HLM多层线性模型讲义1.pptVIP

CSSS 560 Hierarchical Linear Modeling for the Social SciencesCSSS 560/vanduijn/560/ Marijtje van Duijn vanduijn@ Padelford B-5-L; 221-6874 University of Groningen, the Netherlands m.a.j.van.duijn@rug.nl HLM = Special type of regression analysis Simple regression equation y: dependent variable x: explanatory variable e: error i: individual ?: regression coefficient (intercept; slope) Why special: structure of the data Due to sampling Independence (random) sampling replaced by multi-stage (hierarchical) sampling Due to research questions Interested in multiple ‘levels’ of inference Micro/macro distinction Not mutually exclusive What is then so special? Simple two-level regression model What has changed? Structure of the data What are the implications of this change? Model assumptions need to be changed as well How does it work? - example y: 4th grade math score x: IQ measure i indicates 4th grade pupil - level 1 unit of analysis j indicates school (or class) - level 2 unit of analysis ? regression coefficient Structure of the data School 1 2 3 Pupil 1 2 (3 4 5) 1 2 (3) 1 2 Level-2 unit-specific regression equation etc. (one for each level-2 unit) HLM combines these equations by further specification of regression coefficients: z school quality special Regression/Analysis of Variance? ANOVA = Regression ANOVA handles data with ‘group structure’ (cf. nested data) by specifying a ‘group variable’ (factor) fixed effect by specifying a random variable random/mixed effect Motivation: exchangeability of units; random sampling Usually considered a ‘nuisance’, but not in HLM! Summarizing: Hierarchical Linear Models consider Hierarchical (nested) data Separate regressions for each (level 1) unit Regression equations for the regression parameters Variance components, independent between levels, dependent or independent at same level This makes the observations within one unit dependent they always share a vari