My take on luterbacher is here, The dutch and the swiss temperatures from historical sources are a good predictor for central europe, no need to add poor quality proxies.

Esper:

“Die Zusammenfassung von lokalen Jahrringchronologien, die aufgrund der Standardisierungsverfahren keine mehrhundertjàƒ⣨rigen Trends enthalten, um niederfrequente Klimaphàƒ⣮omene wie Mittelalterliches Optimum oder Kleine Eiszeit zu studieren (CROWLEY & LOWERY 2000; MANN et al. 1999), sollte vermieden werden.”

]]>“Combined local treering chronologies which – based on standardising – don’t contain multicentennial trends, should not be used to study lowfrequent climate phenomena like the little ice age or the medieval optimum [sic]”

The assumption of the model is that the independent variables (on the right hand side) contain additive errors, in which case OLS estimates are inconsistent. OLS minimizes the vertical distances from observations to the regression line on the assumption that all the errors are additive on the dependent variable. If a portion of the error term is due to errors in independent variables, the angle of the line connecting the observation to the regression line needs to be adjusted prior to minimizing the sum of squared distances. Varying it from 90 degrees to 0 is equivalent to attributing the fraction of error variance attributed to the independent variable from 0 to 100%. A simple treatment of the problem is to estimate the slope coefficients under the full range of variance attribution, and if the coefficient doesn’t change then E-I-V isn’t likely a problem. But this is not a formal solution.

The approach used in Hegerl et al. assumes the variance of the error on the right hand side is known from other sources. Thus, having assumed the existence of the problem, it is then conveniently assumed away.

The formal solution to the problem is to use an Instrumental Variables estimator, where “Instrumental” is used in the econometrics sense, not as a fancy term for “thermometer”. IV estimation involves projecting the independent variables onto exogenous regressors that are asymptotically uncorrelated with the measurement error but asymptotically correlated with the regressor itself, then using the projection on the right hand side of the regression. If the instrument set is constructed to satisfy the asymptotic conditions, the least squares slope coefficients will be consistent.

I got the impression from the meeting that the “Total Least Squares” estimator, ca. 1878, is the method used in “signal detection” regressions that figure so prominently in the IPCC reports.

]]>I had previously tried to learn what the sites were as an IPCC reviewer, but was told that that was none of my business and that IPCC reviewers were not expected to check studies against data or to request data from authors. (I requested the information both from IPCC secretariat and later from Hegerl, and was told that if I made any further attempt to obtain data from authors, I would be expelled as an IPCC reviewer.) I mentioned this refusal (and a like refusal by D’Arrigo) in my presentation.

Any responses from the Panel to your comments?

]]>