I've searched the web and these forums but couldn't find a subject that resembled my problem. If I missed something then I'd like to apologize upfront.

My problem is this. I have a given probability p, given by an outside agency and the number of successes n from an owned datasource. I'm interested in the population total N and the confidence interval of N.

Background: I know how many passengers there are for a specific bus agency using 2013 data (as in: I have the (micro)data from that specific company). Assumption: this is my n. I also know from transportation figures (from an outside agency) that this specific bus company had a market share (as measured in passengers) of 30% in 2013. Assumption: this is my p. This assumption brings with it certain other assumptions, that is fine.

So now the question is: how many passengers were there in total (assuming this can not be taken from the transportation figures)? ie. what is N and what is the 95% confidence interval surrounding N?

So my questions are:

1) Can I just use the relation: p=n/N, therefore the point estimat for N is: N=n/p? I'd assume so.

2) Can I just use the Clopper-Pearson interval to calculate an interval for p, then using that to get to an interval for N?

3) Suppose the 95% lowerbound for p=plb. Is the 95% lower bound for N then n/plb? Same logic for the 95% upper bound.

4) When calculating the Clopper-Pearson interval, can I just use: n=n, p=p, N=n/p?

5) If any of the answers to my questions 2-4 is 'No', then what is the correct approach?

I realize that using this approach, N is not necessarily an integer. Is that problematic?


1 Answer 1


If we treat the $p$ as a fixed, known quantity (rather than something that was an estimate, as it seems it actually was) and we treat the $n$ as having come from a $\text{binomial}(N,p)$, then I believe this is actually a negative binomial problem.

We can view each passenger as a Bernoulli trial, where 'success' is 'travelled with the particular bus company'. As such, the number of trials (passengers) until the $n$th success is negative binomial $\text{NB}(n,p)$ (in the 'number of trials' form; if we're using the 'number of failures' form we need to add $n$ to the result).

We can use the quantiles of that conditional distribution for $N$ to get a confidence interval.

So for example, with $n=1000$ and $p=0.3$ the interval $(3163,3509)$ contains just over 95% of the probability, and has very close to 2.5% in each tail.

If we treat $p$ as an estimate measured with error, then this becomes a bit more complex, and we'd need some information about the estimation of $p$, or to make some assumptions about it. It might be necessary to use simulation in that case. Dealing with this would tend to make the interval a good deal wider.

However, I'd actually be somewhat more inclined to consider a Bayesian approach for this problem, I think it's a pretty natural fit; incorporating prior information about $N$ and uncertainty in $p$ can both be brought in in fairly natural ways.


Your Answer

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

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