# Estimating Uniform distribution endpoints using data with errors

Suppose I have a random variable $X$ ~ $Unif(0,\theta)$ where I want to estimate $\theta$. I draw a sample $X_1,...,X_n$.One way is to get a point estimate using e.g. maximum likelihood estimation via the sufficient statistic $X_{(n)}$.

However, suppose I want a 95% confidence interval of sorts on the value of $\theta$, how should I construct one?

The reason why I am interested in confidence interval is because of the following issue, which is actually my main question (the above prelude on confidence interval is just a suggestion on possible approach). So actually my sample came with some noise, in a sense that I actually observe $X_1+E_1,...,X_n+E_n$ where $E_i$~$N(0,\sigma_i^2)$ independently from each other and from the $X$'s. How should I make some form of statistical estimation about $\theta$? I would prefer some form of confidence estimate.

You can just use maximum likelihood! First, you give the standard deviation of your normal random variable $E_i$ as $\sigma_i$. If that is really what you want, you are more or less out of luck, since you then have more parameters than observations. So I will presuma that was a slip, and assume identical standard deviations $\sigma$.

Then, each of your observations have the distribution of the sum independent uniform and normal variables. So first we must find that distribution, which is an exercise in convolution. Let $f(x) = \frac{1}{\theta}I_{0 \le x \le \theta}(x)$ be the uniform density and $g(e)$ be the normal density. Calculating the convolution we get \begin{align} f \ast g (x) &= \int f(x-e) \frac{1}{\sqrt{2\pi \sigma^2}} e^{-\frac12(\frac{e}{\sigma})^2} \, de = \\ &= \frac{1}{\theta} \int_{x-\theta}^x \frac{1}{\sqrt{2\pi}\sigma} e^{-\frac12(\frac{e}{\sigma})^2} \, de \\ &= \frac{1}{\theta} \int_{\frac{x-\theta}{\sigma}}^{\frac{x}{\sigma}} \frac{1}{\sqrt{2\pi}} e^{-\frac12 u^2} \, du = \\ &= \frac{1}{\theta} \left( \Phi(x/\sigma) - \Phi((x-\theta)/\sigma) \right) \end{align} where we have used a simple substitution, and where $\Phi$ is the standard normal cumulative distribution function. Now, this can be used as usual for maximum likelihood estimation. Let us write the density function we found above as $h(u; \theta, \sigma)$. Then the likelihood function is $$L(u;\theta, \sigma) = \prod_{i=1}^n h(u_i;\theta,\sigma)$$ so the log-likelihood becomes $$l(u;\theta,\sigma) = -n\log\theta + \sum_{i=1}^n \log \left ( \Phi(u_i/\sigma) -\Phi((u_i-\theta)/\sigma) \right )$$ Maximize this over the unknown parameters!

EDIT after the comments: Then I will continue assuming $\sigma$ is known! You can extend it to your case with distinct known $\sigma_i$. We can make a confidence interval on $\theta$ by likelihood profiling. We just use the (asymptotic) result that $$2(l(\hat{\theta})-l(\theta)) \sim \chi^2_d$$ where $\sim$ is read "distributed as" and $d$ is the dimension of $\theta$, in our case 1. $\hat{\theta}$ is the maximum likelihood estimator.

You can invert this (numerically) to get a confidence interval.

• Opps, I should be clear that all the $\sigma_i$ are actually known. Anyway, I am aware of how to use MLE, I was hoping for a confidence interval kind of method. – renrenthehamster Jun 19 '14 at 15:04
• You can make a confidence interval based on the likelihood! for example, based on likelihood profiling. What I have written above can be easily adjusted for known $\sigma_i$. – kjetil b halvorsen Jun 19 '14 at 15:06
• I am reading your edits. When you suggest using the asymptotic result, is that Wilk's theorem? I am not too sure if Wilk's theorem is valid here because one of the assumptions of Wilk's theorem is that the support doesn't depend on the parameter. – renrenthehamster Jun 20 '14 at 8:23
• In the modified model the support has been "smoothed out" to the whole real line! so that is not a problem! – kjetil b halvorsen Jun 20 '14 at 9:29
• (+1) Caution: for small datasets and many of the $\sigma$ large compared to $\theta$, it's reasonably likely that many of the $x_i$ will be negative. In such cases the likelihood will be maximized at $\theta=0$, where these "asymptotic results" will be incorrect. – whuber Jun 5 '17 at 23:40