Note: I don't have formal training in statistics, so please excuse any mistake I've made in my question.

I have two arrays, each with the time-series data of a binary variable, like this:

A: 001101011010100100101001...
B: 100101000101001100101101...

The time-series are of length $N$ (they can vary). (Edit: Thanks to @Mai for pointing out that I've said "time-series" here, but I only care about their "spatial" correlations, so the samples for each time-step can be treated as independent samples.)

The goal is to determine whether $A$ and $B$ correlate (negative or positive) with one another or not, but to limit the chance of a false positive to $0.0001\%$ (for example). This obviously means raising the probability of a false negative - and that's fine for my use case. Also, to be clear, I don't care about the strength of the correlation - only whether or not they correlate at all.

A simple approach would be to look at the difference between $P(A|B)$ and $P(A)$. If there is a difference, then they correlate. This obviously will give a lot of false positives, since the time-series are of length $N$, which is finite.

So I had the idea that I'd get the absolute value of the difference between $P(A|B)$ and $P(A)$, and then subtract the maximum absolute difference that we'd "expect" to see between those values in $99.9999\%$ of experiments between two independent binary variables with $P(A)$ and $P(B)$ where the time-series are of length $N$. As far as I can see, that would cause the probability of a false positive to be $0.0001\%$.

Is this a reasonable approach? Is there a simpler way? I'm actually dealing with trillions of pairs of binary variables, and so performance is a consideration. If I went with the above approach, then I think I'd need to use a combination of normal and poisson approximations to the binomial distribution to be able to handle the full range of $N$ and $p$ values. Thanks!


1 Answer 1


I think you can do this with a chi-squared test or Fishers exact test!

Just collapse your data to the needed format:

enter image description here

With the chi-squared test you can calculate the probability that your 2 measurements are independent or not. I think this is what you want to know.

  • $\begingroup$ I think the fact that they are time series implies they are not sequences of independent trials. Not sure. $\endgroup$ Commented Nov 13, 2019 at 14:22
  • $\begingroup$ Thanks! This seems like just what I'm after! @Mai I have updated the post to clarify that the samples can be treated independently - thanks! $\endgroup$
    – Mifa
    Commented Nov 13, 2019 at 18:14

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.