Introduction to Multilevel Modeling Using HLM 6

By ATS Statistical Consulting Group

Multilevel data structure

Students nested within schools Children nested within families Respondents nested within interviewers Repeated measures nested within individuals longitudinal data, growth curve modeling

In the example of student nested within schools: Level-1 variables, such as students gender and age Level-2 variables, such as school type and size

How would we analyze such multilevel data?

OLS regression OLS regression with robust standard error Aggregation Disaggregation Ecological fallacy interpreting analyses on aggregated data at the individual level

Ecological Fallacy

See figure 3.1, on page 14 from Multilevel Analysis by Snijders and Bosker

Hierarchical linear model Random Intercept model Yij = 0j + rij 0j = 00 + u0j Written in mixed model format: Yij = 00 + u0j + rij
i is for individuals and j is for schools 0j is the mean of Yij for school j 00 is the average of all the 0js, therefore the grand rij and u0j are normally distributed rij and u0j are independent of each other Parameters to be estimated include regression coefficients and variance components: 00, var(rij) and var(u0j)

Hierarchical linear model Random Intercept and random slope model Yij = 0j + 1jX + rij 0j = 00 + u0j 1j = 10 + u1j Written in mixed model format: Yij = 00 + 10X + u0j + u1jX+ rij
0j is the mean of Yij for school j when X is zero 1j is the slope of X for school j (or the effect of X for school j) rij, u0j and u1j are normally distributed u0j and u1j are assumed to be correlated cross-level error terms are assumed to be independent parameters: 00, 10, var(u0j), var(u1j), cov(u0j, u1j) and var(rij)

Hierarchical linear model Random Intercept and random slope model Level-2 variable(s) to predict intercept and/or slope Yij = 0j + 1jX + rij 0j = 00 + 01W + u0j 1j = 10 + 11W + u1j Written in mixed model format: Yij = 00 + 01W + 10X + 11W*X + u0j + u1j*X + rij
0j is the mean of Yij for school j when X is zero 1j is the slope of X for school j (or the effect of X for school j) 00 is the average intercept 11 is the coefficient for the cross-level interaction term rij, u0j and u1j are normally distributed u0j and u1j are assumed to be correlated Cross-level error terms are assumed to be independent parameters to be estimated: 00, 01, 10, 11, var(u0j), var(u1j), cov(u0j, u1j) and var(rij)

Comparing the assumptions for hierarchical linear models with OLS models OLS Assumptions Linearity: function form is linear Normality: residuals are normally distributed Homoscedasticity: residual variance is constant Independence: observations are independent of each other HLM assumptions Linearity: function forms are linear at each level Normality: level-1 residuals are normally distributed and level-2 random effects us have a multivariate normal distribution Homoscedasticity: level-1 residual variance is constant Independence: level-1 residuals and level-2 residuals are uncorrelated Independence: observations at highest level are independent of each other

Estimation Methods: REML vs. ML

Reading: Section 4.6 Parameter Estimation from Snijder and Bosker REML and ML produce similar regression coefficients REML and ML differ in terms of estimating the variance components If the number of level-2 units is small , then ML variance estimates will be smaller than REML, leading to artificially short confidence interval and biased significant tests. REML is the default estimation method for HLM Likelihood ratio test for nested models When fixed effects are the same, model has fewer random effects , then both REML or ML may be used When one model has fewer fixed effects and possibly fewer random effects, then ML may be used

Issues with Centering

Reading: Section 5.2 The effects of centering from Kreft and De Leeuw In OLS centering is to change the interpretation of the intercept Centering in HLM is not a simple issue Grand-mean centering The raw score model and the grand mean centered model are equivalent linear models. Group-mean centering Most of the times, the group mean centered model and the raw score model are neither equivalent in the fixed part nor in the random part. Combining substantive and statistical reasons in choosing
raw score group-centering with reintroducing the means group-centering without reintroducing the means

An Example
The dataset is a subsample from the 1982 High School and Beyond Survey and is used extensively in Hierarchical Linear Models by Raudenbush and Bryk. It consists of 7185 students nested in 160 schools. The outcome variable of interest is the student-level math achievement score, mathach. Predictor variables
Level-1 (student level) predictor variables:
ses: social-economic-status of a student female 0 = male and 1 = female

Level-2 (school level) predictor variables:

meanses: mean ses at school level, aggregated from student level schtype: type of school: 0 = public and 1 = private, there are 90 public and 70 private schools size: size of a school

Model Building
Reading: Section 6.4 Model specification from Snijder and Bosker Unconditional model: mathachij = 0j + rij 0j = 00 + u0j Random intercept model with level-2 predictor(s): mathachij = 0j + rij 0j = 00 + 01(meanses) + u0j Random intercept and random slope model: mathachij = 0j + 1j(ses) + rij 0j = 00 + u0j 1j = 10 + u1j Full model: mathachij = 0j + 1j(group_mean_centered_ses) + rij 0j = 00 + 01(schtype) + 02(meanses) + u0j 1j = 10 + 11(schtype) + 12(meanses) + u1j

Model 1: Unconditional Means Model

mathachij = 0j + rij 0j = 00 + u0j 00 = 12.636972 var(rij) = 39.14831 var(u0j) = 8.61431 Rho = var(u0j)/(var(u0j) + var(rij)) = 8.61431/(8.61431+39.14831) = .18035673

Final model
mathachij = 0j + 1j(group_mean_centered_ses) + rij 0j = 00 + 01(schtype) + 02(meanses) + u0j 1j = 10 + 11(schtype) + 12(meanses) + u1j
Tau INTRCPT1,B0 SES,B1 2.37996 0.19058 0.19058 0.14892

Final estimation of fixed effects (with robust standard errors) ---------------------------------------------------------------------------Standard Approx. Fixed Effect Coefficient Error T-ratio d.f. P-value ---------------------------------------------------------------------------For INTRCPT1, B0 INTRCPT2, G00 12.096006 0.173699 69.638 157 0.000 SCHTYPE, G01 1.226384 0.308484 3.976 157 0.000 MEANSES, G02 5.333056 0.334600 15.939 157 0.000 For SES slope, B1 INTRCPT2, G10 2.937981 0.147620 19.902 157 0.000 SCHTYPE, G11 -1.640954 0.237401 -6.912 157 0.000 MEANSES, G12 1.034427 0.332785 3.108 157 0.003 ---------------------------------------------------------------------------Final estimation of variance components: ----------------------------------------------------------------------------Random Effect Standard Variance df Chi-square P-value Deviation Component ----------------------------------------------------------------------------INTRCPT1, U0 1.54271 2.37996 157 605.29503 0.000 SES slope, U1 0.38590 0.14892 157 162.30867 0.369 level-1, R 6.05831 36.70313
---------------------------------------------------------------------------------------------------------

Final Model (continued)

mathachij = 0j + 1j(group_mean_centered_ses) + rij 0j = 00 + 01(schtype) + 02(meanses) + u0j 1j = 10 + 11(schtype) + 12(meanses) + u1j 00 = 12.096: the intercept for public schools with meanses =0 (average ses) 01 = 1.226: the change in intercept from a public school to a private school
the intercept for private school with meanses = 0 is 12.096+1.226 = 13.322 the intercept for public school with meanses = 1 is 12.096 + 5.333 = 17.429 the effect of gcses for public schools with meanses = 0 is 2.94 the effect of gcses for private schools with meanses = 0 is 2.94 1.641 = 1.299 the effect of gcses for public schools with meanses = 0 is 2.94 the effect of gcses for public schools with meanses = 1 is 2.94 + 1.034 = 3.974

02 = 5.333: the change in intercept for a one-unit change in meanses 10 = 2.94: the slope of gcses for public schools with meanses = 0. 11 = -1.641: the change in slope from a public school to a private school 12 = 1.034: the change in slope for a one-unit change in meanses

For

For

INTRCPT1, B0 INTRCPT2, G00 SCHTYPE, G01 MEANSES, G02 SES slope, B1 INTRCPT2, G10 SCHTYPE, G11 MEANSES, G12

12.096006 1.226384 5.333056 2.937981 -1.640954 1.034427

0.173699 0.308484 0.334600 0.147620 0.237401 0.332785

69.638 3.976 15.939 19.902 -6.912 3.108

157 157 157 157 157 157

0.000 0.000 0.000 0.000 0.000 0.003

Whats new in HLM 6

The following paragraph is based on: http://www.ssicentral.com/hlm/new.html HLM 6 greatly broadens the range of hierarchical models that can be estimated. It also offers greater convenience of use than previous versions. Here is a quick overview of key new features and options:
All new graphical displays of data. Greater expanded graphics for fitted models. Model equations displayed in hierarchical or mixed-model format with or without subscripts easy to save for publication. Distribution assumptions and link functions are presented in detail. Slightly different and easier way for specifying random effects. Cross-classified random effects models for linear models and non-linear link functions with convenient Windows interface. High-order Laplace approximation with EM algorithm for stable convergence and accurate estimation in two-level hierarchical generalized linear models (HGLM). Multinomial and ordinal models for three-level data. Also see the types of models. New flexible and accurate sample design weighting for two- and three-level HLMs and HGLMs. Easier automated input from a wide variety of software packages, including the current versions of SAS, SPSS, and STATA. Residual files can be saved directly as SPSS (*.sav) or STATA (*.dta) files. Analyses are based on MDM files, replacing the older less flexible SSM format.

Getting ready for using HLM software for multilevel data analysis Creating MDM file
separate level-1 and level2 files for HLM2, or a single file original file can be in different format, such as SPSS, Stata and SAS linking variable can be either numeric or character variables in the analyses have to be numeric mdm file: binary file used for analyses and graphics mdmt file: template file in text format for creating mdm file hlm2mdm.sts: text file containing the summary statistics

Data management
HLM does not have data management capability One has to use other stat package(s) to clean the data and to create variables, such as dummy variables and within-level interaction terms HLM handles cross-level interactions nicely

Choosing preferences and other settings

Demo on using HLM

Input Data and Creating the "MDM" file
from a single SPSS file box-plot scatter plot unconditional means model regression with means-as-outcomes random-coefficient model intercepts and slopes-as-outcomes model Multivariate hypothesis tests on fixed effects Multivariate Tests of variance-covariance components specification Model-based graphs Modeling Heterogeneity of Level-1 Variances Models Without a Level-1 Intercept Constraints on Fixed Effects

data-based graphs

Model Building

Hypothesis Testing, Model Fit

Other Issues

References Multilevel Analysis: An Introduction to Basic and Advanced Multilevel Modeling by Tom Snijders and Roel Bosker Introduction to Multilevel Modeling by Ita Kreft and Jan de Leeuw Multilevel Analysis: Techniques and Applications by Joop Hox Hierarchical Linear Models, Second Edition by Stephen Raudenbush and Anthony Bryk HLM 6 - Hierarchical Linear and Nonlinear Modeling by Raudenbush et al.