Hi Scott, First, I'm not sure how your message got tagged as Spam by Argonne's SpamAssassin (words or phrases selling drugs??? huh???), but that happens well before I can do anything about it. I've forwarded this message on as a false positive for spam. For better archiving, I also included the full message you sent below.
Good point. It is much more straightforward to check the uncertainty than the correlation to determine if that could be responsible for the anomalous S02...so I should have said that Yu-Chuan should check the uncertainty of the S02 parameter. If it is very large (say 0.8), then Ifeffit is in effect reporting that it cannot determine S02 well for some reason. At that point, I think it is helpful to see if there is a high correlation, since that may give a clue as to the reason for the large uncertainty.
Knowing the correlations always helps. And stating a best-fit value without an uncertainty is dangerous (one might take an implicit uncertainty, which may not be what you mean). This is probably a bit off-topic from the original question (where I'd guess self-absorption to be the main issue,....), but let's consider typical fit results: S02= 0.9 +/- 0.1, sigma2= 0.015 +/- 0.005, and a correlation between S02 and sigma2 C_S02_sigma2= +0.90. One could conclude from these that a true value of S02= 1.0 was reasonable. But in order to get to S02= 1.0, sigma2 has to go up to ~= 0.019 (~= sigma2_best + C_S02_sigma2 * delta_sigma2). S02= 0.80 is also reasonable, but this implies sigma2 would drop to around 0.011. The correlation means that, although having either S02= 0.80 OR sigma2= 0.020 would be reasonable,t having both S02= 0.80 AND sigma2= 0.020 is much less likely. The correlation by itself says nothing about the likelihood of having a true value for S02 of, say, 0.5. There is a chance this can happen, but it's small because the uncertainty in S02 is 0.1. The correlation simply tells you how sigma2 would respond if S02 were 0.5, but nothing more.
So this brings up a (possibly contentious) point. In their reporting recommendations, the IXS suggests reporting high correlations, particularly when they are between parameters that do not routinely show high correlations. What is the reason for this suggestion? I'm not criticizing it...just looking for the rationale.
It's hard to speak for the IXS, but I'd say that correlations are recommended to be reported because they're important statistics. The correlations, along with the best-fit values and uncertainties, help more fully describe the range of plausible results. It's generally well-known that the parameters (S02,sigma2) and (E0,R) are highly correlated for a single shell of a single data set, and the implications of these are generally understood, I think. Sometimes other variables are correlated, and some people even do complex fits with multiple data sets or generalized variables that are not the simple XAFS parameters ;). In these cases, it may not be obvious how the variables are correlated. I think the IXS committee was concerned about this, and so recommended reporting correlations in such cases. That seems sensible to me. --Matt
On Sat, 28 Feb 2004, Scott Calvin wrote:
Matt,
Good point. It is much more straightforward to check the uncertainty than the correlation to determine if that could be responsible for the anomalous S02...so I should have said that Yu-Chuan should check the uncertainty of the S02 parameter. If it is very large (say 0.8), then Ifeffit is in effect reporting that it cannot determine S02 well for some reason. At that point, I think it is helpful to see if there is a high correlation, since that may give a clue as to the reason for the large uncertainty.
So this brings up a (possibly contentious) point. In their reporting recommendations, the IXS suggests reporting high correlations, particularly when they are between parameters that do not routinely show high correlations. What is the reason for this suggestion? I'm not criticizing it...just looking for th erationale.
--Scott Calvin Sarah Lawrence College