0
$\begingroup$

I have data on two events over a 25-year span of time. The data look like (example below):

Year  Event_1  Event_2
1990  1000     800
1991  850      750
1992  1050     850
.     .        .
.     .        .

And so on and forth. I'd like to see whether Event_1 and Event_2 are correlated. The figures in the example above are simply frequencies of occurrence of both events. I understand Pearson's product-moment correlation coefficient, but I'm not sure whether that's the best strategy for these data.

I've read about cross-correlation in this other post, but I've never done time-series analysis before. Regarding the data:

  1. I'd like to see if Event_1 and Event_2 over time.

  2. I suspect that Event_1 might cause Event_2, but it's just a guess (it could well be the opposite).

Any reference to the literature to get me started with this will be very appreciated.

$\endgroup$
  • $\begingroup$ How could you look at whether events 1 and 2 are correlated within each year (rather than over years) if both are only observed once? $\endgroup$ – Patrick Coulombe Oct 31 '17 at 17:41
  • $\begingroup$ What are the numbers? Are they the frequencies of the events? It would be very helpful if you told us what the events actually are. $\endgroup$ – Zahava Kor Oct 31 '17 at 18:49
  • $\begingroup$ @PatrickCoulombe I edited OP as that bit was ill-posed. I wanted to mean that if the two events are correlated over time in some sort of trend, does this only hold for "over time" as a whole or also for each year-point? (Not sure if this answers your question). $\endgroup$ – Enoch R. Nov 4 '17 at 10:27
  • $\begingroup$ @ZahavaKor They are frequencies, yes. Just a figure of how many times that event happened in that year. $\endgroup$ – Enoch R. Nov 4 '17 at 10:28
1
$\begingroup$

Seems like a regression problem with year.

Event_1 = Event_2 + Year + Event_2*Year

Well I thought I would expand on this a little more with the commentary below.

You care about looking at the correlation between two events with auto-correlation as an issue. For example, event 1 and event 2 might increase, but that might be because they are correlated and it might be due to auto-correlation.

The trick is to control for this within a model.

You can do something like a simple trend model as well: yt−yt−1=f(xt)+ error. Where yt-yt-1 is just the year to year change. f(xt) is the trend for x and then you have error. Same thing as I described with a small difference in notation.

There is one other thing you can do: a logarithmic model. You will then get an idea if they are proportionately changing.

$\endgroup$
  • $\begingroup$ Thanks, this sounds like a promising point to get started with time-series analysis. Do you happen to know of any reading you'd recommend to get started with these issues? A primer-like text would be amazing. $\endgroup$ – Enoch R. Nov 4 '17 at 10:33
  • $\begingroup$ There are some solid texts from the series for statistics for social scientists. I enjoy those. But that aside, there are some really solid videos on youtube. $\endgroup$ – JWH2006 Nov 6 '17 at 13:17

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.