r57 Subject index Summary f, 87,93,LL8,l2+r25, i, 114123,132,L37-L44 -87,93,97,99-100, This thesis investigatesthe robustnessof estimation methods in covariancestructure analysis (CSA). By means of CSA structural relations between hypothetical constructs can be 42-50, 65-66,69.70, studied. When covariancestructure analysis is applied, in theory a number of assumptions must hold. The central question is: how robust are the estimators of model parameters, the estimators of the standard errors, and the goodness-of-fitstatistic for the model, when il128,134,137-144 one or more of the underlying assumptionsa^reviolated? 86,106 In Chapter 1 the covariancestructure model and a number of estimation methods are described,such as the Maximum Likelihood (ML), the GeneralizedLeast Squa,res(GLS), the Elliptical ReweightedLeast Squares(ERLS), and the Asymptotically Distribution Flee 0, 14&1rg 5 G 5 1 , 5 8 , 6180, 7 - 1 1 3 , i, 111 5 L2ïL28, L44 (ADF) estimation method. This chapter explains how estimates for the model parameters, estimates for the standard errors, and the chi-squaretest statistic for the model fit can be obtained. Thereafter, five robustnessquestions are discussed,namely robustness against: 1) the use of small samples; 2) violation of the postulated distribution of the observedvariables; 3) the analysis of a correlation matrix instead of a covariancematrix; 4) model misspecification; 5) non-linear relations between latent va"riables. The robustness of an estimator against violations of underlying assumptions can be determined empirically by means of a Monte Carlo study. Robustnessstudies with respect to CSA a.re published regularly. In Chapter 2 the results of 25 robustness studies are reviewed. These studies have investigated the effect of small sample size or distributional violations on the behavior of an estimator. They are described systematically and compa.redby means of characteristicsrelated to the population model, the simulated data, the estimation methods, the replications,or the researchsummary statistics. The conclusions of robustness studies frequently seem to contradict. An important cause for contradicting conclusionsis differencesin assessmentcriteria of robustness studies. Becausethese criteria should be no causefor differencesin conclusions,general assessment criteria are defined. They are applied to the presented results in robustness studies whenever possible. The conclusionsof the author(s) are therefore not taken for granted. By means of a meta.analysis the findings of robustnessstudies are summaxized. The characteristics serve as explanatory variables for an analysis concerning the bias of estimators of the model pa,rameters,estimators of the standard errors, and the chi-square statistic for the model fit. In general, population models in robustnessstudies are rather small, that is, they have -- 158 Summaty few degreesof freedom compared with models in applied research. On the basis of the results of the meta.analysis a Monte Carlo simulation study is conducted to fill such gaps in the robustnessresearchand to refine the conclusionsof the meta-analysis. In Chapter 3 the design of this simulation study is discussed,which consists of five population models, eleven continuous distributions of the observedva.riables,four sample sizes (200, 400, 800 and 1600),and four estimationmethods (ML, GLS, ERLS, and ADF). For eachcell of the Monte Carlo design 300 replications a,reanalyzed. A replication is included in the analysis if it provides a convergent and proper solution for each of the four estimation methods. There a.reseveral corrections for estimators of the standard errors and the chi-square Corrections of maximum likelihood estimators investigated in the simulation the study a,re ScaledML X'statistic (Satorra & Bentler, 1994) and the Robust ML esti- statistic. mator of the standa"rderrors (Browne, 1984;Bentler & Dijkstra, 1985). The behavior of the Yuan-Bentler correctedADF / statistic is also studied (Yuan & Bentler, 1994). To simplify the assessmentof the performance an estimator, criteria a,redefined that dichotomize this performance a.sacceptableor unacceptable. Given that the performance of an estimator generally becomes better as the sample size increases,it is essential to know for which sample size this performance becomesacceptable. In a number of ca.ses the behavior of an estimator is unacceptableirrespective of the sample size. The results for one of the five population models are discussedextensively in Chapter 4. The behavior of an estimator is studied per para,uretertype, that is, for the factor loadings, the path coefficients, the factor correlations, and the variances of the measurement errors, because of possible differencesbetween para,metertypes. In general, the overall bias of the GLS and ADF parameter estimators is mainly due to underestimation of the va,riancesof the measurementerrors. When the observedvariables have positive kurtosis, the overall bias of the ML, GLS, a.ndERLS estimator of the standard errors is mainly due to underestimation of the standard errors of the estimated varia,ncesof the measurement errors. In Chapter 4 a number of regressionmodels that predict aspectsof the behavior of an estimator a^represented. The predictors are related to the skewnessand the kurtosis of the observed variables, and the sample size. The regressioncoefficients are estimated by means of the simulated data. In Chapter 5 the influence of the population model on the behavior of an estimator is examined. The examined model effects are related to the size of the factor loadings, the number of factors, the number of indicators per factor, and the existence of structural relations between factors. The design ofthe Monte Carlo study is chosenso that the effect of a specific model cha.racteristiccan be investigated by comparing the simulation results for two different models. Chapter 5 also investigates whether general conclusions across all models ca,nbe formulated. wmary Summaly s of the ah gaps rapter3 models, {00,800 rll of the analysis bhods. ri-square nulation ML estinvior of )4). nedthat brmance entialto of cases An important finding applicable to each population model is that the ML estimator of the model para,rreters is almost unbiased when the sa,mplesize is at least 200. In the case of a small sa,urplethe GLS para,ureterestimator has a much larger bias than the ML pa,ra,ureterestimator when the model has at least twelve observed va"riables. The bias npter 4. ;or loadurement l overall n of the curtosis, nly due rement lr of an tmis of Éedby imator rdiags, rctural : effect reaults acro&9 159 of the ADF para,meterestimator increaseswhen the kurtosis increa^ses.With a positive kurtoeis the bias of the ADF parameter estimator is larger than that of the GLS parameter estimator, irrespective of sa,mplesize. The ML and GLS estimators of the standard errors a,rebiasedwhen the averagekurtosis of the observed rariables deviates from zero. The standard errors a,reunderestimated in the case of a positive average kurtosis and overestimated in the case of negative average kurtosis. The ADF estimation method gives a large underestimation of the standard errors when the sa,rrple size is small relative to the number of observed variables in the model. The Robust ML standaxd error estimator has a smaller bias than the other standard error estimators wheu the average kurtosis is at least 2.0 and the sa,mplesize is at least 400. The chi-square statistic is on averagesmaller when the factor loadings in the population are relatively small, that is, when the observed variables are unreliable measures of the underlying factors. The ML f statistic is not robust against a large skewness,the model is then rejected too often. The Scaled ML X2 statistic grves a better description of the model fit in cases of large non-normality. The ADF f statistic tends to reject models with many degreesof freedom too often, especially when the sample size is small. The Yuan-Bentler corrected ADF I statistic is more robust against small sa,mplesizes than the ADF f statistic. The behavior of these two goodness-of-fit statistics is insensitive to the distribution of the observedva^riableswhen the sa,urplesize is la,rgerelative to the number of degreesof freedom. An iurportant goal of both the meta-analysis and the Monte Carlo study is to find guidelines for applied resea,rchersrega,rding choice of estimation method. These guidelines depend on specific properties ofthe postulated model and the sa;npledata. In general, they indicate how la,rge the sa,rrple size must be to obtain almost unbiased para,rreter estimates, almost unbiased estimates of the standard errors, or an acceptable rejection rate of the chi-square statistic at lhe 5Tolevel when the model is correctly specified. These guidelines are spelled out in Chapter 6. In addition to the required sa.mplesize, the expected sign of the bias of an estimator is supplied when the sarnple size is too small. Several tables and formulas are provided with which the required sa;nple size and the sign of possible bias can be determined. The final chapter also discussesa number ofunder-exposedsubjects, such as the effect ofcategorizing observed ra,riables and model misspecification on the behavior of an estimator.