The background-subtracted light curve (using a background produced by pcabackest) has unexpected variability. Could this be a problem with the background estimator?
The least understood aspect of the background is the component from the activation due to passage through the SAA. For the first few kiloseconds after SAA, the activation is dropping rapidly, and the rate of the drop-off is not modeled with enough precision. This can cause variations in the light curve which are only due to the background; it may also cause an net overestimation in the spectrum. So if you have a long enough observation, you would benefit from excluding the first kilosecond or so after SAA. (This can often be done by simply using maketime on the background lightcurve and giving some maximum value for the count rate.) Extracting all layers and all channels will rarely produce a satisfactory result for the light curve. This is due to the difference in the signal to noise ratio between layer 1 and layers 2 and 3, and also to the fact that at the higher energies, the background dominates. For both of these reasons, any uncertainties in the model will be augmented and you ar
