I'm trying to combine percentiles (50th and 90th) from 2 different datasets. I don't have access to the datasets at the time of combining. Apart from the percentiles I have the number of elements on each dataset.

Now, I know that any attempt at combining them will be only an approximation to the real percentile of the combined datasets, but I just want to know what would be the "most" correct solution.

The options I was considering were: - Weighted average - Take the percentile for the biggest dataset

The output will be used to build a graph, as I get these percentiles periodically.

  • $\begingroup$ What else do you know about the two datasets? Are they normal? Do they have the same mean and standard deviation? Without some additional knowledge about the nature of the datasets there isn't anything to work from. $\endgroup$ – Ian Miller Nov 15 '15 at 12:47
  • $\begingroup$ They are not normal and the datasets are totally different, so they have different mean and standard deviation. They are just measurements on the time it takes to execute different requests to a website, but from different computers $\endgroup$ – hortega Nov 15 '15 at 15:00
  • $\begingroup$ What do you want the combination to do? Does it make sense to combine the datasets? You can add them. You can multiply them, too If they are time taken on different computers, it probably doesn't make sense unless you are trying to assess the time a user spends waiting. In that case, reporting percentiles may not be useful-people get frustrated at slow response even if it happens much less than 50% of the time. You should make your objective clearer. $\endgroup$ – Ross Millikan Nov 15 '15 at 15:44
  • 1
    $\begingroup$ I want the combined statistics of the dataset to build a graph that will be easier to look at than than having as many graphs as datasets. I want users of this graphs to be able to assess the performance of their service in real time. As I said on the question, I don't have access to the datasets so I can't combine them. $\endgroup$ – hortega Nov 16 '15 at 14:30
  • $\begingroup$ So a few thoughts on website response times: - Response times are generally bounded between 0ms and 5000ms, and often center strongly between 20ms and 200ms. This has the benefit of keeping your graph looking nice. It also means that you can overlay the two datasets in the same graph - and thus bypass the problems associated with merging differing percentiles. - People are often concerned with the "worst case" when dealing with response time. It may be better to just show the larger of the values. This also avoids merging the datasets. If you must, I suspect a weighted average to be best. $\endgroup$ – Mark Roberts May 22 '17 at 19:04

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.