Introduction to Frailty Model

This passage discusses frailty models used in survival analysis, particularly in the context of lifetime distributions, to account for uncertainty in the hazard rate of an event or outcome. It explores the use of mixture distributions, introduces a frailty random variable to quantify the uncertainty, and demonstrates how certain choices of baseline hazard functions lead to specific distributions such as the Weibull, exponential, Burr, and Pareto distributions.

Some author2 min read

Frailty models can be found in the field of survival analysis, particularly in the context of analyzing lifetime distributions. These models are used to account for the uncertainty associated with the hazard rate of an event or outcome. They are a type of mixture distribution that combines multiple components to generate new distributions. We can introduce a frailty random variable λ\lambda, which we will use to quantify the uncertainty over the hazard rate.

Specifically, let's consider the conditioned hazard function hX∣Λ(x∣λ)h_{X|\Lambda}(x|\lambda) which equals to λa(x)\lambda a(x) where a(x)a(x) is the baseline hazard function. We can also define the incomplete baseline hazard accumulation function A(x)=∫0xa(t)dtA(x) = \int_0^x a(t)dt. Then, the survival function SX∣Λ(x∣λ)S_{X|\Lambda}(x|\lambda) is given by SX∣Λ(x∣λ)=exp⁡(−A(x)λ)S_{X|\Lambda}(x|\lambda) = \exp(-A(x)\lambda). And the marginal survival function SX(x)S_X(x) is given by SX(x)=EΛ[SX∣Λ(x∣λ)]=EΛ[exp⁡(−A(x)λ)]S_X(x) = E_{\Lambda}[S_{X|\Lambda}(x|\lambda)] = E_{\Lambda}[\exp(-A(x)\lambda)], which is the moment generating function of Λ\Lambda evaluated at −A(x)-A(x).

When a(x)=γxγ−1a(x) = \gamma x^{\gamma-1}, then A(x)=xγA(x) = x^{\gamma}, our conditional probability will be SX∣Λ(x∣λ)=exp⁡(−xγλ)=exp⁡(−(xλ−1/γ)γ)S_{X|\Lambda}(x|\lambda) = \exp(-x^{\gamma}\lambda) = \exp\left(- \left(\frac{x}{\lambda^{-1/\gamma}} \right)^{\gamma} \right) . This is the Weibull distribution.

In particular when γ=1\gamma =1, then we will have SX∣Λ(x∣λ)=exp⁡(−xλ)S_{X|\Lambda}(x|\lambda) = \exp(-x\lambda), which is an exponential distribution.

Now if we have a Weibull distribution as our condition probability and lambda follows gamma distribution, we will have a Burr distribution as our marginal probability. We can show it by using the following.

First the moment generating function of a gamma distribution is the following:

MΛ(−t)=∫0∞exp⁡(−xt)xa−1exp⁡(−x/θ)Γ(a)θadxM_{\Lambda}(-t) = \int_0^{\infty} \exp(-x t) \frac{x^{a-1}\exp(-x/\theta)}{\Gamma(a)\theta^a} dx MΛ(−t)=1Γ(a)θa∫0∞xa−1exp⁡(−xt−x/θ)dx=1Γ(a)θa(θθt+1)a∫0∞((θt+1θ)x)a−1exp⁡(−(θt+1θ)x)d((θt+1θ)x)M_{\Lambda}(-t) = \frac{1}{\Gamma(a)\theta^a} \int_0^{\infty} x^{a-1}\exp(-x t-x/\theta) dx \\ =\frac{1}{\Gamma(a)\theta^a} \left(\frac{\theta}{\theta t+1}\right)^{a}\int_0^{\infty} \left( \left(\frac{\theta t+1}{\theta}\right)x\right)^{a-1}\exp\left(- \left(\frac{\theta t+1}{\theta}\right)x\right) d\left(\left(\frac{\theta t+1}{\theta}\right)x\right)

Where we can replace (θt+1θ)x\left(\frac{\theta t+1}{\theta}\right)x with uu, so we will have

MΛ(−t)=1Γ(a)θa(θθt+1)a∫0∞ua−1exp⁡(−u)du=(1θt+1)aM_{\Lambda}(-t) =\frac{1}{\Gamma(a)\theta^a} \left(\frac{\theta}{\theta t+1}\right)^{a}\int_0^{\infty} u^{a-1}\exp\left(- u\right) du = \left(\frac{1}{\theta t+1}\right)^{a}

Woops, so the moment generating function for gamma distribution is the following;

MΛ(t)=(11−θt)aM_{\Lambda}(t) = \left(\frac{1}{1-\theta t}\right)^{a}

And now we also know that the marginal probability follows the following:

SX(x)=MΛ(−A(x))S_X(x) = M_{\Lambda}(-A(x))

In this case, we have A(x)=xγA(x) = x^{\gamma}, so by substitution, we will have

SX(x)=(11+θxγ)aS_X(x) = \left(\frac{1}{1+\theta x^{\gamma}}\right)^{a}

which is a Burr distribution.

If we set γ\gamma to be one, we will have

SX(x)=(11+θx)aS_X(x) = \left(\frac{1}{1+\theta x}\right)^{a}

,

which, in turns, is a Pareto distribution