There are four Dawson YT sites in the Canadian data set, three of which are used in the GHCN station list:

CA002100400 64.05 -139.43 320 DAWSON
CA002100LRP 64.05 -139.13 370 DAWSON
CA002100402 64.05 -139.13 370 DAWSON A GSN 71966

2100400 daily data starts in 1897 and ends in 1979 (http://www.climate.weatheroffice.ec.gc.ca/climateData/dailydata_e.html?timeframe=2&Prov=CA&StationID=1534)
2100LRP daily data starts in 1995 and is current (http://www.climate.weatheroffice.ec.gc.ca/climateData/dailydata_e.html?timeframe=2&Prov=CA&StationID=10194)

2100402 daily data starts in 1976 and ends in 2007 (http://www.climate.weatheroffice.ec.gc.ca/climateData/dailydata_e.html?timeframe=2&Prov=CA&StationID=1535) This is one of the stations that I noticed data issues with regards resolution. Starting June of ’95, the daily Tmin and Tmax readings are in X.0 or X.5.

I suspect 2100LRP and 2100402 are the same station but different equipment since Env. Can has given the LRP station the WMO ID# 71966 but GHCN is giving 2100402 the WMO ID# 71966.

Just a little confusing…

Relatedly there is the question of what happens if you do the average that way?

I can see that this could reduce the variance of the data. I’d be very curious to know if it also changed the accuracy. [Perhaps I need to figure out how to get the GISS data into R and run the tests unfortunately I've yet to get R working for me]

1. Do not assume that normal distribution statistics apply simply because a curve looks bell-shaped. Test the maths.

2. Aggregation of many observations will generally give a better precision estimate, but might say nothing about accuracy (bias) improvement.

3. Where possible, use standard units of measurement. Imagine that the ranges derived above could be anomalies versus a stipulated period, or Celcius, or degrees F, or degrees absolute. Each choice gives a different appearance to the plot.

Re the question of massaging of Canadian data before its is processed more. I have a paper “Updating Australia’s high-quality annual temperature dataset” by Paul Della-Marta, Dean Collins and Karl Braganza. Rec’d June 2003. Published Aust. Met. Mag. 53 (2004) 75-93. I have it in paper form only.

Some quotes: On UHI correction “Ultimately a subjective decision was made for each station as to whether it was likely to have been influenced by urbanisation.” Point of illustration – this procedure, repeated many times, might have evened out the plus and minus precision deviations, but it might have done nothing for bias).

On station quality “However, at present, the information in the database is only reliable since the mid to late 1990s” (GS: This might have improved since).

On this thread “Some records contain(ed) many years in which annual mean values were estimated by Torok and Nicholls (1996) using a reference series based on highly correlated neighbour series”.

It is unfair to cite these comments out of context, but one might see some similarity with the procedures of the thread above. At last inquiry, I was invited to go to the BOM library to read the PhD thesis of S.J. Torok 1996 from Melbourne University. I gained the impression that much Australian temperature data handling is based on this thesis.

if we assume 1000 random estimateions with an error of 1 degree, and the error is randomly and normally distributed, the error in the average is:
individual error / sqrt(number measurements) = .03 degrees.
This won’t take into account any systematic errors/biases, such any tendancy for measurements to be more likely missing in snow storms etc.

If there is a systematic bias due to the ‘snow storm effect’, this may not effect the trend unless the bias changes over time. For instance if there are more or less snow storms now than 20 years ago, or if measurers today are more/less devoted and less/more likely to skip a measurement today in a snow storm than they were 20 years ago.

Yes, I had the same thought with a different twist. There may be more likelyhood of missing data on cold/windy/stormy/snowed-in days than on a balmy winter days.