Ross McKitrick
Professor of Economics 
Department of Economics and Finance
University of Guelph
519-824-4120 x53051
ross.mckitrick [at] uoguelph.ca
  • Home
  • Environmental Economics
    • Textbook
    • Theory of Policy Design
    • Environment and Health
    • Air Pollution and Growth
  • Global Warming
    • Climategate
    • Economics/Climate Policy
    • Emission Trends
    • Fingerprinting
    • General Overviews / Misc
    • IPCC Reviews and Critiques
    • Model Testing / Hiatus
    • Paleoclimate/Hockey Stick
    • Submissions/Responses to Gov't Inquiries
    • Temperature-Indexed Tax
    • Taken by Storm
    • Temperature Data Quality
  • Ontario Energy Policy
  • CV, talks, bio, etc.
  • EPEQ
  • Op-Eds
  • Science and Public Policy
  • Global Financial Crisis
  • Other Research
  • Earth Hour
  • EPEQ
  • YourEnvironmentSources

Global Warming: Fingerprinting Methodology



ACADEMIC PAPERS

CHECKING FOR MODEL CONSISTENCY IN OPTIMAL FINGERPRINTING: A COMMENT

I have several papers underway assessing the statistical methods used by IPCC authors over the past 20 years for attributing climate changes to greenhouse gas emissions. The first is a critique of the seminal paper in the field, published in 1999 by Myles Allen and Simon Tett in Climate Dynamics. My comment has likewise been published in Climate Dynamics. 

  • McKitrick, Ross R (2021) Checking for Model Consistency in Optimal Fingerprinting: A Comment. Climate Dynamics August 2021 https://doi.org/10.1007/s00382-021-05913-7 
​
Although I requested Allen and Tett be asked to provide a response the journal did not solicit one. Once it was accepted and it was clear no response had been obtained I contacted Allen and Tett and offered to withhold my paper from publication until they had a chance to prepare a reply, an offer which they declined. As of the present (February 2023) they have not published a response other than the non-technical one published through the GWPF (see entry to the right).  

UPDATE  (May 2022): Chen et al. (Peking University) have released a pre-print commenting on my paper (available here). As far as I know it isn't published. I was asked to review it shortly after it appeared and my comments are here. 



UPWARD BIAS  IN OPTIMAL FINGERPRINTING FROM USE OF TLS

Continuing with my examination of the optimal fingerprinting methodology used in climatology to attribute climate change to GHG's I have an article in Climate Dynamics comparing OLS and TLS estimators.

  • McKitrick, Ross R (2022) On the choice of TLS versus OLS in climate signal detection regression. Climate Dynamics 10.1007/s00382-022-06315-z 

I have posted a blog at Climate Etc. offering a non- (or less-) technical explanation of the findings. Basically, TLS was proposed 20 years ago as a solution to attenuation bias in OLS, which can cause the coefficients in a fingerprinting regression to be underestimated. The OLS bias may thwart "detection" of the GHG signal and since fingerprint coefficients are used in carbon budget calculations it may overstate the "allowable" CO2 emissions associated with a warming target. The problem is that in many circumstances TLS over-corrects and imparts an upward bias, thus exaggerating the size of the forcing signal. I run a series of Monte Carlo experiments and show that TLS is not automatically preferred to OLS and can easily be more biased, but in the opposite direction. This propensity was known in the 1990s. Econometricians never use TLS as far as I know, we use IV regression to fix attenuation bias since unlike TLS it can be shown to be consistent.  

TOTAL LEAST SQUARES BIAS WHEN EXPLANATORY VARIABLES ARE CORRELATED

Continuing my exploration of the statistical elements of the IPCC climate attribution methodology I have a couple of papers under review at journals in which I use Monte Carlo simulations to analyse the properties of Total Least Squares (TLS, the preferred regression method) under conditions typical in a signal detection regression. There is very little underlying theory about when TLS yields consistent or unbiased results. In a single-variable model with a random explanatory variable TLS corrects a downward bias in Ordinary Least Squares (OLS, the standard regression method). But in many other cases it over-corrects or introduces new biases, and consistency results are not available without imposing unrealistic and untestable assumptions. In one paper I examine the consequences of omitted variables bias, and I will disseminate that paper separately. In this paper I look at what happens if the explanatory variables are allowed to be correlated (as they are in signal detection regressions). The results are, frankly, bizarre. I have posted a draft of the paper on the Earth and Space Science pre-print archive here:
​
  • McKitrick, Ross R (2022) Total Least Squares Bias when Explanatory Variables are Correlated. Earth and Space Science Open Archive. https://doi.org/10.1002/essoar.10510919.1 

I don't know what to make of the results and I would welcome comments. Unfortunately the table formatting in the archived version is wonky so here is a better version. The turn-key R code is in the Supplement provided, and is also here. 
​​
OTHER

NON- (OR LOW-) TECHNICAL DISCUSSIONS OF "CHECKING FOR MODEL CONSISTENCY IN OPTIMAL FINGERPRINTING: A COMMENT"
​
  • McKitrick, Ross (2021) Suboptimal Fingerprinting? A debate about climate statistics. London: Global Warming Policy Foundation. October 18, 2021.

 I have published through the Global Warming Policy Foundation a non-technical explanation of my paper. At the accompanying website the GWPF has reproduced comments Myles Allen provided to earlier media inquiries, to which I have added a reply, and Richard Tol has supplied a commentary on the exchange.

UPDATE: Here is a non-technical Backgrounder to try and make the material more accessible. 
​
I have published a (somewhat) non-technical summary at Judith Curry's blog (PDF here). Optimal Fingerprinting has long been the dominant tool in climatology for attributing climate changes to greenhouse gases. It is a matrix-weighted generalized least squares (GLS) regression model, and as such is based on tools familiar to economists, although changed in non-standard ways. In my article I show that those changes destroy the properties of consistency and unbiasedness associated with regular GLS methods. Unfortunately the problems have been concealed by exclusive reliance on a test statistic introduced by Allen and Tett that is meaningless for checking specification errors. As a result, none of the applications of this method over the past 20 years can be considered to have yielded reliable results.
  
 
Proudly powered by Weebly