Inverse Gaussian Distribution Family Function

Estimates the two parameters of the inverse Gaussian distribution by maximum likelihood estimation.

inv.gaussianff(lmu = "loglink", llambda = "loglink",
      imethod = 1, ilambda = NULL,
      parallel = FALSE, ishrinkage = 0.99, zero = NULL)

Arguments

lmu, llambda

Parameter link functions for the $\mu$ and $\lambda$ parameters. See Links for more choices.

ilambda, parallel

See CommonVGAMffArguments for more information. If parallel = TRUE then the constraint is not applied to the intercept.

imethod, ishrinkage, zero

See CommonVGAMffArguments for information.

Details

The standard (“canonical”) form of the inverse Gaussian distribution has a density that can be written as $$f(y;\mu,\lambda) = \sqrt{\lambda/(2\pi y^3)} \exp\left(-\lambda (y-\mu)^2/(2 y \mu^2)\right)$$ where $y>0$, $\mu>0$, and $\lambda>0$. The mean of $Y$ is $\mu$ and its variance is $\mu^3/\lambda$. By default, $\eta_1=\log(\mu)$ and $\eta_2=\log(\lambda)$. The mean is returned as the fitted values. This VGAM family function can handle multiple responses (inputted as a matrix).

Value

An object of class "vglmff" (see vglmff-class). The object is used by modelling functions such as vglm, rrvglm and vgam.

References

Johnson, N. L. and Kotz, S. and Balakrishnan, N. (1994). Continuous Univariate Distributions, 2nd edition, Volume 1, New York: Wiley.

Forbes, C., Evans, M., Hastings, N. and Peacock, B. (2011). Statistical Distributions, Hoboken, NJ, USA: John Wiley and Sons, Fourth edition.

Author

T. W. Yee

Note

The inverse Gaussian distribution can be fitted (to a certain extent) using the usual GLM framework involving a scale parameter. This family function is different from that approach in that it estimates both parameters by full maximum likelihood estimation.

Examples

idata <- data.frame(x2 = runif(nn <- 1000))
idata <- transform(idata, mymu   = exp(2 + 1 * x2),
                          Lambda = exp(2 + 1 * x2))
idata <- transform(idata, y = rinv.gaussian(nn, mu = mymu, Lambda))
fit1 <-   vglm(y ~ x2, inv.gaussianff, data = idata, trace = TRUE)
#> Iteration 1: loglikelihood = -3404.5408
#> Iteration 2: loglikelihood = -3400.1357
#> Iteration 3: loglikelihood = -3400.1076
#> Iteration 4: loglikelihood = -3400.1075
#> Iteration 5: loglikelihood = -3400.1075
rrig <- rrvglm(y ~ x2, inv.gaussianff, data = idata, trace = TRUE)
#> RR-VGLM    linear loop  1 :  loglikelihood = -3420.0968
#>    Alternating iteration 1 ,   Convergence criterion  =  1 
#>     ResSS  =  1758.4748488 
#>    Alternating iteration 2 ,   Convergence criterion  =  1.293016e-16 
#>     ResSS  =  1758.4748488 
#> RR-VGLM    linear loop  2 :  loglikelihood = -3400.3441
#>    Alternating iteration 1 ,   Convergence criterion  =  1 
#>     ResSS  =  2032.777272 
#>    Alternating iteration 2 ,   Convergence criterion  =  0 
#>     ResSS  =  2032.777272 
#> RR-VGLM    linear loop  3 :  loglikelihood = -3400.1085
#>    Alternating iteration 1 ,   Convergence criterion  =  1 
#>     ResSS  =  2034.3110727 
#>    Alternating iteration 2 ,   Convergence criterion  =  0 
#>     ResSS  =  2034.3110727 
#> RR-VGLM    linear loop  4 :  loglikelihood = -3400.1075
#>    Alternating iteration 1 ,   Convergence criterion  =  1 
#>     ResSS  =  2034.2935559 
#>    Alternating iteration 2 ,   Convergence criterion  =  0 
#>     ResSS  =  2034.2935559 
#> RR-VGLM    linear loop  5 :  loglikelihood = -3400.1075
coef(fit1, matrix = TRUE)
#>             loglink(mu) loglink(lambda)
#> (Intercept)    2.001970       2.0153819
#> x2             1.041353       0.9964811
coef(rrig, matrix = TRUE)
#>             loglink(mu) loglink(lambda)
#> (Intercept)    2.001963       2.0153949
#> x2             1.041368       0.9964544
Coef(rrig)
#> A matrix:
#>                    latvar
#> loglink(mu)     1.0000000
#> loglink(lambda) 0.9568706
#> 
#> C matrix:
#>      latvar
#> x2 1.041372
#> 
#> B1 matrix:
#>             loglink(mu) loglink(lambda)
#> (Intercept)    2.001963        2.015395
summary(fit1)
#> 
#> Call:
#> vglm(formula = y ~ x2, family = inv.gaussianff, data = idata, 
#>     trace = TRUE)
#> 
#> Coefficients: 
#>               Estimate Std. Error z value Pr(>|z|)    
#> (Intercept):1  2.00197    0.06354  31.507  < 2e-16 ***
#> (Intercept):2  2.01538    0.08995  22.407  < 2e-16 ***
#> x2:1           1.04135    0.11385   9.147  < 2e-16 ***
#> x2:2           0.99648    0.16028   6.217 5.06e-10 ***
#> ---
#> Signif. codes:  0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
#> 
#> Names of linear predictors: loglink(mu), loglink(lambda)
#> 
#> Log-likelihood: -3400.108 on 1996 degrees of freedom
#> 
#> Number of Fisher scoring iterations: 5 
#> 
#> Warning: Hauck-Donner effect detected in the following estimate(s):
#> '(Intercept):1'
#>