It has been shown that a sum of independent observations from a logarithmic series distribution gives a negative binomial distribution when the number of terms in the sum has a Poisson distribution.
I would like the earliest reference to an equivalent way of looking at this, which I think is more intuitive (I assume I wasn't the first to think of it). Suppose that events involving k individuals have a Poisson distribution with rate proportional to p^k / k. If these events are independent then the total number of individuals involved has a negative binomial distribution.