first and second derivative of κ, respectively. The variance function allows us to express the variance in terms of the mean. It is defined as

$${\bf V}(\mu) = \ddot{\kappa}(\dot{\kappa}^{-1}(\mu)).$$

We have not been able to find an analytical expression for κ̇⁻¹(μ). Thus, the function unifed.kappa.prime.inverseimplements it using the Newton-Raphson method for solving equations. It takes a value between 0 and 1 and it returns the value of θ that corresponds to that mean. There are some numerical stability problems around 0 and 0.5.

In order to solve the problem around 0.5, this function returns 0 (which is the image of 0.5), for every μ with |μ − 0.5| < 0.0001.

To address the problems around 0, unifed.kappa.prime.inverse returns −κ̇⁻¹(1 − μ) for μ < 0.1. The justification for this is that for every θ

κ̇(θ) = 1 − κ̇(−θ).

In the package the variance function is implemented by the function unifed.varf. The following code plots of the values of the variance function.

x <- seq(0,1,length.out=100)
y <- unifed.varf(x)
plot(x,y,type="l",xlab=expression(mu),ylab=expression(bold(V)(mu)),main="Unifed Variance Function")

We do not have an analytical expression for the unit deviance. The package provides the function unifed.deviance that computes the deviance numerically. It uses unifed.kappa.prime.inverse and the fact that the deviance can be expressed as

d(y, μ) = 2[y{κ̇⁻¹(y) − κ̇⁻¹(μ)} − κ(κ̇⁻¹(y)) + κ(κ̇⁻¹(μ))].

Unifed GLM - An Illustrative Example

The package provides the function unifed which returns a family that can be used with the glm function of R. This section gives an example of how to prepare a dataset and fit a unifed GLM.

The data from this section appears in De Jong and Heller (2008). It is based on 67,856 one–year auto in- surance policies from 2004 or 2005. It comes with the package in the variable car.insurance and it is lazily loaded. The original csv file can be downloaded from the companion site of the book, as the dataset called Car. The following text is the description of the variables provided by the website of the book

This data set is based on  one-year vehicle insurance
policies taken out in 2004 or 2005. There are 67856 policies, of
which  4624 (6.8%) had at least one claim. 

Variables:

veh_value   vehicle value, in $10,000s
exposure    0-1
clm     occurrence of claim (0 = no, 1 = yes)
numclaims   number of claims
claimcst0   claim amount (0 if no claim)
veh_body    vehicle body, coded as
              BUS
              CONVT = convertible  
              COUPE 
              HBACK = hatchback              
              HDTOP = hardtop
              MCARA = motorized caravan
              MIBUS = minibus
              PANVN = panel van
              RDSTR = roadster
              SEDAN    
              STNWG = station wagon
              TRUCK           
              UTE - utility
veh_age age of vehicle: 1 (youngest), 2, 3, 4
gender      gender of driver: M, F
area        driver's area of residence: A, B, C, D, E, F
agecat      driver's age category: 1 (youngest), 2, 3, 4, 5, 6

We are interested in modeling the exposure; which is the proportion of time in a year in which the insurance policy is in-force for a given client. We use gender, agecat, area and veh_age as the explanatory variables.

car.insurance$agecat <- factor(car.insurance$agecat)
car.insurance$veh_age <- factor(car.insurance$veh_age)

agg.car.data <- aggregate( cbind(exposure,rep(1, dim(car.insurance)[1] )) ~ gender + agecat + area + veh_age,
                          data=car.insurance,
                          FUN=sum)

colnames(agg.car.data)[colnames(agg.car.data) == "V2"] <- "weight"
colnames(agg.car.data)[colnames(agg.car.data) == "exposure"] <- "class.exposure"
agg.car.data$class.exposure <- agg.car.data$class.exposure / agg.car.data$weight

library(data.table)
car.insurance <- data.table(car.insurance)
car.insurance[,agecat:=factor(agecat)]
car.insurance[,veh_age:=factor(veh_age)]

agg.car.data <- car.insurance[,.(class.exposure=sum(exposure)/.N,
                              weight=.N),
                              by=.(gender,agecat,area,veh_age)]

exposure.model <- glm(class.exposure ~ gender + agecat + area + veh_age,
                 family=unifed(),
                 weights=weight,
                 data=agg.car.data)

summary(exposure.model,dispersion=1)

## 
## Call:
## glm(formula = class.exposure ~ gender + agecat + area + veh_age, 
##     family = unifed(), data = agg.car.data, weights = weight)
## 
## Coefficients:
##              Estimate Std. Error z value Pr(>|z|)    
## (Intercept) -0.331898   0.019711 -16.838  < 2e-16 ***
## genderM      0.028770   0.008995   3.198 0.001382 ** 
## agecat2      0.001109   0.018361   0.060 0.951818    
## agecat3      0.053024   0.017834   2.973 0.002947 ** 
## agecat4      0.058287   0.017770   3.280 0.001038 ** 
## agecat5      0.104217   0.018921   5.508 3.63e-08 ***
## agecat6      0.069233   0.020958   3.303 0.000955 ***
## areaB        0.023933   0.013491   1.774 0.076057 .  
## areaC        0.001392   0.012120   0.115 0.908564    
## areaD        0.005330   0.015666   0.340 0.733668    
## areaE        0.011977   0.017545   0.683 0.494834    
## areaF        0.087916   0.021438   4.101 4.11e-05 ***
## veh_age2     0.170794   0.013775  12.398  < 2e-16 ***
## veh_age3     0.161287   0.013262  12.162  < 2e-16 ***
## veh_age4     0.154869   0.013429  11.533  < 2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## (Dispersion parameter for unifed family taken to be 1)
## 
##     Null deviance: 585.47  on 287  degrees of freedom
## Residual deviance: 297.86  on 273  degrees of freedom
## AIC: NA
## 
## Number of Fisher Scoring iterations: 3

plot(predict(exposure.model, type="response"),
     residuals(exposure.model, type="deviance"),
     xlab="Predicted value",
     ylab="Deviance residuals")

exposure.model.2 <- glm(exposure ~ gender + agecat + area + veh_age,
                        family=unifed(),
                        data=car.insurance)

exposure.model <- glm(class.exposure ~ gender + agecat + area + veh_age,
                      family=unifed(),
                      weights=weight,
                      control=list(epsilon=1e-20,maxit=1e5),
                      data=agg.car.data)

exposure.model.2 <- glm(exposure ~ gender + agecat + area + veh_age,
                        family=unifed(),
                        control=list(epsilon=1e-20,maxit=1e5),
                        data=car.insurance)

functions{
#include unifed.stan
}

data{
  int<lower=1> M; //Rows in the design matrix
  int<lower=1> P; //Columns in the design matrix
  matrix[M,P] X;
  vector<lower=0,upper=1>[M] y;
  int<lower=1> ws[M]; //Number of observations in each class
}


parameters{
  vector[P] beta;
}

transformed parameters{
  vector[M] theta;
  theta = X*beta;   
}


model{ 
  beta ~ normal(0,20);
  unifed_glm_lp(y, theta , to_vector(ws));      
}

generated quantities{
  vector[M] replicated_samples=rep_vector(0,M);
  vector[M] mu;
  
  for(i in 1:M){
    int Nobs;
    Nobs=ws[i];
    for(n in 1:Nobs ){
      replicated_samples[i]+=unifed_rng(theta[i]);
    }
    replicated_samples[i]/=Nobs;
    mu[i]=unifed_kappa_prime(theta[i]);
  }}

X <- model.matrix( ~ gender + agecat + area + veh_age, agg.car.data)

model.data <- list(M=dim(X)[1],
                   P=dim(X)[2],
                   X=X,
                   y= agg.car.data$class.exposure,
                   ws= agg.car.data$weight)

library(rstan)
model.list <- stanc_builder("unifed_example.stan",
                            isystem=unifed.stan.folder())

m1 <- stan(model_code=model.list$model_code,
           data=model.data,
           warmup=3e3,
           iter=2e4,
           chains=4)

Comparison to the Beta Regression

The beta regression (Ferrari and Cribari-Neto 2004) is a model for applications with a response variable in (0,1). In R it is implemented in the package betareg (Cribari-Neto and Zeileis 2010). That package includes the enhancements of the model from Simas, Barreto-Souza, and Rocha (2010), where explanatory variables are also used for the dispersion parameter.

The density of the beta distribution is much more flexible than the density of the unifed distribution. To see this compare the plot of the unifed densities shown here with the one provided for the beta in Ferrari and Cribari-Neto (2004).

In the cases where both a unifed GLM and a beta regression give a similar fit, the parsimony principle suggests to take the unifed GLM since it has less parameters.

On the computational side, when using an unifed GLM one can apply the properties of reproductive EDFs for data reduction.