I have two databases which include amount of passengers between two routes. One is the full dataset, while the other is supposedly a 10% sample.

So for example, the full database will display a route (AB)= 4.000 passengers while the 10% sample should display route (AB) = 400, obviously. However that doesn't always happen, but I still want to be able to validate with a certain confidence interval how much can my 10% sample variate.

I've been looking at different methods but so far bootstrapping for example doesn't seem to fit here as I can't figure out how to resample my route info. Meaning that I can resample the routes but then I'll have an absurd number for a certain route.

Edit: About the last paragraph. From what I understand from bootstrapping, I would resample the amount of passengers traveling on those routes. However, this is a fixed number meaning that there's no shuffling. The only resampling I could do would involve attributing a new value for the amount of passengers on route AB. For example resampling those numbers would either be based on the other routes, meaning that I could end with a number of 1 or 5 million. Or I would just resample that value based on a percentage which then I'm not sure how it would allow me to validate the 10% sample.

Saying so, I'm not sure how to validate if a 10% sample is indeed a 10% sample as there are no characteristics on the routes being analysed. It's just a sum of passengers, without any other characteristics.

  • $\begingroup$ Your last paragraph is not clear to me. What is the obstacle you are facing? What is "absurd"? $\endgroup$ – rolando2 Mar 1 '18 at 12:53
  • $\begingroup$ @rolando2 I've added some more info. The problem is that I basically have to show that the 10% sample is representative, and indeed 10%. Meaning that it falls into a confidence interval (lets say 99%) of what would be possible to be a 10% sample. $\endgroup$ – FilipeTeixeira Mar 1 '18 at 13:23

You could use a one way chi-squared analysis. Suppose you have 5 routes and their proportions of the total data set are .3, .3, .2, .1, .1. In your sample, suppose you get 100, 100, 70, 30, 25. Then, in R:

props <- c(.3, .3, .2, .1, .1)
samp <- c(100, 100, 70, 30, 25)
chisq.test(samp, p = props)

tests the hypothesis that the proportions in sample are equal to those in the population.

Similar methods are available in SAS and probably most other statistics packages.


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.