msm (version 1.6.7)

2phase: Coxian phase-type distribution with two phases

Description

Density, distribution, quantile functions and other utilities for the Coxian phase-type distribution with two phases.

Usage

d2phase(x, l1, mu1, mu2, log=FALSE)
  p2phase(q, l1, mu1, mu2, lower.tail=TRUE, log.p=FALSE)
  q2phase(p, l1, mu1, mu2, lower.tail=TRUE, log.p=FALSE)
  r2phase(n, l1, mu1, mu2)
  h2phase(x, l1, mu1, mu2, log=FALSE)

Arguments

x,q

vector of quantiles.

p

vector of probabilities.

n

number of observations. If length(n) > 1, the length is taken to be the number required.

l1

Intensity for transition between phase 1 and phase 2.

mu1

Intensity for transition from phase 1 to exit.

mu2

Intensity for transition from phase 2 to exit.

log

logical; if TRUE, return log density or log hazard.

log.p

logical; if TRUE, probabilities p are given as log(p).

lower.tail

logical; if TRUE (default), probabilities are P[X <= x], otherwise, P[X > x].

Value

d2phase gives the density, p2phase gives the distribution function, q2phase gives the quantile function, r2phase generates random deviates, and h2phase gives the hazard.

Alternative parameterisation

An individual following this distribution can be seen as coming from a mixture of two populations:

1) "short stayers" whose mean sojourn time is \(M_1 = 1/(\lambda_1+\mu_1)\) and sojourn distribution is exponential with rate \(\lambda_1 + \mu_1\).

2) "long stayers" whose mean sojourn time \(M_2 = 1/(\lambda_1+\mu_1) + 1/\mu_2\) and sojourn distribution is the sum of two exponentials with rate \(\lambda_1 + \mu_1\) and \(\mu_2\) respectively. The individual is a "long stayer" with probability \(p=\lambda_1/(\lambda_1 + \mu_1)\).

Thus a two-phase distribution can be more intuitively parameterised by the short and long stay means \(M_1 < M_2\) and the long stay probability \(p\). Given these parameters, the transition intensities are \(\lambda_1=p/M_1\), \(\mu_1=(1-p)/M_1\), and \(\mu_2=1/(M_2-M_1)\). This can be useful for choosing intuitively reasonable initial values for procedures to fit these models to data.

The hazard is increasing at least if \(M_2 < 2M_1\), and also only if \((M_2 - 2M_1)/(M_2 - M_1) < p\).

For increasing hazards with \(\lambda_1 + \mu_1 \leq \mu_2\), the maximum hazard ratio between any time \(t\) and time 0 is \(1/(1-p)\).

For increasing hazards with \(\lambda_1 + \mu_1 \geq \mu_2\), the maximum hazard ratio is \(M_1/((1-p)(M_2 - M_1))\). This is the minimum hazard ratio for decreasing hazards.

General phase-type distributions

This is a special case of the n-phase Coxian phase-type distribution, which in turn is a special case of the (general) phase-type distribution. The actuar R package implements a general n-phase distribution defined by the time to absorption of a general continuous-time Markov chain with a single absorbing state, where the process starts in one of the transient states with a given probability.

Details

This is the distribution of the time to reach state 3 in a continuous-time Markov model with three states and transitions permitted from state 1 to state 2 (with intensity \(\lambda_1\)) state 1 to state 3 (intensity \(\mu_1\)) and state 2 to state 3 (intensity \(\mu_2\)). States 1 and 2 are the two "phases" and state 3 is the "exit" state.

The density is

$$f(t | \lambda_1, \mu_1) = e^{-(\lambda_1+\mu_1)t}(\mu_1 + (\lambda_1+\mu_1)\lambda_1 t)$$

if \(\lambda_1 + \mu_1 = \mu_2\), and

$$f(t | \lambda_1, \mu_1, \mu_2) = \frac{(\lambda_1+\mu_1)e^{-(\lambda_1+\mu_1)t}(\mu_2-\mu_1) + \mu_2\lambda_1e^{-\mu_2t}}{\lambda_1+\mu_1-\mu_2}$$

otherwise. The distribution function is

$$F(t | \lambda_1, \mu_1) = 1 - e^{-(\lambda_1+\mu_1) t} (1 + \lambda_1 t)$$

if \(\lambda_1 + \mu_1 = \mu_2\), and

$$F(t | \lambda_1, \mu_1, \mu_2) = 1 - \frac{e^{-(\lambda_1+\mu_1) t} (\mu_2 - \mu_1) + \lambda_1 e^{-\mu_2 t}}{\lambda_1+\mu_1-\mu_2}$$ otherwise. Quantiles are calculated by numerically inverting the distribution function.

The mean is \((1 + \lambda_1/\mu_2) / (\lambda_1 + \mu_1)\).

The variance is \((2 + 2\lambda_1(\lambda_1+\mu_1+ \mu_2)/\mu_2^2 - (1 + \lambda_1/\mu_2)^2)/(\lambda_1+\mu_1)^2\).

If \(\mu_1=\mu_2\) it reduces to an exponential distribution with rate \(\mu_1\), and the parameter \(\lambda_1\) is redundant. Or also if \(\lambda_1=0\).

The hazard at \(x=0\) is \(\mu_1\), and smoothly increasing if \(\mu_1<\mu_2\). If \(\lambda_1 + \mu_1 \geq \mu_2\) it increases to an asymptote of \(\mu_2\), and if \(\lambda_1 + \mu_1 \leq \mu_2\) it increases to an asymptote of \(\lambda_1 + \mu_1\). The hazard is decreasing if \(\mu_1>\mu_2\), to an asymptote of \(\mu_2\).

References

C. Dutang, V. Goulet and M. Pigeon (2008). actuar: An R Package for Actuarial Science. Journal of Statistical Software, vol. 25, no. 7, 1-37. URL http://www.jstatsoft.org/v25/i07