Ned Foley and Charles Stewart have posted this draft on SSRN. Here is the abstract:

We conduct statistical analysis of a phenomenon identified by Foley (2013), the rise in the number of votes counted after Election Day (“overtime votes”) and the growing tendency of these votes to disproportionately favor Democrats in presidential elections (the “blue shift.”) We provide a historical description of these two time series, from 1948 to 2016, and establish that the timing of the persistent growth in the blue-shifted overtime vote began with the 2004 election. Changes in the time series are broadly consistent with changes in electoral practices, especially in recent years.

We perform statistical analysis to better understand variability across states in the overtime vote and the blue shift in the 2016 presidential election. We discover that variation in the size of the overtime vote is associated with the number of mail and provisional ballots, and with the partisanship of the state. The blue-shift is associated with state partisanship, but not with the number of mail and provisional ballots.

This analysis has relevance to post-Election-Day dynamics, especially in elections where a close election-night results that favor one candidate may yield to close final-canvass results that favor the other candidate. Among the concluding remarks, we discuss the conditions under which overtime votes counted in the 2020 presidential election could prove especially problematic.