climate-sensitive tree-ring records that have been thoroughly documented in the refereed scientific literature. Second, we chose the longest of those records in order to maximize the temporal coverage, and tried, as far as possible, to select samples from representative geographical areas.

A number of problems are inherent in any climate reconstruction. These problems are discussed in detail in the original published papers; however, we will briefly highlight here the more critical ones as they may affect the results presented below. Because we have analyzed changes in variability, temporal changes in the composition of the tree-ring network used for the reconstructions will, in general, affect the high-frequency variance, and to some degree the low-frequency variance as well. In interpreting the temperature reconstructions shown here, we have taken into account these potential sources of biases as much as possible.

Regardless of whether one uses the width of the annual growth rings or the maximum latewood density to reconstruct a particular climate variable (generally, growing-season temperature and/or precipitation), a process of standardization is required to account for different rates of tree growth as a function of age. This standardization is achieved by fitting a growth curve to the series of annual values, usually a cubic smoothing spline (see Cook and Kariukstis, 1990), and taking residuals about the smoothed curve. The degree of smoothing and the functional form of the growth curve partly determine the spectral properties of the residual series. In order to develop a climate reconstruction from these records, it is necessary to formulate a transfer function to convert the tree-growth index into, say, a temperature index. The procedure usually involves a calibration phase, in which a set of regression coefficients is derived that convert the tree-growth parameter into a climatic estimate, and a verification phase (see, e.g., Cook