probability model for comparison. When wealth is instrumented, its coefficient becomes negative
but is still insignificant. The Hausman test fails to reject the joint null of weak exogeneity and no
measurement error in wealth (p-value is .24), and according to the Sargan overidentification test
our instruments are valid (p-value of .36). This seems to suggest that the concerns about reverse
causation and measurement error are not warranted empirically. Notice that the “Chair’s language”
variable remains significant in all cases and its coe∏cient is even bigger in the 2SLS regression.
[Insert table 5]
Apart from including group fixed effects, the analysis so far has not attempted to relate ability to
borrow to any specific group characteristic. Table 5 addresses this point by adding to the individual
controls listed in column 2 of table 4 (not displayed) a number of group level controls. The first
two are measures of resources available in the group: group profits per capita (which may serve as
capital to advance loans) and average wealth of group members (in case members draw upon their
own personal capital to lend to others). Neither variable has a significant impact on access to credit.
The remaining variables are indexes of heterogeneity within the group.
Fragmentation indexes measure the likelihood that two randomly drawn member belong to dif-
ferent ‘categories’ and are computed as follows:
Fragmentj = 1 — ɪɪ s2kj (2)
k
where j represents a group, k a category, and each term skj is the share of category k in the total
membership of group j . For ethnic fragmentation, for example, the categories are Kikuyu, Luhya,
Luo, Kalenjin, Kamba, Kisii, Meru, Mijikenda, Masai, Somali, Nubian and Other. In this case, skj
is the fraction of group members speaking each of the above languages. The other fragmentation
indexes are computed in a way analogous to (2) but instead of using ethnicity we use the fraction
of individuals in a group with the same education level (No formal education, Class 1-8, Form 1-4,
Technical or College), and residential location.10 All indexes are constructed in a way that increasing
values correspond to more heterogeneity.
10The different ‘categories’ in the residential fragmentation index were the sub-areas where members resided within
each slum. For example, Korogocho was divided into Grogon, Gitadhuru, Highridge, Nyayo, Kisumu Ndogo, Ngomougo,
and so on for the other settlements.
13