0
$\begingroup$

Apologies if this question is quite simple, but I struggle with stats and wondered if you could help.

I have a data set which has days of the year (1-365/366) from 2004-2013 and a variable with the number of earthquakes per day per year for this same period. I have subset my data into two: one into summer earthquakes and one into winter earthquakes as my research question is looking at whether glacial ice loss in summer increases the number of earthquakes compared to that in the winter months.

I am wanting to see if there is a significant difference between summer and winter? However understanding what test to do is proving tricky.

Any help would be greatly appreciated!

$\endgroup$
1
$\begingroup$

This approach would throw away most of the time detail in the data and is thus unnecessarily coarse. There is no physical reason to expect that there will be a step from winter to summer or that any winter or summer is homogeneous. Rather, I would expect if anything smooth dependence on time of year (except insofar as melting might not be continuous through the year, depending on location). (I am an author on various papers in glaciology, but contributing always on the more statistical side.)

The only justification I can imagine for treating time of year as just summer or winter might be that it seems to bring the problem within the range of techniques taught in an introductory statistics course. If you are a student on such a course, then the rest of this answer is more than you want, but might be of interest to anyone interested by the question.

An appropriate model, at least as a starting point, might be a Poisson regression of daily counts on time of year represented as sine and cosine terms. Make sure that days with no earthquakes are explicit in the data as zeros.

Thus assuming a day of year variable running from 1 to 366 I would construct

fraction of year $=$ (day of year $-$ 0.5) / (365 or 366) $=: f$

and then terms $\sin 2\pi f, \cos 2\pi f, \sin 4\pi f, \cos 4 \pi f, \dots$ for a Poisson regression. Whether there is dependence on time of year is then directly tested by whether the coefficients on sine and cosine terms are jointly zero. The model should be of direct scientific interest as indicating not just whether there is a seasonal effect, but also what form it takes.

The model could further be extended by having calendar year or some appropriate temperature or other climatic variable(s) as extra predictors.

One discussion of such trigonometric (Fourier, periodic) regression with tutorial flavour is accessible here.

Returning to your approach: a further limitation would seem to be that at most you have summer and winter counts for 10 years, which by any standards is a very small sample. (Naturally we take it that the data are all those available, a fact you cannot mitigate.) I suppose that you could put those into a $t$ test or Wilcoxon-Mann-Whitney test. I have to be queasy about that because it would ignore any time dependence between years. Also, the structure is odd: you have a kind of pairing, but only arbitrarily (do you pair a summer with the previous winter or the following winter?). So, the summers and winters are perhaps best treated as unpaired, but as said that ignores any time dependence. The more I think about the sketched approach, the more problematic it seems.

See also other threads mentioning trigonometric regression, e.g.

How to prove statistical significance between the 4 seasons

$\endgroup$
1
$\begingroup$

Thanks for your response. I'll go into a bit more depth regarding my research as above was a very simplified description. I am looking at whether there is a seasonal pattern of earthquakes related to general ice mass gain and loss in Iceland. My hypothesis is that with ice loss during the ablation season, earthquake activity will increase due to an 'unloading' like mechanism (similar to small scale GIA). The 10 years of data is all I could get so yes, this is a limitation.

I have done numerous graphs that seem to show a peak in seismic activity between days 230-310 (julian days) which is the general consensus for each year.

Therefore I wanted to subset this section of data and compare whether this was statistically significant to the rest of the year? This was one of my first thoughts so just wanted to see if this was viable and what test to use?

Your response is very interesting to my previous question, any more thoughts?

kind regards!

$\endgroup$
  • $\begingroup$ This should be an EDIT to your question, as it is not really an answer. The idea of taking the part of the data when an effect is most obvious and testing that for significance against the rest of the year is quite unsound. It's like comparing the tallest people in the population with everyone else. Otherwise your extra detail is helpful, but doesn't change my advice to model the effect across the entire year. $\endgroup$ – Nick Cox Mar 8 '17 at 16:19
  • $\begingroup$ Testing hypotheses is over-rated any way. You can see there is an effect that makes physical sense, so the top question is to quantify the effect, not to sanctify its existence. $\endgroup$ – Nick Cox Mar 8 '17 at 16:30

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.