Q1. I want to test if my data are the result of a Poisson process. I need to check if: (1) The number of arrivals N(t) in a subinterval of length t obeys the Poisson(λt) distribution or (2) The inter-arrival times obey the Exponential(λ) distribution.

Let's suppose I go for the 2nd option: I would implement a Goodness-of-Fit Test for checking whether the inter-arrival times can be fitted by the Exp distribution. My question is: could I do the test for all the events in my dataset in one time instead that for each subinterval? In few words: if the DeltaTimes are exponentially distributed within each interval, is it correct to expect them to be exponentially distributed for the total interval (= length of the dataset)?

Q2. Inter-arrival times of Poisson events with constant rate X are independent random variables, each exponentially distributed with mean 1/X. My question is: if I verify that the inter-arrival times are exponentially distributed and if I also verify that they are random (with a test for randomness), could I conclude that my data result from a Poisson process with constant rate?


1 Answer 1


You have a tradeoff between power (e.g. the ability to see as quickly as possible when your process is not following a homogeneous Poisson process) and flexibility (the ability to see that it was a Poisson process the whole time, but perhaps the rate changed at some point, making it non-homogenous).

Note that Poisson processes have several features; there's not just that you have an exponential inter-event time but that the rate is constant and events occur independently of each other.

If you want to test for a homogenous process, looking at the inter-event times makes sense, and in that case you would maximize power by combining all your inter-event times into one sample. You could take counts in many sub-intervals and test for Poisson-ness but I don't think that will give you more power than testing for an exponential-distributed inter-event time.

You could gain substantial power if you have a specific alternative in mind and tailor your test against that alternative. Failing being able to nominate a specific sort of alternative there are several possible tests for an exponential distribution. Indeed, D'Agostino & Stephens' book (Goodness of fit Techniques) has an entire chapter on tests for the exponential -- and it doesn't include smooth tests.

There's some additional specifics on testing (and diagnostics) for exponential distributions in this thread

However you might also consider looking for some time dependence or non-homogeneity that might not be so obvious in the marginal distribution of inter-event times.

Note that however you test it - whether with one test or several tests of different aspects of the process - failure to reject doesn't mean that you have a Poisson process. It simply means you couldn't detect that it wasn't.

That is, the answer to Q2 is "No". You can't verify that the inter-arrival times are exponentially distributed, nor can you verify that they are random. You may be able to see that they probably aren't but failure to demonstrate that doesn't mean that they are. Maybe the effect was just smaller than you could reliably detect at this sample size.

In practice you will almost never have a true Poisson process but may in many cases have something well-approximated by one. With very large samples you'll be very likely to reject even trivial deviations from a Poisson process, even when it would be suitable for every practical purpose you had in mind for it. [Both rejection and failure to reject may not tell you what you need to know.]

For this reason you may be better to look at diagnostic plots that would let you examine deviations from the exponential or Poisson as well as looking at time-dependence or changes in rate over time. These will focus more nearly on how much deviation there is from the ideal case rather than on significance, which will be more useful from a modelling point of view (i.e. it will come a bit closer to answering a question like "will this Poisson-process model be a suitable approximation for my present purpose?")

  • $\begingroup$ Thank you @Glen_b. I completely agree with your statement about hypothesis testing. However: if I do not reject the hypothesis that my DeltaT are exponentially distributed (Lilliefors test) + that they are random (Runs Test) + I observe a straight line in the QQ plot, may I conclude that my data are quite compatible with a Poisson process hypothesis? PS. Which diagnostic plot would you suggest to check for a constant rate in time? $\endgroup$
    – Roland
    Dec 4, 2017 at 19:05
  • $\begingroup$ The sorts of tests you are considering - like tests of goodness of fit and serial dependence - don't answer a question like "is this close enough for my purposes?". Diagnostic displays come closer to doing that (in that you're looking more at effect size than significance) but still don't look at "how much is too much (for my applications)?" $\endgroup$
    – Glen_b
    Jul 31, 2020 at 3:39

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.