Testing for bias in the climate record.

calculate the distribution of such an influence, trends at each of the 144 U.S. stations were simulated with an annual cycle calculated from observed daily data (the first five harmonics captured most of the variance in the annual cycle) as well as the autocorrelation and variance of the standard monthly means. An auto-regressive model (5) of order 1 (AR-1 model) was fit to the residuals of the daily temperatures minus the value of the sum of the harmonics, and the variance of each month was calculated in the same manner. The data at each station were simulated with the use of the mean appropriate for the day of the year calculated from the harmonics and the perturbation about this mean from the AR-1 model. Trends were then calculated using the same calendar offset as calculated from precession of the Earth's orbit, as described above, and subtracted from the trends without the offset. The results were stratified by the magnitude of the persistence and variance term in the AR-I model. A 2 x 2 continency table was developed with class limits based on the median value of persistence and variance for all stations and all months. A significant portion of the difference in trends was found to be a result of the vagaries of weather and had nothing to do with precession of the orbit (Fig. 1). We show the result for the minimum temperature (Fig. 1), but results are similar for the maximum and mean temperature. The effect of perihelion should be most apparent in the longest daily data set available, the CET. After incorporating the perihelion dates into the CET data set, spring (March, April, and May) warmed relative to the standard trends by 0.42°C over the 219 years and autumn cooled by 0.46°C. Using standard monthly data, we found that winter warmed relative to summer over 1772 to 1990 by I.1 °C; allowing for perihelion reduced this difference to 1.0°C. For the last 100 years (1891-1990), trend differences for the CET series are within the +0.250C range, falling in the center of the U.S. distributions (Fig. 1). Inspection of the differences in trends calculated with and without the shift in calendar dates indicated that a bias was introduced during the transition seasons. It averaged between 0.050 and 0.10°C per century, with a positive bias during spring and a negative one during the autumn. This was due to shifting the calendar to later in the season. During summer and winter the slight shift of 2 days showed little bias, and even a 5-day offset in the CET time series produced only small differences. In sum, we find Thomson's concerns about the manner in which climatologists have calculated trends on monthly and annual time series to be of little consequence during the instrumental climate record.