sfOctober 10.

   I couldn't get good per capita regressions. The instruments are not
very good fit ,as far as R2, though the t-stats are big.

  I did get good results with levels regressions, even getting
plausible elasticities.

 We trim off the top 5% of districts. That is necessary-- they are
just too big, and would dominate regressions.





BELOW ARE OLD NOTES ON PER CAPITA REGRESSIONS:


It would make more sense to use Budget per Crime, maybe. THere,
getting a good instrument is tough, though.

 I cna't get any good instrument for Budget per Population or budget
per crime.

We have to trim the sample. The distribution of value is just too
skewed.

A logged IV regression needs 2 stages of regression, and bootstrapping
of the standard errors.

 We badly need that key.

Maybe numbe rof assistant prosecutors is better for us than budget.

STAFF and TOTALAREA make no sense as explanatory variables. Well,
STAFF is endogenous.

Heckman procedure: TWOSTEP in ivprobit does not allow us to get
predicted probability of being positive.


My Heckman procecure comes out with things significantiwth the wrong
sign!


Twostep does NOT allow marginal effects elasticities to be computed.

The WINRATE elasticities are amazingly tiny, even for significant
variables. The PROSRATE elasiticies are too big.

This is a big mess.
 

Localpay and percapita99 are fine as instruments for BUDGET; just not
for budpop.

so try