When rays from the camera are refracted to different media, the radiance
they carry is scaled depending on the relative indices of refraction. It's
worthwhile to apply Russian roulette without including this scaling, since
it saves us from terminating rays that are actually about to refract out of
a medium and have the effect of radiance scaling cancelled out.

Test scene that shows the difference:

Film "image" "integer xresolution" 400 "integer yresolution" 400

Sampler "halton" "integer pixelsamples" 1
Integrator "path" "float rrthreshold" .8 "integer maxdepth" 7

LookAt 0 0 -3   0 0 0   0 1 0
Camera "perspective"

WorldBegin

LightSource "spot" "rgb I" [1000 1000 1000] "point from" [0 15 0] "point
to" [0 0 5]

AttributeBegin
Material "uber" "rgb Kd" [ 0 0 0] "rgb Ks" [ 0 0 0 ] "float index" 1
  "rgb Kt" [0 0 0 ] "rgb opacity" [.00001 .00001 .00001]
Shape "trianglemesh" "point P" [-9 -9 -2.5   9 -9 -2.5    0 9 -2.5 ]
  "integer indices" [0 1 2]
Shape "trianglemesh" "point P" [-9 -9 -2.25   9 -9 -2.25    0 9 -2.25 ]
  "integer indices" [0 1 2]
Shape "trianglemesh" "point P" [-9 -9 -2   9 -9 -2    0 9 -2 ]
  "integer indices" [0 1 2]
Shape "trianglemesh" "point P" [-9 -9 -1.5   9 -9 -1.5    0 9 -1.5 ]
  "integer indices" [0 1 2]
AttributeEnd

AttributeBegin
Material "glass"
Shape "sphere"
AttributeEnd

Material "matte" "rgb Kd" [.4 .4 .8 ]
Shape "trianglemesh" "point P" [-40 -40 5   40 -40 5    0 40 5 ]
  "integer indices" [0 1 2]

WorldEnd