7,649
edits
Changes
Jump to navigation
Jump to search
Urban Start-up Agglomeration and Venture Capital Investment (view source)
Revision as of 12:05, 25 October 2019
, 12:05, 25 October 2019→Analysis Exploration
Within cities, and with year fixed effects, the only things that matter are the number of singleton/multitons and the number of hulls. The more singletons/multitons, the greater the growth; and the more hulls, the lower the growth. This result goes away if we control for the full set of geographic characteristics. Without city fixed effects, only avgdisthm is significant (and negative).
===Maximum R2===
The Maximum R-squared analysis relies on finding layers by hull count. Specifically, for a certain hull count, say 2, we find the lowest-highest layer occurrence. That is we first find the last layer (i.e., the highest layer) that had a hull count of 2, then if that occurrence is a part of a sequence, we find the first time that in the sequence.
Possible examples include Portland, OR (in say 2018), or Burlington, VT in 2015
SELECT * FROM HullsBase WHERE place='Burlington' AND statecode='VT' AND year=2015;
place statecode year layer numclusters
Burlington VT 2015 1 1
Burlington VT 2015 2 1
Burlington VT 2015 3 1
Burlington VT 2015 4 2
Burlington VT 2015 5 2
Burlington VT 2015 6 1
For Burlington, the lowest-highest layers are 4 for hull count 2 and 6 for hull count 1. Note that there is presumably a layer 7 with hull count 0.
We want to regress, within each city, our measures on performance, and then select the lowest-highest layer -- in effect the number of hulls -- which maximizes the R-squared of the regression. We are going to use 1995 to 2018 inclusive, which is 24 years. We want to include scale variables (i.e. 3 vars: growthinv17l numdealsl numstartupsl), and the (almost full -- we can't do avgdisthm) set of explanatory vars (8 vars nosinglemulti nopair nohull totsinglemulticount totpaircount tothullcount tothullarea totpairlength), which would total 11 vars plus an intercept. To reduce this, we can do a PCA first.
The alternative method is to create residuals from the outcome variable (growth vc) after taking out the variation explained by the scale effect. We can then use these residuals as the outcome variable in the city-by-city regressions (again, perhaps with a PCA). The advantage to this method is that we will pick layers based on their R2 of just the agglomeration effect, rather than agglomeration and scale together.
reg growthinv17lf growthinv17l numdealsl numstartupsl i.year i.placeid if lowesthighestflag==1 & year>=1995 & year <.
predict growthinv17lfres, residuals
Using the first method, the PCA yields 3 or maybe 4 components. We'll use 4.
===Instrumental Variable(s)===
