The stand-out finding is reliability:
True random designs show larger effect sizes that favor PBL over traditional learning too.
The consistent trend seems to be that we are hamstringing the PBL literature base with weak research designs and little attention to measurement. When we pay attention to those things, and presumably reduce measurement error and a priori group differences--PBL shows improved student outcomes. Almost double what we find as a norm.
Figure design shamelessly stolen from Brett Shelton.