========================================================================
SELECTION-MARGIN TEST: how well do individual observables predict ...
========================================================================
Outcome                             sample N    LPM R²   pseudo-R²
------------------------------------------------------------------------
(a) Internet adoption (full sample)      6948     0.230       0.263
(b) Online learning (users only)        5783     0.117       0.109
(b2) Online shopping (users)            5783     0.117       0.090
(b3) Online gaming (users)              5783     0.153       0.138
------------------------------------------------------------------------
Access-margin observables explain 2.0x as much variation (LPM R²) as the within-user learning choice.
  -> Selection on observables is concentrated at the ACCESS margin,
     supporting the claim that within-user use-type choice is less
     contaminated by the ability/income sorting that drives adoption.
========================================================================
