If the time is measured in discrete periods, a model that is often used for the time $X$ failure of an item is:$$P_{\theta}[X=k]=\theta^{k-1}(1-\theta), k=1,2,...$$ where $0<\theta<1$. Suppose that we only record the time of failure, if failure occurs on or before time $r$ and otherwise just note that the item has lived at least $r+1$ periods. Hence, we observe $Y_1,...,Y_n$, which are iid and have common frequency function: $$f(k,\theta)=\theta^{k-1}(1-\theta), k=1,...,r$$ $$f(r+1,\theta)=1-P_\theta[X\leq r]=1-\Sigma_{k=1}^r(1-\theta)=\theta^r$$ The $r+1$ notation, means survival for at least $r+1$ periods. Let $M=$number of indices i such that $Y_i=r+1$. I want to derive that the MLE of $\theta$ based on $Y_1,...,Y_n$ is: $$\hat{\theta}(\mathbf{Y})=\frac{\Sigma_{i=1}^nY_i-n}{\Sigma_{i=1}^nY_i-M}$$ but I'm pretty confused about it.

Note: This process is known as Censored Geometric Waiting Time.

Any help would be appreciated!


Brace yourself, Math ahead !

The likelihood is defined to be the probability to observed your data under a given parameters : $P(Y_1,...,Y_n|\theta)$

Let's consider that all event are iid:

$P(Y_1,...,Y_n|\theta) = \prod_1^n{P(Y_i|\theta})$

Now let's split the case where $Y_i = r+1$ from the rest:

$P(Y_1,...,Y_n|\theta) = \prod_{Y_i<r+1}{P(Y_i|\theta}) * \prod_{Y_i=r+1}{P(Y_i|\theta})$

Now we know the expression of the probability for both of these product:

$P(Y_1,...,Y_n|\theta) = \prod_{Y_i<r+1}{P(X=Y_i|\theta)} * \prod_{Y_i=r+1}{P(X>=Y_i|\theta)}$

$P(Y_1,...,Y_n|\theta) = \prod_{Y_i<r+1}{(\theta^{Y_i-1}*(1-\theta))} * \prod_{Y_i=r+1}{\theta^{Y_i-1}}$

$P(Y_1,...,Y_n|\theta) = \prod_{Y_i<r+1}{(\theta^{Y_i-1}*(1-\theta))} * \prod_{Y_i=r+1}{\theta^{r}}$

$P(Y_1,...,Y_n|\theta) = \prod_{Y_i<r+1}{(\theta^{Y_i-1}})*(1-\theta)^{n-M} * \theta^{M*r}$

From Now on it is easier to take the log : $L(\theta) = log(P(Y_1,...,Y_n|\theta))$

$L(\theta) = \sum_{Y_i<r+1}[(Y_i-1)*log(\theta)]+(n-M)*log(1-\theta) + M*r*log(\theta)$

$L(\theta) = log(\theta) * [\sum_{Y_i<r+1}Y_i - \sum_{Y_i<r+1}[1]] +(n-M)*log(1-\theta)+ M*r*log(\theta)$

$L(\theta) = log(\theta) * [\sum_{Y_i<r+1}Y_i - (n-M)] +(n-M)*log(1-\theta)+ M*r*log(\theta)$

Now we take the derivative with respect to $\theta$:

$L'(\theta) = \frac{\sum_{Y_i<r+1}Y_i - (n-M)}{\theta} -\frac{n-M}{1-\theta} + \frac{M*r}{\theta}$

We then look for $\theta$ such that $L'(\theta)=0$

$\frac{\sum_{Y_i<r+1}Y_i - (n-M)}{\theta} -\frac{n-M}{1-\theta} + \frac{M*r}{\theta} = 0$

$\frac{\sum_{Y_i<r+1}Y_i - (n-M)}{\theta} + \frac{M*r}{\theta} = \frac{n-M}{1-\theta}$

$\frac{1-\theta}{\theta} = \frac{n-M}{M*r+\sum_{Y_i<r+1}Y_i - (n-M)}$

$\frac{1-\theta}{\theta} = \frac{n-M}{M*(r+1)+\sum_{Y_i<r+1}Y_i - n}$

$\frac{1-\theta}{\theta} = \frac{n-M}{\sum_{Y_i=r+1}Y_i+\sum_{Y_i<r+1}Y_i - n}$

$\frac{1-\theta}{\theta} = \frac{n-M}{\sum_{1}^nY_i - n}$

$\frac{1-\theta}{\theta} = \frac{n-M}{\sum_{1}^nY_i - n}$

$\frac{1}{\theta} = \frac{n-M}{\sum_{1}^nY_i - n} + 1$

$\frac{1}{\theta} = \frac{n-M+\sum_{1}^nY_i - n}{\sum_{1}^nY_i - n}$

$\frac{1}{\theta} = \frac{\sum_{1}^nY_i - M}{\sum_{1}^nY_i - n}$

$\theta = \frac{\sum_{1}^nY_i - n}{\sum_{1}^nY_i - M}$

Done !


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.