It was suggested that there could be some geometric dependence to the data-simulation mismatch.  Two possible sources of geometric dependence were proposed: eta dependence in the tower gains, and phi dependence in the BSMD strip gains.  I examined both of these possibilities.

1. Eta dependence

I split the dataset into four bins in |eta| (0<|eta|<.25, and so on: both photons were required to fall into the same bin) and performed the usual fitting procedure.  As you can see, the same structures observed in the overall fit are still visible.  Figure 1 shows the fits and their components, and figure 2 data-fit/data.  Each row is an eta bin, and each column a pT bin.

Figure 1:

Figure 2:

2. Phi dependence

The original suggestion was to exclude the BSMDP strips at the two highest and lowest values for each module.  This would require reprocessing everything, however, so instead I approximated it by determining, for each pion candidate, the BSMDP strip pointed to by both its photons, and excluding the candidate if that strip was at one of the two highest or lowest phi values.  Since this might not get rid of all the photons that included one or more of the strips to be excluded in their BSMDP cluster, I also did the fitting using a second, tighter cut that eliminated any pion candidate with a photon that pointed to a BSMDP strip at one of the four highest or lowest phi values.  Since clusters can have a maximum of five strips, two on each side of the peak, this tighter cut should ensure that none of the strips at the phi edge of the module are used.  Figures 3-4 show the result for the first cut, and figures 5-6 for the second.  In both cases, the same structures as are seen in the full data fit are visible.

Figure 3:

Figure 4:

Figure 5:

Figure 6: