Many measures based on egocentric network data, such as age compositio
n or (local) network density, can be viewed as 'aggregate' measures: t
hey are mean values of the alter attributes or the dyadic attributes t
hat fall within a given respondent's egocentric network. Internal cons
istency methods of classical test theory are not suitable for assessin
g the reliability of such measures: such methods presume a 'crossed' d
esign for data collection in which each respondent is scored on the sa
me set of indicators. In designs for gathering egocentric network data
, alters are instead 'nested' within respondents; moreover the number
of alters may differ across respondents. This paper evaluates the reli
ability of composition and density measures via analysis-of-variance a
pproaches to reliability known as generalizability theory. Reliability
estimates are presented for egocentric measures based on the 1985, 19
87, and 1988 General Social Surveys and for the 1977-1978 Northern Cal
ifornia Community Study. Ethnoreligious composition, political composi
tion, density, and composition of a network by 'friends' or co-members
of organizations are measured with relatively high reliability, even
for a relatively small number of alters. Other measures require more a
lters to attain adequate reliability, and some, such as sex compositio
n, remain problematic even when the number of alters grows quite large
. The sensitivity of reliability estimates to differences in instrumen
t design is examined using design variations in the surveys studied.