---
title: "Model fitting"
author: Eunseop Kim
output: 
  rmarkdown::html_vignette:
    fig_width: 7
    fig_height: 5
vignette: >
  %\VignetteIndexEntry{Model fitting}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include=FALSE}
knitr::opts_chunk$set(collapse = TRUE, comment = "#>", dpi = 300)
```
```{r, echo=FALSE}
library(melt, warn.conflicts = FALSE)
```

## Fitting an `EL` object
The melt package provides several functions to construct an `EL` object or an 
object that inherits from `EL`:

* `el_mean()` for the mean.
* `el_sd()` for the standard deviation.
* `el_lm()` for linear models.
* `el_glm()` for generalized linear models.

We illustrate the usage of `el_mean()` with the `faithful` data set. 

```{r, eval=TRUE}
data("faithful")
str(faithful)
summary(faithful)
```

Suppose we are interested in evaluating empirical likelihood at `c(3.5, 70)`.

```{r}
fit <- el_mean(faithful, par = c(3.5, 70))
class(fit)
showClass("EL")
```

The `faithful` data frame is coerced to a numeric matrix. Simple print method 
shows essential information on `fit`. 

```{r}
fit
```

Note that the maximum empirical likelihood estimates are the same as the sample 
average. The chi-square value shown corresponds to the minus twice the empirical
log-likelihood ratio. It has an asymptotic chi-square distribution of 2 degrees 
of freedom under the null hypothesis. Hence the $p$-value here is not exact. The
convergence status at the bottom can be used to check the convex hull 
constraint. 

Weighted data can be handled by supplying the `weights` argument. For non-`NULL`
`weights`, weighted empirical likelihood is computed. Any valid `weights` is 
re-scaled for internal computation to add up to the total number of 
observations. For simplicity, we use `faithful$waiting` as our weight vector.

```{r}
w <- faithful$waiting
(wfit <- el_mean(faithful, par = c(3.5, 70), weights = w))
```

We get different results, where the estimates are now the weighted sample 
average. The chi-square value and the associated $p$-value are based on the same 
limit theorem, but care must be taken when interpreting the results since they 
are largely affected by the limiting behavior of the weights.