plot_GrowthCurve() ... large error bars
Closed this issue · 2 comments
Request by Dirk Mittelstraß via email to RLumSk
Background
Data sets with bad OSL curve SNR lead to LxTx values with large error bars. Nonetheless, analysing such data sets leads often to reasonable De values but also to heavily overestimated De error bars (for example: De = 123.45 s; De.Error = 12345.67 s).
Problem
Probably the following happens: For De error estimation via Monte Carlo, the LxTx values are randomly varied according to their individual error bar and assuming normal distribution. In a significant portion of Monte Carlo runs the random LxTx value sets lead to diverging De values. These diverging events propagate into the standard deviation calculation and lead to a diverging De error bar.
Solution
Set the randomisation bandwidth of the LxTx values not according to the normal distribution but to, for example, 1/100 of the standard deviation. The resulting confidence interval of De intervals can be transformed to the De error by multiplication with 100. Diverging events should become unlikely while good-SNR data sets should get about the same De error as previously.
Reminder: Simulation needed.
As discussed via email with Dirk and Rex Galbraith, the current implemention is sensible;
alternatively a maximum likelihood fitting with the De as unknown parmeter could be implemented.
However, this is nothing on the agenda for the package development at this very moment.