I have a process that generates a sequence of events, where each event is of a certain type (let's say $A$, $B$, and $C$). The length of each generated sequence is random.

I wish to count the number of events of each type and calculate the probability that an event is of a certain type. For example, for the sequence $\left ( A, A, B \right )$, the proportion of $A$ events would be $\frac{2}{3}$, the proporition of $B$ events would be $\frac{1}{3}$, and the proportion of $C$ events would be $\frac{0}{3} = 0$.

Now, if I repeat the process $n$ times, I can calculate the sample mean for each of the probabilities. My question is: How can I calculate a confidence interval for these means? Can I use the usual normal procedure as long as my sample is big enough? It seems that I would risk getting an interval that goes below 0 or above 1, which would be meaningless. Also, it would intuitively seem that longer sequences carry more information (better estimates of the probabilities) and should be weighted accordingly.


I will assume that the number of occurrences for A, B, C is marginally binomial for each event type. For large n the normal approximation with the continuity correction can work as an approximate confidence interval. If the true proportion is close to either 0 or 1 the normal approximation will be less accurate and could have lower endpoint negative in the case p is close to 0 and the upper bound could go above 1 when p is close to 1.

Now the Clopper-Pearson method provides a relatively fast method for calculating exact binomial confidnce intervals for any n. So I would suggest going with the exact method. For a good reference, look at Statistical Intervals by Hahn and Meeker

  • $\begingroup$ The events are not really independent -- does this change anything? $\endgroup$ – Vegard Sep 20 '12 at 8:19

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.