We need to decide the best way to implement a global chi-square filter that will compare model SANS data to experimental contrast variation data. Currently, chi-square filter is only used for one contrast at a time and the chi-square value calculated is only for one contrast at a time (although results for multiple contrasts can be saved in the same folder). I have performed global chi-square filtering "by hand" in the following ways:
1) calculate the average chi-square from the values at each contrast.
2) find the subset of structures that has the lowest chi-squared ("bottom of the well") at each contrast and then search for structures that are present in all of these subsets.
Both methods result in an ensemble of structures that best fit the data at all contrasts. However, the two methods don't give the same ensemble, although there is some overlap. The average chi-square method is perhaps a little easier to automate since it can be done in one step. The two-step approach works better with user intervention, i.e., to select the chi-square range that defines the "bottom of the well", which likely is different at each contrast. This can probably be automated, but is that really what we want to do?
Should we offer both options?
We need to decide the best way to implement a global chi-square filter that will compare model SANS data to experimental contrast variation data. Currently, chi-square filter is only used for one contrast at a time and the chi-square value calculated is only for one contrast at a time (although results for multiple contrasts can be saved in the same folder). I have performed global chi-square filtering "by hand" in the following ways: 1) calculate the average chi-square from the values at each contrast. 2) find the subset of structures that has the lowest chi-squared ("bottom of the well") at each contrast and then search for structures that are present in all of these subsets. Both methods result in an ensemble of structures that best fit the data at all contrasts. However, the two methods don't give the same ensemble, although there is some overlap. The average chi-square method is perhaps a little easier to automate since it can be done in one step. The two-step approach works better with user intervention, i.e., to select the chi-square range that defines the "bottom of the well", which likely is different at each contrast. This can probably be automated, but is that really what we want to do? Should we offer both options?