The Making of the Hockey Stick - The Hockey Stick and the Climate Wars - page 32

networks of tree ring data (e.g., North America, Eurasia) in terms of a small number of representative
PC series. With the tree ring data that dominate the midlatitude continents 11 represented by a few
leading patterns, the handful of coral series from the tropics and ice core series from the Arctic
would have commensurate representation in the dataset. As far as each proxy's opportunity to help
tease out the past patterns of climate variation was concerned, it was now a fair fight.
Our statistical method established a relationship between the proxy data (which extend several
centuries back in time) and the modern surface temperature record during the period of the twentieth
century where both datasets overlap. This procedure required finding a way to represent efficiently
the information from the modern instrumental record. Spanning a century, the instrumental surface
temperature dataset contained more than a thousand months of data at more than a thousand locations
around the globe. 12 But only a handful of patterns could potentially be captured by our noisy proxy
dataset. The temperature data thus also needed to be represented in terms of a modest number of its
most prominent underlying patterns. We recognized that the same tool we had used to simplify the
representation of dense tree ring data networks—PCA—could again prove useful. Using PCA, we
could represent the key information in the instrumental temperature dataset with just a dozen or fewer
distinct patterns. The leading temperature pattern related to the overall warming of the globe, while
subsequent patterns related to phenomena such as El Niño, the North Atlantic Oscillation, and the
Atlantic Multidecadal Oscillation. 13
Figure 4.1: PCA Example: Spatial and Temporal Variations in Temperature Data
This example shows how PCA can be used to characterize efficiently a global temperature dataset (top) in terms of a small number of
patterns in space and time. In this case, just two patterns (middle and bottom) characterize the data. The horizontal axis is time in years,
and the vertical axis depicts the relative departure of temperature in degrees from the average over the baseline, 0 (in this case, the
average over the full hundred-year period).
With the aid of our statistical method, we simply needed to figure out the combination of these
dozen or so temperature patterns that most closely matched the behavior of the proxy series we had
during the twentieth-century period of overlap. Once those relationships had been established, they