Ra. Johnson et al., GROWTH-CURVES OF DEVIANT-BEHAVIOR IN EARLY ADOLESCENCE - A MULTILEVELANALYSIS, Journal of quantitative criminology, 13(4), 1997, pp. 429-467
Multilevel growth curve models provide a means of analyzing individual
differences in the growth of deviance, allow a number of theories to
be integrated in a single model, and can help to unify research on dev
iant/delinquent/criminal careers at different stages of the life cycle
. Building on the distinction between ''population heterogeneity'' and
''state dependence'' as alternative explanations of persistent indivi
dual differences in deviance (Heckman, 1981; Nagin and Paternoster, 19
91), we show that models with two levels can be used to represent and
analyze a variety of criminological theories. The first level (level 1
) uses repeated measurements on individuals to estimate individual-lev
el growth curves. The second level treats the level 1 growth curve par
ameters (e.g., slope, intercept) as outcome variables and uses time-in
variant factors to explain variation in these parameters across indivi
duals. We illustrate this approach by estimating a model of growth in
deviance drawn from Gottfredson and Hirschi's deviant propensity theor
y. An innovative feature is the assumption that adolescents' expected
growth curves of deviance follow a classical Pearl-Verhulst logistic g
rowth model (Pearl, 1930). The results suggest that five risk factors-
parental psychiatric problems, lack of parental support, living arrang
ements with zero or one parent in residence, low family income, and ma
le gender-have strongly positive effects on deviant propensity. For ex
ample, adolescents with no supportive parents, and no other risk facto
rs, have expected asymptotic levels of deviance (peak levels attained
at about age 18) that are about twice as high as those of adolescents
with no risk factors. Yet more than two-thirds of the individual-level
variability in growth curves is unexplained by the five risk factors.
This unobserved heterogeneity would remain hidden in analyses using c
onventional structural equations models and the same explanatory varia
bles.