 |
The ICC is
important to indicate to what extent the clustering of data, if not taken into account,
will violate the assumption of independent observations within groups and result in
inflated standard errors of coefficients, and a greater likelihood of seeing nonexistent
relationships (Type I errors). Even as small an intraclass correlation as 0.01, with about
50 observations within each area, will inflate the Type I error from the posited a =.05 to about a =.11 (Kreft and DeLeeuw,
1998). Secondly, a model was constructed which explained within-area variance, followed
by a model explaining between-area variance (Bryk and Raudenbush, 1992). Four inequality
variables were specified as main effects for the final model, in which cross-level
interactions with significant micro variables were fitted (Snijders and Bosker, 1999).
Continuous variables were centered around the grand mean to minimize cross-level
correlations, all other variables were dichotomized. |
