stratify and stratify_s methods exclude units with propensity score equal to zero

[diego-mazon]

This happens regardless of trimming.

causal.reset()
print causal.summary_stats
print "There are 453 control units and 547 treatment units \n"

causal.est_propensity_s()
score = causal.propensity['fitted']
print "minimum propensity score = ", score.min()
print "number of units with null propensity score = ", len(score[score == 0])
print "\n"

print "stratify using implemented algorithm"
causal.stratify_s()
print "blocks", causal.blocks
print(causal.strata)
print "Why 7 controls only?!"
print "\n"

print "stratify using a single segment [0, 1]"
causal.stratify()
print "blocks", causal.blocks
print(causal.strata)
print "Why 7 controls only?!"
print "\n"


print "Replace null scores with 1e-100"
score[score == 0] = 1e-100
print "minimum propensity score = ", score.min()
causal.stratify_s()
print "blocks", causal.blocks
print(causal.strata)
print "453 controls as should be"

Summary Statistics

                   Controls (N_c=453)         Treated (N_t=547)             
   Variable         Mean         S.d.         Mean         S.d.     Raw-diff

---

          Y        0.031        1.389        6.097        1.287        6.066

                   Controls (N_c=453)         Treated (N_t=547)             
   Variable         Mean         S.d.         Mean         S.d.     Nor-diff

---

         X0       -0.021        0.976        7.004        0.980        7.184
         X1       -0.059        1.038       -0.054        0.993        0.004
         X2        0.506        0.501        0.497        0.500       -0.017
         X3        0.552        9.884        0.592       10.317        0.004
         X4        1.119        1.019        0.956        1.023       -0.160

There are 453 control units and 547 treatment units

minimum propensity score = 0.0
number of units with null propensity score = 446

stratify using implemented algorithm
blocks [0, 1]

Stratification Summary

          Propensity Score         Sample Size     Ave. Propensity   Outcome

Stratum Min. Max. Controls Treated Controls Treated Raw-diff

     1     0.000     1.000         7       547     0.000     1.000     2.984

Why 7 controls only?!

stratify using a single segment [0, 1]
blocks [0, 1]

Stratification Summary

          Propensity Score         Sample Size     Ave. Propensity   Outcome

Stratum Min. Max. Controls Treated Controls Treated Raw-diff

     1     0.000     1.000         7       547     0.000     1.000     2.984

Why 7 controls only?!

Replace null scores with 1e-100
minimum propensity score = 1e-100
blocks [0, 1]

Stratification Summary

          Propensity Score         Sample Size     Ave. Propensity   Outcome

Stratum Min. Max. Controls Treated Controls Treated Raw-diff

     1     0.000     1.000       453       547     0.000     1.000     6.066

453 controls as should be

[diego-mazon]

I would fix it by modifying subset function in causal.py

def subset(p_low, p_high):
            # deal with first non-single-point segment separately 
            # to include vanishing pscore units within it
            if p_low == 0 and p_high != 0:  
                 return (p_low <= pscore) & (pscore <= p_high)
            else:
                return (p_low < pscore) & (pscore <= p_high)

I think that pscore=0 units should be included in this segment rather than in (0, 0] or [0, 0] segments.