This course combines the theory of individual and average causal effects in the sense of J. Neyman and D. B. Rubin with analysis techniques of structural equation modelling. All designs and models for the analysis are developed for the purpose to learn about individual, conditional and/or average causal effects. Unlike other courses on the analysis of treatment effects, it uses structural equation modelling (with or without latent variables) instead of analysis of variance techniques, the General Linear Model or related techniques. As will be shown, this will enable us to learn not only about average and conditional effects, but, in specific models, also about individual causal effects.

This course is a synthesis of different traditions in methodology: Rubin's approach to causality, the Campbellian tradition of quasi-experimentation and internal validity, and structural equation modeling, especially latent state-trait modeling, latent change modeling and latent growth curve modeling.

Although this workshop does not require experience and knowledge in structural equation modelling (SEM), we do not recommend this workshop as a first introduction to SEM, if the motivation is to have an introduction into SEM. For this purpose we rather suggest our workshop "Erste Einführung in die Analyse von Strukturgleichungsmodellen mit LISREL" held in German in November 2005. Similar introduction in English is: "Analysis of Structural Equation Models with Lisrel 8.54" held in March 2004. (These workshops are still available in the internet and on DVDs).

The present workshop, "Introduction to the Analysis of Causal Effects with EffectLite, LISREL and Mplus" aims at those interested in data analysis in experimental and quasi-experimental studies involving covariates such as one or several pretests, a discrete treatment variable, and one or several outcome variables. EffectLite analyzes a generalized multivariate analysis of variance and covariance. It creates LISREL or Mplus input files, reads and interprets the results, computes some statistics, and produces an output file containing the most important results. EffectLite does not assume homogeneity of variances (in the univariate case with a single outcome variable) or covariance matrices (in the multivariate case with two or more outcome variables) of the outcome variables between groups. Furthermore, it allows analyzing mean differences between groups with respect to (a) several manifest outcome variables, (b) one or more latent outcome variables, and (c) a mixture of the two kinds of outcome variables. The results include estimates of conditional and average effects with respect to several manifest covariates or with respect to one or more latent covariates, and a mixture of the two kinds of covariates. The covariate(s) may also be qualitative. In this case we estimate and test average effects for non-orthogonal analysis of variance designs, provided that the covariates are specified as qualitative indicator variables. If the covariates fulfil certain assumptions, the program estimates and tests the average causal effects.

In the workshop we will:

present the theory of individual and average causal effects,

show how to use EffectLite, LISREL and Mplus for this class of models, and

show how to use LISREL and Mplus for individual causal effect models.

Um die Videos anzuschauen, benötigen Sie den , den Sie hier herunterladen können. Die Videos werden bereitgestellt über die Digitale Bibliothek Thüringen (dbt). Die grün verlinkten Videos und Materialien sind kostenfrei abrufbar. Klicken Sie auf den grünen Link, um das Video zu sehen oder die Datei herunterzuladen! Um alle Materialien und Videos abrufen zu können, müssen Sie sich einloggen.

Individual and Average Causal Effects (Theory)

Motivation: The Simpson paradox

Basic concepts: Individual and average causal effects (ICEs and ACEs)

Fixed and random effects regression

Generalization to more than two treatment conditions

Identifying the average causal effect via mean differences (Theory)

The prima facie effect

Two kinds of biases

The role of randomization

Heterogeneity of variances between treatment groups

Estimating and testing average causal effects via structural equation modeling (Applications using LISREL and Mplus)

One outcome variable, two or more treatment groups

Several outcome variables, two or more treatment groups

One latent outcome variable, two or more treatment groups

Several latent outcomes

Conditional Causal Effects (Theory)

Conditional (average) causal effects

Conditional mean differences

Identifying conditional and average causal effect via conditional mean differences (Theory)

Sufficient conditions for unbiasedness of conditional mean differences

Understanding the sufficient conditions

Estimating and testing conditional and average causal effects via structural equation modeling (Applications using LISREL and Mplus)

Three kinds of covariates: manifest continuous, manifest nominal, latent continuous

Analysis of covariance, regression with interactions, nonorthogonal analysis of variance

One outcome variable, two or more treatment groups, one covariate

Several outcome variables, two or more treatment groups, one covariate

One latent outcome variable, two or more treatment groups, one covariate

Several latent outcomes and several latent covariates

Analysis of Individual and Average Causal Effects (Theory and Applications)

A Single Group, two Pre-test and two Post-test Occasions

Two Groups, two Pre-test and two Post-test Occasions