452 V. Maksimovic and G. Phillips
Shin and Stulz’s results suggest that conglomerates may invest less efficiently than
single-segment firms, and that, while firm’s internal financial markets are integrated,
the integration is partial so that the markets are not allocatively efficient. These studies,
based on COMPUSTAT data, stand in marked contrast to the findings ofMP (2002)
using LRD data, who find that conglomerate investment is, on the whole, efficient.
More recent work has tried to reconcile the findings of these papers. As is often the
case in research on conglomerates, the issues center on the thorny issue of measurement
of the within firm quantities, in this case investment and Tobin’sq.
A key variable which is difficult to measure at the conglomerate-segment level is
Tobin’sq. As discussed above, the COMPUSTAT based literature attempts to proxy
Tobin’sqfor a segment by using observedqs of “comparable” firms.Whited (2001)
directly tests whether the findings of the COMPUSTAT based literature can be attributed
to measurement error caused by the use of segments’qs based on estimated derived
from “comparable” single-segment firms.
Whited’s arguments can be illustrated with equation(1). As noted above, we can-
not observeqdirectly, but must use a proxy, perhaps based on the average Tobin’sqs
of single segment firms operating in segmentj’s industry.Whited (2001)models the
consequences of the use of a noisy proxy on the estimates of coefficients ofβin equa-
tion(1)andβ,δandφin equation(2)above. Suppose that the relation between the
proxy,pand the Tobin’sqtakes the following form:
pj=α+qj+εj.
We can eliminatezfrom this system by regressing all the variables onzand using the
residuals. For simplicity we can also initially fold the variables CFjand CF−jwith the
other exogenous variables intoz. Doing so we obtain
̃ij= ̃qjβ+ζ ̃j, (6)
p ̃j= ̃qj+ ̃εj. (7)
These equations can be used to generate a set of eight moments such asE( ̃
i^2 j)
=β^2 E(
q ̃j^2)
+E
(
ζ ̃j^2)
,E( ̃ijp ̃j)=βE(
q ̃^2 j)
,
E
(
p ̃^2 j)
=E
(
q ̃j^2)
+E
(
ε ̃^2 j)
,
E
(
̃ijp ̃j^2)
=βE(
q ̃j^3)
,E
(
̃ij^2 p ̃j)
=β^2 E(
q ̃j^3)
, etc.
The estimation technique consists of replacing the eight left-hand side moments with
their sample estimates and then using GMM to find a vector of six right-hand side un-
observable quantities(β, E(q ̃j^2 ), E(εj^2 ), E(ζj^2 ), E(q ̃j^3 ), E(q ̃j^4 )). This vector is one that
comes closest to minimizing the distance between the left-hand and right-hand sides of
equations, when evaluated using the minimum variance GMM weighting matrix derived
byErickson and Whited (2000).
The estimate of sensitivity of investmentβis obtained fromβ=E(i ̃^2 jp ̃j)/E( ̃ijp ̃^2 j).
Given the estimate ofβ, the remaining moment conditions can then be solved to give the