counts.Rmd

# Unbounded Count Models

Unbounded count models are models in which the response is a natural number with no upper bound,
$$
y_i = 0, 1, \dots, \infty .
$$

The two distributions most commonly used to model this are

-   Poisson
-   Negative Binomial

## Poisson

The [Poisson distribution](https://en.wikipedia.org/wiki/Poisson_distribution):
$$
y_i \sim \dpois(\lambda_i) ,
$$

In Stan, the Poisson distribution has two implementations:

-   `r stanfunc("poisson_lpdf")`
-   `r stanfunc("poisson_log_lpdf")`

The Poisson with a log link. This is for numeric stability.

In **rstanarm** use `r rdoc("rstanarm", "stan_glm")` for a Poisson GLM.

-   **rstanarm** vignette on [count models](https://cran.r-project.org/web/packages/rstanarm/vignettes/count.html).

## Negative Binomial

The [Negative Binomial](https://en.wikipedia.org/wiki/Negative_binomial_distribution) model is also used for unbounded count data,
$$
Y = 0, 1, \dots, \infty
$$
The Poisson distribution has the restriction that the mean is equal to the variance, $\E(X) = \Var(X) = \lambda$.
The Negative Binomial distribution has an additional parameter that allows the variance to vary (though it is always larger than the mean).

The outcome is modeled as a negative binomial distribution,
$$
y_i \sim \dBinom(\alpha_i, \beta)
$$
with shape $\alpha \in \R^{+}$ and inverse scale $\beta \in \R^{+}$, and $\E(y) = \alpha_i / \beta$ and $\Var(Y) = \frac{\alpha_i}{\beta^2}(\beta + 1)$.
Then the mean can be modeled and transformed to the
$$
\begin{aligned}[t]
\mu_i &= \log( \vec{x}_i \vec{\gamma} ) \\
\alpha_i &= \mu_i / \beta
\end{aligned}
$$

**Important** The negative binomial distribution has many different parameterizations.
An alternative parameterization of the negative binomial uses the mean and a over-dispersion parameter.
$$
y_i \sim \dnbinalt(\mu_i, \phi)
$$
with location parameter $\mu \in \R^{+}$ and over-dispersion parameter $\phi \in \R^{+}$, and $\E(y) = \mu_i$ and $\Var(Y) = \mu_i  + \frac{\mu_i^2}{\phi}$.
Then the mean can be modeled and transformed to the
$$
\begin{aligned}[t]
\mu_i &= \log( \vec{x}_i \vec{\gamma} ) \\
\end{aligned}
$$

### Stan

In Stan, there are multiple parameterizations of the negative binomial distribution:

-   `neg_binomial_lpdf(y | alpha, beta)`with shape parameter `alpha` and inverse scale parameter `beta`.
-   `neg_binomial_2_lpdf(y | mu, phi)` with mean `mu` and over-dispersion parameter `phi`.
-   `neg_binomial_2_log_lpdf(y | eta, phi)` with log-mean `eta` and over-dispersion parameter `phi`

Also, `rstanarm` supports Poisson and [negative binomial models](https://cran.r-project.org/web/packages/rstanarm/vignettes/count.html).

### Link functions

In many applications, $\lambda$, is modeled as some function of covariates or other parameters.

Since $\lambda_i$ must be positive, the most common link function is the log,
$$
\log(\lambda_i) = \exp(\vec{x}_i' \vec{\beta})
$$
which has the inverse,
$$
\lambda_i = \log(\vec{x}_i \vec{\beta})
$$

### Example: Bilateral Sanctions

A regression model of bilateral sanctions for the period 1939 to 1983 [@Martin1992a].
The outcome variable is the number of countries imposing sanctions.

```{r results='hide'}
mod_poisson1 <- stan_model("stan/poisson1.stan")
```

```{r}
data("sanction", package = "Zelig")
```

## Negative Binomial

The [Negative Binomial](https://en.wikipedia.org/wiki/Negative_binomial_distribution) model is also used for unbounded count data,
$$
Y = 0, 1, \dots, \infty
$$
The Poisson distribution has the restriction that the mean is equal to the variance, $\E(X) = \Var(X) = \lambda$.
The Negative Binomial distribution has an additional parameter that allows the variance to vary (though it is always larger than the mean).

The outcome is modeled as a negative binomial distribution,
$$
y_i \sim \dnegbinom(\alpha_i, \beta)
$$
with shape $\alpha \in \R^{+}$ and inverse scale $\beta \in \R^{+}$, and $\E(y) = \alpha_i / \beta$ and $\Var(Y) = \frac{\alpha_i}{\beta^2}(\beta + 1)$.
Then the mean can be modeled and transformed to the
$$
\begin{aligned}[t]
\mu_i &= \log( \vec{x}_i \vec{\gamma} ) \\
\alpha_i &= \mu_i / \beta
\end{aligned}
$$

**Important** The negative binomial distribution has many different parameterizations.
An alternative parameterization of the negative binomial uses the mean and a over-dispersion parameter.
$$
y_i \sim \dnbinomalt(\mu_i, \phi)
$$
with location parameter $\mu \in \R^{+}$ and over-dispersion parameter $\phi \in \R^{+}$, and $\E(y) = \mu_i$ and $\Var(Y) = \mu_i  + \frac{\mu_i^2}{\phi}$.
Then the mean can be modeled and transformed to the
$$
\begin{aligned}[t]
\mu_i &= \log( \vec{x}_i \vec{\gamma} ) \\
\end{aligned}
$$

In Stan, there are multiple parameterizations of the

-   `neg_binomial_lpdf(y | alpha, beta)`with shape parameter `alpha` and inverse scale parameter `beta`.
-   `neg_binomial_2_lpdf(y | mu, phi)` with mean `mu` and over-dispersion parameter `phi`.
-   `neg_binomial_2_log_lpdf(y | eta, phi)` with log-mean `eta` and over-dispersion parameter `phi`

Also, `rstanarm` supports Poisson and [negative binomial models](https://cran.r-project.org/web/packages/rstanarm/vignettes/count.html).

### Example: Economic Sanctions II {ex-econ-sanctions-2}

Continuing the [economic sanctions example](#ex-econ-sanctions-2) of @Martin1992a.

### References

For general references on count models see @GelmanHill2007a [p. 109-116], @McElreath2016a [Ch 10], @Fox2016a [Ch. 14], and @BDA3 [Ch. 16].