Suppose we have the following count data:

|   X   | Freq |
|   0   |  18  |  
|   1   |  32  |  
|   2   |  17  |
|   3   |  10  |
|   4   |   5  |
|   5   |   2  |
| TOTAL |  84  |

We want to construct a likelihood ratio test to see if a Poisson distribution is suitable to describe the data. The likelihood and log-likelihood equations for a Poisson distribution are:

$$ L(\lambda) = \prod_{y=1}^{84} \frac{e^{-\lambda}\lambda^y}{y!} $$

$$ l(\lambda) = \sum_{y=1}^{84}\bigg(-\lambda + ylog(\lambda) - log(y!) \bigg)$$

We can solve for the MLE $\hat{\lambda}$ as follows:

$$ \frac{dl(\lambda)}{d\lambda} = \sum_{y=1}^{84}\bigg(-1 + \frac{y}{\lambda}\bigg) = 0 \rightarrow \hat{\lambda} = \frac{\sum_{y=1}^{84}y}{84} = \frac{\sum_{i=1}^{84}x_i f_i}{84} = \frac{126}{84} = 1.5 $$

The alternative hypothesis is that the data follow a multinomial distribution. The likelihood and log-likelihood equations for a Poisson distribution are:

$$ L(p_0,p_1,p_2,p_3,p_4,p_5) = {n\choose{f_0,f_1,f_2,f_3,f_4,f_5}} p_0^{f_0} p_1^{f_1} p_2^{f_2} \pi_3^{f_3} p_4^{f_4} p_5^{f_5} $$

$$ l(p_0,p_1,p_2,p_3,p_4,p_5) = log{n\choose{f_0,f_1,f_2,f_3,f_4,f_5}} + \sum_{i=0}^{5} f_i \times log(p_i) $$

Using the method of Lagrange Multipliers (with constraint $\sum_{i=0}^{5} p_i = 1$), it can be shown that the MLE's $\hat{p_i}$ are equal to:

$$ \hat{p_i} = \frac{f_i}{n} $$

I know that the likelihood ratio statistic is defined as $ D = -2(l_0 - l_A) \sim \chi_{(6-1)-1}^2 \sim \chi_4^2$, where $l_A = l(p_0,p_1,p_2,p_3,p_4,p_5)$, i.e. the log-likelihood equation for the multinomial distribution.

However, I am not sure what $l_0$ should be. Do I assume that the distribution of the null hypothesis follows a multinomial distribution with $p_i$ ~ Poisson (Case 1), or do I assume that the null hypothesis follows Poisson distribution (Case 2)? I've tried both and I do not get the same answer.

Case 1: $ l_0 = log{n\choose{f_0,f_1,f_2,f_3,f_4,f_5}} + \sum_{i=0}^{5} f_i \times log\bigg(\frac{e^{-\hat{\lambda}}\hat{\lambda}^i}{i!}\bigg) $

Case 2: $ l_0 = \sum_{y=1}^{84}\bigg(-\lambda + ylog(\lambda) - log(y!) \bigg) $


You seem to be using the generalised likelihood ratio test. However, this only works when the alternative hypothesis is a more general version of the null hypothesis, for example when the null hypothesis is that $\lambda = 1$ and the alternative hypothesis is that $\lambda$ is unconstrained (can be anything but 1). Since you're testing Poisson against multinomial, the generalised LRT doesn't really work.

You can still use a chi-squared test to conduct a chi-squared test of goodness of fit, but replace the alternative hypothesis with 'The data do not follow the Poisson distribution'. Here, the chi-squared test statistic $\chi^2 = \sum_{i}\frac{O_i- E_i}{E_i}$ roughly follows the $\chi^2_{6-1-1}$ distribution for large $n$ (where $O_i$ and $E_i$ refer to the observed and expected values of the frequency in the i-th bin).*

Also, I think that if you know the source of the data, you should know whether Poisson or multinomial is appropriate, since they're applied to quite different situations.

*Technically, Pearson's chi-squared test is an approximation of the generalised likelihood ratio test, so you'd still be using that (in a sense).


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.