Thursday, June 30, 2016

Seasonal Temperature Correlations

Last week reader Gary posed the question of whether there is a connection between this summer's weather in Fairbanks and the extreme heat in the desert southwest of the Lower 48.  Intrigued by the question, I created a set of spatial correlation maps for monthly mean temperatures through the year.  To illustrate how this works, the map below shows the correlation of June mean temperature between Fairbanks airport and every other point on the map.  As expected, the correlation falls off with distance, and the scale of good correlation is quite small at this time of year.  Temperatures in south-central Alaska only have a modest correlation with Fairbanks temperatures, and the correlation is about zero for stations on the eastern North Slope.

Here's a map showing the results for all of North America.  I created a smoothed contour map rather than showing all the points, because there are so many stations in the lower 48 that all the markers overlap.  We see that Fairbanks temperatures do indeed have a modest negative correlation with temperatures in parts of the southwest US at this time of year, and there is a small positive correlation with the northern Plains and upper Midwest.


Below are the maps for each month of the year (click to enlarge).  A surprising result is that the scale of very good correlations (R>0.9) across Alaska and western Canada is greatest in September; I would have expected this to occur in winter.


























Update July 2: below are the corresponding maps for monthly precipitation.  Note that I've used the rank correlation coefficient rather than the Pearson correlation coefficient, because precipitation amounts can be highly non-Gaussian in drier areas even on a monthly timescale.

Also note that there are far fewer stations reporting precipitation than temperature across the north (at least in the GHCN data set), so the continent-wide map shows large areas of missing data, and the results of the smoothing algorithm (Cressman analysis) are not realistic in Alaska.  I recommend paying attention only to the very broad-scale features of the maps.  But it's interesting to see that September again has a relatively good correlation of anomalies across Alaska.


























12 comments:

  1. I don't understand how you are creating the correlations. Correlations are meaningless without trends. So what trends are you using?

    ReplyDelete
    Replies
    1. They are correlations over time, 1981-2010. So we have 30 data points for Fairbanks June temp, 30 data points for Barrow June temp, and we find the correlation coefficient. Repeat for thousands of stations.

      Does that help? Sorry for the confusion.

      Delete
    2. Forgive my confusion. I interpret your explanation in two possible ways:

      1. Given the 30 year period, the warmer it is in Fairbanks during any given month, the colder the blue areas are. Thus a warm January in Fairbanks is cold in Orlando, while a cold winter here is warm there.

      2. Over the 30 years period, the Florida temps have been getting colder with respect to Fairbanks.

      I suspect that option 1 is right but it's not clear.

      Delete
    3. Eric I think your option 1 is the closest. My understanding is that the colors are not related to specific warm/cold but instead related to the correlative relationship between Fairbanks' temps with temps elsewhere.

      My question is, what are the original data points? Are they deviation from the long term mean? For instance, if it's a day in January in Fairbanks that is 20 degrees cooler than normal, does that make a day in Tanana that is 20 degrees cooler than normal a "1.0"? And a day in Quebec City that is 20 degrees warmer than normal a "-1.0"?

      Delete
    4. Option 1 and Andy's explanation are correct.

      The data points are just mean monthly temperature. The correlation calculation takes out the sample (i.e. 1981-2010) mean:

      https://en.wikipedia.org/wiki/Pearson_product-moment_correlation_coefficient

      The variance at each site doesn't matter, as we are looking at correlation. So for example the following two series are perfectly correlated (R=+1.0) even though the variance is different.

      3 -1 -5 9 2
      6 -2 -10 18 4

      Delete
  2. Great map set Richard! May I share?

    ReplyDelete
    Replies
    1. Yes, of course. It would be trivial to create similar maps for any other location if you're interested.

      Delete
  3. I like the visual presentation of correlation between different NA zones...even when it's a negative less than 0.

    Which to me means there's a relationship although opposite in temperature...when we're warm; they're cold apparently.

    How about one for precipitation?

    Gary

    ReplyDelete
    Replies
    1. Yes, some places are inversely correlated. A ridge upstream is usually associated with a trough downstream at a certain preferred wavelength, and vice versa.

      Shouldn't be too difficult to add precipitation, although I'll have to use Spearman correlation rather than Pearson.

      Delete
    2. It'll be interesting to potentially compare temps and precip via your visual methods.

      Or...simply regress sales of umbrellas per capita by location (smile).

      I looked up Pearson/Spearman but it's been 48 yrs since Stat classes so I got the pre-exam shakes and quit. Now with plug-in PC programs its easier than hand computation was in the age before calculators

      Gary

      Delete
    3. I've added the precip correlation maps. The paucity of data makes them less informative than the temperature maps, but there are a couple of interesting features.

      Delete
  4. Thank you for adding the precip correlation maps! It's a bummer that NCEI has precip (v2) and temps (v4) in separate files with separate identifiers. Makes it hard to join data.

    ReplyDelete