Inference About Trends in Temperature Data After Controlling for Serial Correlation and Heteroskedastic Variance

Global surface temperature anomalies post-1958 are examined for evidence of significant trends. The routine application of a linear least-squares method is critiqued. Testing for model misspecification in the form of serial correlation and heteroskedasticity in the residuals leads to consideration of a piecewise trend with lagged anomalies and a linear model of the error variance. While the naï ve least-squares model suggests a highly significant trend post-1979, correcting the known specification errors leads to the finding of a reduced trend magnitude and evidence that the trend is insignificant, i.e. within the bounds of random noise. This points to the need for careful treatment of time series variables in order to avoid spurious inference. Acknowledgments I thank Kirill Kondratiev for his invitation for me to address this topic.