$X_i$ are i.i.d exponential, mean $\lambda^{-1}$ for $1 \leq i \leq n$ and, the values are measured such that $X_i = c$ if $X_i \geq c$ and $X_i$ otherwise. Show that MLE of $\lambda = \frac{n-T_n}{S_n+cT_n}$ where $S_n= \sum_{j=1}^n X_jI(X_j < c)$ and $T_n = \sum_{j=1}^n I(X_j \geq c)$


The likelihood function is given by $L(\lambda | x_i, x_2, \ldots x_n) = \lambda^n e^{-\lambda \sum x_i} = \lambda^n e^{-\lambda(S_n+cT_n)}$ $\implies$ $log L = n log \lambda -\lambda(S_n+cT_n) $

and hence $\hat\lambda=\frac{n}{S_n+cT_n}$ which is different from the expected answer.


Indicator variables to the rescue.

Few observations:

$P(X_i < x) = 1 - \lambda e^{-x}$ if $ 0 \leq x < c$

$P(X_i = c) = P(X_i \geq c) = e^{-\lambda c}$ (This one looks counter intuitive at first look, but that is what the $X_i=c$ if $X_i \geq c$ returns)


$\begin{align}L( \lambda| x_i) &= \prod_{i=1}^{n} (\lambda e^{-\lambda x_i}) I(0 \leq x_i < c) \times (e^{-\lambda c})I( x_i \geq c) \\&= \lambda^{\sum_{i=1}^n I(0 \leq x_i < c)}e^{-\lambda(\sum_{i=1}^n x_iI(0 \leq x_i < c)} \times e^{-\lambda c \sum_{i=1}^n I(x_i \geq c)} \\&= \lambda^{n-T_n}e^{-\lambda S_n} \times e^{-\lambda c T_n}\end{align}$

$\log L= (n-T_n) \log \lambda -\lambda(S_n+cT_n)$

$\frac{\partial \log L}{\partial \lambda} = \frac{n-T_n}{\lambda}-(S_n+cT_n)$

which gives $\hat \lambda = \frac{n-T_n}{S_n+cT_n}$

| cite | improve this answer | |

I tired it a bit, but did not get a neat MLE like yours. In any case, I thought to post my attempt.

We observe $y_i=\min\{c,X_i\}$ which is again iid. I think since we do not observe $X_i$ we can not plug in its pdf in the likelihood device, rather the pdf of $y_i$, which is what we observe. The pdf of $y_i$ is $$f_i(y)=\lambda e^{-\lambda y}(1-e^{-\lambda c})^{-1}, 0\leq y\leq c$$ now we can write the likelihood function \begin{align} L(\lambda)&=\prod_{i=1}^{n}f_i(y_i)\\ &=\lambda^{n}(1-e^{-\lambda c})^{-n}\prod_{i=1}^{n}e^{-\lambda y_i}\\ &=\lambda^{n}(1-e^{-\lambda c})^{-n}\prod_{i=1}^{n}e^{-\lambda X_iI(X_i<c)}\prod_{i=1}^{n}e^{-\lambda cI(X_i\geq c)}\\ \end{align} and the log likelihood \begin{align} l(\lambda)&=n\log(\lambda)-n\log(1-e^{-\lambda c})-\lambda c\sum_{i=1}^{n}I(X_i\geq c)-\lambda \sum_{i=1}^{n}X_iI(X_i<c)\\ \end{align}

\begin{align} \frac{d}{d\lambda}l(\lambda)&=\frac{nc}{1-e^{c\lambda}}-S_n-cT_n+\frac{n}{\lambda}\\ \end{align}

Note that $$\lim_{\lambda \to 0}=-S_n+\frac{c}{2}(n-2T_n)$$ and $$\lim_{\lambda \to \infty}=-S_n-cT_n)$$

therefore a solution exists and will be unique (by monotonicity) if $-S_n+\frac{c}{2}(n-2T_n)>0$ or that $\bar{y}<\frac{c}{2}$ otherwise this does not have a solution on $\mathbb{R}$.

If a solution exists, then solving $\frac{d}{d\lambda}l(\lambda)=0$, however, seems not to result in your neat MLE!

| cite | improve this answer | |
  • $\begingroup$ I think I did it wrong because the PDF of $X_i$ is simply not $\lambda e^{-\lambda x_i}$ as $P(X_i=c)=1$ for $x_i \geq c$. I think the PDF you have used also does not reflect this $\endgroup$ – user239438 May 13 '15 at 17:36
  • $\begingroup$ since $X_i$ is continuous, $P(X_i=c)=0$, my pdf should not be wrong ... but I will look at it later again. $\endgroup$ – Math-fun May 13 '15 at 17:55

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.