knitr::opts_chunk$set( collapse = TRUE, comment = "#>" )
The purpose of this vignette is to explain the withdrawal/dropout modelling assumptions used in the package.
Suppose that survival time $T$ follows a piecewise exponential model with rates $(\lambda_1, \lambda_2,\ldots, \lambda_k$) and changepoints $(t_1,\ldots,t_{k-1})$.
Suppose that dropout time $C$ is independent and follows an exponential distribution with rate $\mu$.
We are interested in the probability that no event is observed by time $t$.
$$pr\left( \left\lbrace T >t \text{ and }C > t\right\rbrace \text{or} \left\lbrace T > C \text{ and }C < t\right\rbrace \right) = pr\left( T >t \text{ and }C > t\right) + pr\left( T > C \text{ and }C < t \right)$$
The first term:
$$pr\left( T >t \text{ and }C > t\right) = \exp(-\sum \lambda_i g(t, t_i))\exp(-\mu t)$$
where $g(t, t_i) = \max(0, \min(t - t_{i - 1}, t_i - t_{i-1}))$.
The second term:
$$ pr\left( T > C \text{ and }C < t \right) = \int_0^t pr(T > c)f(c)dc$$
Consider the case with just one change point, first when $t < t_1$:
$$\int_0^t pr(T > c)f(c)dc ~~= \int_0^{t}\exp(-\lambda_1c)\mu\exp(-\mu c)dc$$ $$ = \left( \frac{\mu}{\mu + \lambda_1} \right)\left[1 - \exp\left\lbrace-(\mu + \lambda_1)t\right\rbrace\right]$$
Now, when $t > t_1$:
$$\int_0^t pr(T > c)f(c)dc ~~= \int_0^{t_1}\exp(-\lambda_1c)\mu\exp(-\mu c)dc + \int_{t_1}^t\exp\left\lbrace-\lambda_2(c - t_1) - \lambda_1t_1\right\rbrace\mu\exp(-\mu c)dc$$
The first term on the right-hand-side is:
$$ \left( \frac{\mu}{\mu + \lambda_1} \right)\left[1 - \exp\left\lbrace-(\mu + \lambda_1)t_1\right\rbrace\right].$$
The second term on the right-hand-side is equal to:
$$-\left( \frac{\mu}{\mu + \lambda_2} \right) \left[ \exp\left\lbrace -\lambda_2(t - t_1)-\lambda_1t_1-\mu t\right\rbrace - \exp\left\lbrace-(\mu + \lambda_1)t_1 \right\rbrace \right] $$
$$= \left( \frac{\mu}{\mu + \lambda_2} \right) \exp\left\lbrace-(\mu + \lambda_1)t_1\right\rbrace\left[1 - \exp\left\lbrace-(\mu + \lambda_2)(t - t_1)\right\rbrace\right]$$
And it's easy to see how this formula is extended for more changepoints. This formula makes sense: we calculate the probability that $\min(T,C)$ occurs in an interval $t_{i-1}, t_i$, multiply this by the probability the minimum was $C$ rather the $T$ (which is just $\mu / (\mu + \lambda_i)$), and then we sum over the intervals up to time $t$.
devtools::load_all("..") surv_pieces(t = 5, change_points = 10, lambdas = c(log(2) / 15, log(2) / 15), mu = 0.12, dropouts = FALSE)
If, instead, we were interested in the probability that a patient has not dropped out by time $t$ we would simply switch
$$\frac{\mu}{\lambda_i + \mu}$$
for
$$\frac{\lambda_i}{\lambda_i + \mu}$$
in the above formulae. We can do this with the surv_pieces
function by change dropouts
to TRUE
:
surv_pieces(t = 5, change_points = 10, lambdas = c(log(2) / 15, log(2) / 15), mu = 0.12, dropouts = TRUE)
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.