My company gathers all sorts of series for page views, unique visitors and many other user data (collected by our web servers). As I see it, these sets of series are discrete since they are all collected by computer systems and software and there is no 'intermediate' value between the collected numbers. Also because of that I don't have a way to measure its 'estimation error'.

These series are compared monthly and my company moves forward according to the percentage increase/decrease in the values. But my managers never take in consideration its standard deviation nor its 'possible error'. The only statistical analysis they do is the percent variation between series.

What would be a sounder, proper way to analyses the monthly variation between each series? I aim to derive the 'real' variation in the number of visitors, page-views, etc.


What you have is called panel data: multiple time series which are "connected" in some way. Panel data are commonly analyzed in economics (where the connected time series would be GDP, unemployment rates etc.). I don't have a reference handy, but the Wikipedia page may be helpful. Dig into this. ANOVA is not helpful for time series or panel data.

Alternatively, you can start by reading up on time series analysis and forecasting (e.g., with this free online textbook), ignoring any connections between your series, at least at the beginning.

On the other hand, while your data is discrete in the sense that there cannot be 2.3 or 5.7 page views, you are probably dealing with sufficiently high volumes that you can safely use standard methods which assume a (non-discrete) normal distribution on errors.

  • $\begingroup$ about the last paragraph, by "dealing with sufficiently high volumes that you can safely use standard methods which assume a (non-discrete) normal distribution on errors" what methods you have in mind? $\endgroup$
    – Draconar
    Nov 13 '12 at 17:58
  • $\begingroup$ Basically, the basic panel data or time series analysis (ARIMA etc) methods you read about in textbooks all assume normality. There are special methods like INAR for count data, i.e., for time series that take on values like 0, 1, 2, 3, ... However, if your values are usually "high enough" (>20), you should be able to use the basic approaches and not need to worry about the more complex ones. $\endgroup$ Nov 13 '12 at 19:12

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.