7,649
edits
Changes
Jump to navigation
Jump to search
Estimating Unobserved Complementarities between Entrepreneurs and Venture Capitalists (view source)
Revision as of 13:27, 23 April 2021
, 13:27, 23 April 2021no edit summary
|Has paper status=In development
}}
==Current Work==
Xunkie Zheng wrote "Notes on Estimation of Venture Model", which is in the project folder.
Note on that doc:
*Startup-Specific Variables:
**VE adds executives round after round, so firms with more rounds will have more executives. We should only use Founders and whether or not they are serial from the executive measures.
**We should be using patents here?
**This appears ok in section 4 but not in section 1
**It somewhat repeats in section 5.1
**PC Total Investor suffers from similar issues
*The dataset might not be informational...
**Interaction terms related to VC Exp could have over 70% of zeros
*Same city indicator? Also, log distance. Transformation is awkward.
*We may want to redefine PC Rank so that the higher its value the higher the rank of the city
Discuss normalization:
Distance is standardized to have zero mean and unit variance with respect to the whole
dataset. PC Rank is normalized to be between 0 and 1 with respect to the whole dataset.
Other variables are normalized to be between 0 and 1 with respect to each matching market.
See Table 1 for summary statistics for realized matches before data transformation. See
Table 2 for summary statistics for realized matches after data transformation.
Discuss the structural estimation results in Table 3:
Two coefficient estimates in Model 6 have unexpected sign
Discuss:
In general, the average market complementarity in Table 8, the average observed market
production in Table 9, and the average observed market complementarity in Table 10 should
be the highest for the original market, then after removing the worst VC, then after removing
the mediocre VC, and then after removing the best VC. If the latter three counterfactual
values are similar, it is due to the small magnitude of the coefficient compared to other
coefficients in the model. The counterfactual values in Models 3 and 5 are higher than
the original one in Table 8, it is because there is a dominant negative coefficient in the
model. Also, the large standard deviation of unobserved complementarity makes unobserved
complementarity and so is the whole complementarity of some matches to be large and
negative. After one firm is removed, the number of total matches in the optimal assignment
decreases by 1 so that the match with large and negative unobserved complementarity is
dropped. However, since the observed production or observed complementarity of the same
match may not be negative, such case only appears in Table 8.
To be checked:
*VC Exp is “matchprevindu20” in the codebook, which is defined as “Matched VCs prior experience investing in companies (i.e., count of companies) in the same pccode20 as the PortCo”. There might be a mistake in constructing this variable. The VC Exp for counterfactual matches are greater than or equal to the value for realized matches.
Tables:
1
2
3 Structural Estimation Results
Model 3
Model 6 has two unexpected signs
4 Change in production with respect to one standard deviation of the (interaction) variable
5 Change in production with respect to one standard deviation of one variable in the interaction term while setting the other variable at 0.5
6
7 Correlation between (two) firm-specific variables in the interaction term in realized matches
8 average market complementarity
9 average observed market production
10 average observed market complementarity
The estimation strategy is two-step simulated method of moments. The solver is "adaptive_de_rand_1 bin_radius_limited" in the
Julia package "BlackBoxOptim.jl", which executes Differential Evolution algorithm.
Julia code:
*https://github.com/robertfeldt/BlackBoxOptim.jl (support parallel CPU)
==Previous Work==
