Beruflich Dokumente
Kultur Dokumente
November 2009
At the individual level, inject a new technology (use of the right quantity of fertilizer, gizmo to help that) randomly to some farmers. Make gizmo available at low cost in the shop. See whether it diuse among friends of the injection points Prediction: no diusion in the usual networks. More diusion in village with farmers coop and discount program.
Environment
75 villages, rural Karnataka Within 2-3 hours outside of Bangalore Little pre-existing presence of micronance institutions Limited access to formal credit technologies Industries include agriculture (nger millet, rice, coconut), sericulture (silk worm rearing) BSS (Bharatha Swamukti Samsthe) is a young micronance institution (MFI) seeking to penetrate these villages Presents a unique opportunity to study diusion of micronance across the network
Data
Full census of the village:
Village questionnaire:
village leadership, presence of ngos, SHGs, geographical features (rivers, mountains, roads)
Household census:
demographic information, GPS coordinates, amenities (roof type, latrine), caste, occupation, etc.
Follow up survey:
Detailed individual questionnaire:
Stratied the households by religion (Hindu, Muslim, Christian) Stratied by geographic sub-locations Randomly sample subject to stratication half of the eligible households Basic household information Social network data on thirteen dimensions including visitors to the home, religious network, advisors, nancial network (who they borrow or lend small amounts of money or kerosene/rice from), etc.
Data
Sample Village
Sample Village
Data
Household versus individual network as unit of analysis What network to use? Usual identication issues (omitted variable bias etc.)
Descriptive Statistics
Table 1 : Descriptive statistics % general merit or obc % with savings % in self-help groups Average degree Average clustering Average path length Average eigenvector centrality (leaders) Average eigenvector centrality (population) Mean 0.676 0.414 0.357 15.691 0.530 1.976 0.072 0.050 Std. Dev. 0.167 0.136 0.165 2.615 0.046 0.132 0.015 0.008 Min 0.034 0.143 0.025 11.797 0.417 1.747 0.042 0.031 Max 0.991 0.643 0.635 22.462 0.657 2.249 0.123 0.067
Are the basic properties of the network correlated with take up?
Size Clustering Average Degree Variance of Degree distribution Degree of schism: p (1 p ) where p is fraction of network people who have negative 2d EV value.
Result
Table 2: Correlation between network characteristics and MF take up Take Up (Village level) Average degree -0.0041 -0.0043 [0.0057] [0.0058] Var. of degree distribution -0.0005 -0.0004 [0.0008] [0.0008] Average path length Average clustering Second eigenvalue of stochastized First eigenvalue of adjacency Schism Size Adj. R squared F stat p-value No. obs -0.0006 [0.0002] 0.2741 0.0030 38 -0.0005 [0.0002] 0.2552 0.0077 38 0.0922 [0.2467]
0.0236 [0.3144] 0.0386 [0.1749] -0.0036 [0.0245] 0.0009 [0.0012] -0.0005 [0.0002] 0.1968 0.0535 38
Questions of interest
Question 1: Does the fact that these leaders are leaders in the network sense matter for take up: We use eigenvector centrality of leaders as a measure of network leadership : a measure of importance which reects the importance of the people you know. Question 2: Does it make a dierence whether or not these people take up themselves (i.e., is is the centrality of the stained leaders that matters, or centrality in general?
Results
Table 4: Main regression Centrality Stain Stained Dummy*Centrality Size Village network control Village other controls (SHG, saving) Village centrality Adj. R squared F stat p-value No. obs -0.0007 [0.0003] N N Y 0.3862 0.0004 38 -0.0007 [0.0008] Y N Y 0.3611 0.0091 38 -0.0007 [0.0003] N Y Y 0.4435 0.0003 38 -0.0004 [0.0008] Y Y Y 0.4137 0.0053 38 1 1.1661 [0.76708] 0.3522 [0.1732] 2 1.4330 [0.8496] 0.2909 [0.1900] 3 1.8227 [0.79711] 0.2396 [0.1744] 4 1.8671 [0.8935] 0.2043 [0.1941] 5 2.2123 [0.9114] 0.5454 [0.3092] -0.6436 [0.4592] -0.0006 [0.0008] Y Y Y 0.4346 0.0050 38
Interpretation
Centrality of leaders matters a lot Whether or not leaders are stained also matters: but it is not really identied Centrality of leaders matter just as much, even if none of them take up: an information story rather than other peer eects stories.
Caveats
Leader centrality may still tell us something about the nature of the village which is correlated with subsequent take up of micronance
We regress village level variables on centrality Table level of savings signicant... We control for village level network and background characteristics in the regressions.
Table 3 : Is the centrality of the leader correlated with village characteristics? Eigenvector centrality of leader set Age -0.0029 -0.0031 [0.0021] [0.0020] Education 0.0013 0.0000 [0.0032] [0.0032] % GM or OBC -0.0158 -0.0216 [0.0152] [0.0153] No. beds -0.0224 [0.0300] Electricity -0.0077 [0.0247] Latrine 0.0086 [0.0112] Savings -0.0030 [0.0306] SHG -0.0265 [0.0253] Adj. R squared 0.0327 0.0705 -0.0432 F stat p-value 0.2546 0.1991 0.6924 No. obs 38 38 38
-0.0016 [0.0450] 0.0018 [0.0284] -0.0237 [0.0128] -0.0391 [0.0027] 0.0056 [0.0042] -0.0002 [0.0169] -0.0333 [0.0461] -0.0059 [0.0349] 0.0067 0.4359 38
Conclusions
Basic network properties such as degree, clustering, average path length, second eigenvalue all seem to be largely irrelevant
Next Steps
Does injection point within a network matter for adoption in this part of the network, particularly when the network has strong schism? Test node-level data hypotheses Test time series level hypotheses Study group formation patterns