Please bear with me on this question, it's a follow on from my other question on a similar topic but in this question I wish to understand how to handle periodicity when testing a hypothesis with categorical data.

I have discrete measurements of the number of call attempts on a cell in a wireless network. The counts are categorised into 21 different classes and these 21 classes represent the various distance each user was when they tried to establish a call i.e. Class 0 represents 0-234m from the cell site, Class 1 represents 234-600m and Class 13 represents 15000m-18500m from the same site. A count of 300 in Class 1 represents 300 call attempts 234m-600m away from the cell site. As you can see, the distances each class represents are non uniform.

As you can imagine, cellular traffic shifts about to other cells in a network over the course of a day and my measurement on Cell X at 7am is vastly different to my measurement also on Cell X at 2pm and my Cell X that covers an business district has a very different profile from Mon-Fri compared to Sat and Sun. There are multiple seasonalities (15 min, hourly, daily, weekly, monthly) to be considered depending on the sampling period and the location of the physical cell site. A cell could have hourly, daily and weekly periodicity.

This bring me to my point, over a number of weeks I wish to make changes to every cell in the network (12000+ cells) for a small fraction of time, ideally in the middle of the night to minimise network disruption. In simple terms, if my geographic cell size is small before the change, I plan to make it geographically much bigger and allow more customers to access it and therefore I should have more counts in the latter classes which represent distances further away from the cell site. Inversely, if i have a very large geographical cell I intend to shrink it such that I lose all the distant traffic on that cell and the counts of call attempts in the latter classes will dramatically decrease.

To my question, once I implement my change at X a.m for a cell, how can I be confident that my change actually caused the change in the counts for each of the classes rather than the seasonality within the data. I know for traditional time series data we can first and second order difference the data to ensure stationarity but is this possible with categorical time series?


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.