The Good, The Bad and The Ugly

Statistics for an Excellent Dataset

Statistics for a Pretty Good Dataset

Statistics for Pretty Bad Dataset

---

Methods to Calculate Data Quality Statistics

R linear = SUM ( ABS(I - I(mean)) ) / SUM (I)

R square = SUM ( (I - I(mean)) ** 2 ) / SUM (I ** 2)

Chi**2 = SUM ( (I - I(mean)) ** 2 ) / (Error ** 2 * N / (N-1) ) )

R linear is the same as Rmerge when the data is reduced in P1.
Otherwise R linear is the same as Rsym.
For reflections with only a single measurement (N=1), these statistics are not computed.

According to the HKL2000 Manual: Chi**2 is the weighted ratio of the difference between the observed and average value of I and I(mean), squared, divided by the square of the error model times a factor correcting for the correlation between I and I(mean). It depends on an explicit declaration of the expected error in the measurement. So the user of the program is part of the Bayesian reasoning process behind the error estimation.

Loren's comment: But the equation given does not indicate that there is a weight. And introducing user bias into your error analysis is nothing to brag about.

Starting on page 89 of the HKL2000 is a discussion of these statistics.

HKL2000 Manual (pdf file)

After looking at the statistics you may be left wondering what "nonunf" means. Who knows? It is not mentioned in the HKL2000 Manual. If you do know what it means, please tell Loren.