1
$\begingroup$

I'm studying several time series. The variables of interests are dummy variables which take value 1 when a certain event happened, 0 otherwise. I want to find the best distribution to describe these data. I'm using the function fitdistr() of R, which allows me to use as discrete distributions the Poisson and the negative binomial. I've found using a chi-square test that the Poisson describe really well these data (p-values all over 0.90, so I do not reject the hypothesis that the data are distributed following a Poisson).

My question is: can a Poisson distribution be considered a good model to describe binary data? The hypothesis testing suggests that it is, I want to know if there are theoretical drawbacks or statistical errors in such fitting. Thank you.

$\endgroup$
5
$\begingroup$

If you know the only allowed values are 0 and 1, really the only sensible approach is to use the Bernoulli distribution (equivalent to a binomial distribution with $N=1$). The Poisson does converge to Bernoulli as the mean approaches 0, but if the mean value of the sample is greater than a small value (say 0.1), the Poisson will predict values >1 that can't occur in your data. For example for a Poisson with mean=0.1, $P(0)=0.905; P(1)=0.0905; P(2)=0.00452; P(3)=0.000151$ etc., while the Bernoulli would have $P(0)=0.9; P(1)=0.1$.

| cite | improve this answer | |
$\endgroup$
  • $\begingroup$ Indeed there are series with very small mean values (0.0042, 0.02) where the fitting of the poisson is perfect (P-value equal to 1). I have also cases as you pointed out in your example where the poisson gives probability until 4. I have two questions: 1) my series has 215 observations it makes sense fitting a bernoulli where the number of trials is just 1? 2) a binomial would makes sense? 3) the observation may have a sort of dependence, is that a problem? 4) is so conceptually wrong fitting a poisson even if the fitting is very good? Thank you. $\endgroup$ – Kolmogorovwannabe Feb 25 '19 at 7:16

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.