Truncated Gaussian Regression Models in R

Overview

The R package truncreg provides estimation of regression with a truncated Gaussian response via maximum likelihood.

In short, a linear model with homoscedastic Gaussian (normally distributed) errors is assumed for a latent untruncated response \(y_i^*\):

\[ y_i^* = x_i^\top \beta + \varepsilon_i \qquad \varepsilon_i \sim \mathcal{N}(0, \sigma^2) \]

The observed outcome variable \(y_i\) is then the truncated version of the latent variable \(y_i^*\), restricted to be positive: \(y_i^* | y_i^* > 0\). Instead of \(0\) other truncation points could also be used.

The function truncreg() estimates the regression coefficients via maximum likelihood and returns an object of class "truncreg". Standard methods are provided for extracting information: print(), summary(), predict(), coef(), vcov(), model.matrix(), model.frame(), residuals(), fitted(), logLik(). Furthermore, there is a prodist() method, enabling the workflow provided by the distributions3 and topmodels packages.

A more flexible implementation of truncated regression with more features is available in the crch package. In particular, this provides more response distributions, heteroscedastic errors, interval truncation, and also censoring in addition to truncation.

Installation

The stable version of truncreg is available from CRAN:

install.packages("truncreg")

The latest development version can be installed from R-universe:

install.packages("truncreg", repos = "https://zeileis.R-universe.dev")

License

The package is available under the General Public License version 3 or version 2

Get started

Simulation of a simple artificial sample from a truncated normal distribution.

set.seed(1071)
n <- 1000
x <- runif(n, -5, 5)
y <- 1 + 1 * x + rnorm(n, sd = 3)
y[y <= 0] <- NA
d <- data.frame(x = x, y = y)

Estimate the unknown coefficients and error variance via maximum likelihood.

library("truncreg")
m <- truncreg(y ~ x, data = d)
summary(m)
## 
## Call:
## truncreg(formula = y ~ x, data = d)
## 
## BFGS maximization method
## 39 iterations, 0h:0m:0s 
## g'(-H)^-1g = 4.39E-11 
##  
## 
## 
## Coefficients :
##             Estimate Std. Error t-value Pr(>|t|)    
## (Intercept) 0.733149   0.386431  1.8972   0.0578 .  
## x           1.099983   0.098068 11.2166   <2e-16 ***
## sigma       2.999813   0.154238 19.4493   <2e-16 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
## 
## Log-Likelihood: -1218.4 on 3 Df

Estimated coefficients, variance-covariance matrix, confidence intervals, number of observations, log-likelihood, and information criteria.

coef(m)
## (Intercept)           x       sigma 
##   0.7331488   1.0999830   2.9998128
vcov(m)
##             (Intercept)            x        sigma
## (Intercept)  0.14932880 -0.031668241 -0.046290620
## x           -0.03166824  0.009617287  0.008963986
## sigma       -0.04629062  0.008963986  0.023789280
confint(m)
##                   2.5 %   97.5 %
## (Intercept) -0.02424175 1.490539
## x            0.90777371 1.292192
## sigma        2.69751234 3.302113
nobs(m)
## [1] 587
logLik(m)
## 'log Lik.' -1218.385 (df=3)
AIC(m)
## numeric(0)
BIC(m)
## [1] 2455.895