Evaluating Small Sample Approaches for Model Test Statistics in Structural Equation Modeling

Size: px
Start display at page:

Download "Evaluating Small Sample Approaches for Model Test Statistics in Structural Equation Modeling"

Transcription

1 Multivariate Behavioral Research, 9 (), Copyright 004, Lawrence Erlbaum Associates, Inc. Evaluating Small Sample Approaches for Model Test Statistics in Structural Equation Modeling Jonathan Nevitt Department of Agriculture and Applied Economics Virginia Polytechnic Institute and State University Gregory R. Hancock Department of Measurement, Statistics and Evaluation University of Maryland, College Park Through Monte Carlo simulation, small sample methods for evaluating overall data-model fit in structural equation modeling were explored. Type I error behavior and power were examined using maximum likelihood (ML), Satorra-Bentler scaled and adjusted (SB; Satorra & Bentler, 1988, 1994), residual-based (Browne, 1984), and asymptotically distribution free (ADF; Browne, 198, 1984) test statistics. To accommodate small sample sizes the ML and SB statistics were adjusted using a k-factor correction (Bartlett, 1950); the residual-based and ADF statistics were corrected using modified and F statistics (Yuan & Bentler, 1998, 1999). Design characteristics include model type and complexity, ratio of sample size to number of estimated parameters, and distributional form. The k-factor-corrected SB scaled test statistic was especially stable at small sample sizes with both normal and nonnormal data. Methodologists are encouraged to investigate its behavior under a wider variety of models and distributional forms. Structural equation modeling (SEM) has become a versatile and widely used data analytic method for evaluating causal and predictive hypotheses in the behavioral sciences. Historically it has been a large sample technique, with minimum sample size guidelines ranging from five to ten cases per estimated model parameter (e.g., Bentler & Chou, 1987), depending upon the method of estimation employed. Unfortunately, practitioners are often unable to obtain sufficient numbers of cases to meet such minimum guidelines, let alone satisfy an estimation method s distributional assumptions. Thus, there has been increased demand for methods that perform optimally at smaller sample sizes and under varied distributional conditions. The methodological community has responded accordingly, with several major SEM software Correspondence concerning this article should be directed to the first author at Department of Agricultural and Applied Economics, Virginia Tech, Blacksburg, VA , or to jnevitt@vt.edu. MULTIVARIATE BEHAVIORAL RESEARCH 49

2 programs incorporating techniques that could be more viable under suboptimal conditions. While practitioners might be eager to employ such robust methods, their inclusion in software packages is often intended initially to facilitate methodologists scrutiny. The purpose of the current investigation is to provide just such scrutiny within the context of a factorially designed Monte Carlo investigation. Background For a system of p measured variables [yielding p* = p(p + 1)/ unique variances and covariances], let X i = (x i1,..., x ip ) for i = 1,..., n be a sample from X = (x 1,..., x p ), with sample mean vector X, sample covariance matrix S, population mean vector, and population covariance matrix 0. Then, a covariance structure model represents the elements of 0 as functions of q free model parameters in vector, with null hypothesis H 0 : 0 = ( ). An hypothesized model may be fit to a p p sample covariance matrix, and for any vector of model parameter estimates ( û ) the hypothesized model can be used to evaluate the model-implied covariance matrix, ( û ) = Ŝ. The goal in parameter estimation is to obtain a vector of parameter estimates such that a function of the discrepancy between Ŝ and S is minimized. The maximum likelihood (ML) function is the most commonly employed discrepancy function and is defined as: -1 (1) ˆF ML = ln Ŝ ln S + tr(s Ŝ ) p, with associated test statistic () T ML = (n 1) ˆF ML, asymptotically distributed as a central with p* q degrees of freedom (df) under multivariate normality and when H 0 is true (see, e.g., Hayduk, 1987). With respect to issues of sample size and normality in SEM, empirical research has generated a large body of literature and an understanding of the behavior of ML estimation (and its associated test statistic) with realistic data forms (Bentler & Yuan, 1999; Boomsma, 198, 1985; Curran, West, & Finch, 1996; Fouladi, 1998, 1999, 000; Gerbing & Anderson, 1985; Hu & Bentler, 199, 1995; Hu, Bentler, & Kano, 199; Tanaka, 1984, 1987; Yuan & Bentler, 1997, 1998, 1999). Key findings indicate that at small sample sizes: (a) rates of non-convergence and/or improper solutions can be high; (b) parameter estimates exhibit only marginal bias; (c) parameter standard errors become attenuated; (d) T ML, and hence Type I error rates based on T ML, become 440 MULTIVARIATE BEHAVIORAL RESEARCH

3 inflated; and (e) nonnormality exacerbates inflation in Type I error rates based on T ML. Findings also commonly suggest that sample size adequacy is best measured by the ratio of subjects-to-estimated parameters (n: q) rather than by sample size in an absolute sense. To combat distortion in the model test statistic, two fundamentally divergent approaches have emerged and are examined in this investigation: abandoning ML for distribution-free estimation methods, and using ML for parameter estimation but then adjusting the test statistic to account for the effects of sample size and nonnormality. Each of these approaches is discussed in turn below. 1 Browne s (198, 1984) asymptotically distribution-free (ADF) method estimates model parameters by minimizing () ˆF ADF = (s ŝ ) 1 Ĝ (s ŝ ), where s = vech(s) (i.e., a p* 1 column vector of the unique elements of S), where ŝ = vech( Ŝ ), and where (4) =Cov{vech[(X )(X ) ]}, is a symmetric p* p* population fourth-order moment weight matrix. Browne (198) proposed estimating using only the sample data [requiring estimation of p*(p* + 1)/ unique elements in the weight matrix]. Let (5) Y i = vech[(x i X )(X i X ) ]. Then an estimator for is (6) Ĝ = S Y = Cov(Y i ), the sample covariance matrix of Y i. Under the null hypothesis, the associated test statistic (7) T ADF = (n 1) ˆF ADF, asymptotically follows a central distribution with p* q df. Monte Carlo experiments have demonstrated that at large sample sizes (e.g., n 5,000) T ADF yields Type I error rates at the nominal level (Chou, Bentler, & Satorra, 1 Bootstrap resampling is another approach for obtaining robust statistics in SEM. While bootstrapping appears to be a viable alternative under nonnormal data conditions, it is not directly intended to address the problem of small samples and thus was not pursued in this investigation. MULTIVARIATE BEHAVIORAL RESEARCH 441

4 1991; Curran et al., 1996; Hu et al, 199; Muthén & Kaplan, 199; Tanaka, 1984). However, with large models and small to moderate sample sizes ADF is problematic, leading to high rates of non-convergence and improper solutions, and to inflated Type I error rates associated with inflated T ADF values when models do converge. ADF s ability to yield a correct test statistic under nonnormal conditions at large sample sizes inspired Yuan and Bentler (1997, 1999) to develop corrections to T ADF for small sample sizes. They noted that in the regression literature cross products of model residuals are often used when estimating asymptotic covariances, and proposed an estimator for the weight matrix in ˆF ADF as n (8) Gˆ = [ 1/ ( n 1 )] ( ˆ)( ˆ i i ) Y s Y s. i = 1 Incorporating this estimator into ˆF ADF is tantamount to rescaling the original ADF test statistic such that (9) T ADF1 = [1/{1 + [T ADF /(n 1)]}]T ADF, which follows a central distribution with p* q df (when H 0 is true). Further motivated to improve small sample performance associated with T ADF, Yuan and Bentler (1999) proposed another modification to the ADF test statistic, appealing to Fisher s F distribution. They offered a transformation of T ADF based upon the logic of the transformation applied to Hotelling s T statistic in MANOVA (see Stevens, 1996, p. 155, for a review of the T statistic). Observing that T is a quadratic form, similar in structure to the ADF fit function, they proposed rescaling T ADF to an F-distributed statistic, (10) T ADF = [n (p* q)]/[(n 1)(p* q)]t ADF, with numerator and denominator df of p* q and n (p* q), respectively. Yuan and Bentler (1999) used Monte Carlo simulation to investigate the small sample performance of T ADF, T ADF1, and T ADF, systematically varying distributional forms and sample sizes. They reported that over the range of sample sizes investigated T ADF1 and T ADF maintained adequate control of Type I error rates as compared to T ADF, and yielded adequate power (for rejecting misspecified models) with both normal and nonnormal data. Inspired by these limited, yet promising, results, the adjusted ADF test statistics have been incorporated into recent releases of SEM software: T ADF1 is currently available in EQS 5.7 (Bentler, 1996), and the F-distributed statistic T ADF will be incorporated into EQS MULTIVARIATE BEHAVIORAL RESEARCH

5 Another strategy to control for nonnormality and potentially small samples is to estimate model parameters using ML and then assess datamodel fit using a test statistic that has been corrected. These corrections take two basic forms: (a) adjusting T ML for sample size and nonnormality, and (b) constructing distribution-free test statistics based on ML-estimated model parameters. Both approaches are reviewed as follows. The most well known corrections to T ML were developed by Satorra and Bentler (1988, 1994), with two modifications to T ML (generically referred to as SB statistics here) that make adjustments based on the degree of nonnormality in the sample data. Define ŝ& as the p* q matrix of partial derivatives of the p* elements in ŝ with respect to the q model parameters (i.e., the Jacobian matrix), evaluated at the final model parameter estimates. Let W be the symmetric p* p* matrix of unique fourth-order moments 1 1 obtained by ˆ ˆ S S, and let Uˆ = W ˆ ˆ ˆ ˆ Wss & & Ws& s& W, (11) ( ) 1 be the residual weight matrix of those inverted fourth-order moments. Then, (1) = ( * )/ ( ˆ ) TSB1 p q tr US Y TML. The asymptotic distribution of T SB1 is generally unknown; however, when H 0 is true its first moment matches a central distribution with p* q df. The T SB statistic adjusts the model df to yield (1) d tr( ˆ ) / tr( ˆ ˆ ) = USY USYUS Y, for the model df. Then as with T SB1, T ML is scaled as (14) / ( ˆ ) TSB = d tr US Y TML. Like T SB1, the asymptotic distribution of T SB is generally unknown; however, both its first and second moments match that of a central distribution with d df (when H 0 is true). Simulation research has indicated that both SB adjustments to T ML can be effective in controlling Type I error rates under some experimental conditions (Chou et al., 1991; Curran et al., 1996; Fouladi, 1998, 1999, 000; Hu et al., 199; Nevitt & Hancock, 001; Yuan & Bentler, 1998). The scaled T SB1 is currently available in EQS 5.7 (Bentler, 1996) and LISREL 8. (Jöreskog & Sörbom, 1996), while variants of T SB1 and T SB have been incorporated into Mplus 1.0 (Muthén & Muthén, 1998). MULTIVARIATE BEHAVIORAL RESEARCH 44

6 Fouladi (1998, 1999, 000) proposed another correction to T ML, combining a small-sample correction developed by Bartlett (1950) with the SB corrections for nonnormality. Within the context of exploratory factor analysis, Bartlett (1950) offered a k-factor correction to the ML test statistic (where k represents the number of latent factors in the model), suggesting that at small sample sizes (15) T = (n p/ k/ 11/6) ˆF ML, more closely follows a central distribution (with p* q df) than the usual T ML statistic. This adjusted statistic is equivalent to applying a multiplicative correction to T ML (or to any test statistic) of the form (16) c = 1 [(p + 4k + 5)/6(n 1)]. Fouladi (1998, 000) applied this correction factor to T ML to improve its small sample performance with normal data. She also investigated the k-factor correction applied to T SB1 and T SB with nonnormal data and reported that this scaling correction can be effective in controlling Type I error rates under some experimental conditions (Fouladi, 1999). Interestingly, for all three studies she only examined the measured variable path analysis case which has no latent factors (i.e., k = 0). While ADF estimation and T ML scaling adjustments appear to provide some protection against distortion in data-model fit statistics when modeling nonnormal data at small sample sizes, both approaches demonstrate either practical or theoretical limitations. For example, ADF estimation can be carried out only when sample sizes are above an absolute lower bound of n = p* (see, e.g., Bentler & Yuan, 1999). With respect to T ML scaling, the SB adjustments to T ML have been criticized for lacking known asymptotic results with nonnormal data; the k-factor correction as proposed by Fouladi (1998, 1999, 000) is mostly uninvestigated. To address the potential shortcomings of ADF estimation and T ML scaling adjustments, Yuan and Bentler (1998) turned to a mostly ignored residual-based test statistic offered by Browne (1984), noting that there exists a p* (p* q) matrix s& ˆ c whose columns are orthogonal to ŝ& (and to each other). Then the residual-based test statistic (asymptotically distributed as a central with p* q df when H 0 is true) proposed by Browne (1984) is (17) T B ˆ ˆ ˆ 1 ˆ Y = ( n 1) eˆ s& ( ) ˆ c s& cs s& c s& ce, 444 MULTIVARIATE BEHAVIORAL RESEARCH

7 where eˆ = s s ˆ is a p* 1 column vector of residual variances and covariances. Yuan and Bentler (1998) contended that T B could be applied to any consistent estimator for, even when data are not normally distributed, and argued that ML is a prudent choice. The T B test statistic is available to practitioners within LISREL 8. (Jöreskog & Sörbom, 1996). The advantage to T B is that it is an asymptotically distribution-free test statistic with theoretical elegance not found with T SB1, T SB, or the k-factor corrected test statistics in that it follows a known sampling distribution without assuming multivariate normality of the data. Moreover, Bentler and Yuan (1999) indicated that T B has smaller minimum sample size requirements than T ADF with a lower bound of n = df + 1. Unfortunately, these authors found that T B has a propensity to over-reject true models at n < 5000, thus showing Type I error performance remarkably similar to that of the unadjusted T ADF test statistic. The poor performance of T B at small samples led Yuan and Bentler (1998) to propose corrections to the test statistic that are analogous to the small sample corrections to adjust T ADF. First, they proposed replacing the fourth-order moment matrix in T B (i.e., S Y ) with the improved weight matrix given in Equation 8. Incorporating this estimator into T B is equivalent to rescaling it such that (18) T B1 = {1/[1 + (nt B )/(n 1) ]}T B, with T B1 following a central distribution with p* q df when H 0 is true. Yuan and Bentler (1998) also noted that the equation for T B takes on a quadratic form also reminiscent of Hotelling s T statistic. They proposed rescaling T B to an F-distributed statistic, (19) T B = {[n (p* q)]/[(n 1)(p* q)]}t B, with numerator and denominator df of p* q and n (p* q), respectively. The residual-based test statistics will be incorporated into EQS 6.0. Yuan and Bentler (1998) examined the residual-based statistics in a small Monte Carlo experiment, varying combinations of distributional form and sample size. T B showed a strong tendency to over-reject true models at all but the largest sample size with the F-distributed T B demonstrating the best overall performance. The strong performance of T B evidenced by Yuan and Bentler (1998) at the smallest sample size investigated (n = 150, n:q = 4.5:1) led Bentler and Yuan (1999) to conduct an experiment designed to push the test statistics to their sample size lower bounds. Some important findings from their study are: (a) consistent with previous investigations, T ML yielded MULTIVARIATE BEHAVIORAL RESEARCH 445

8 inflated Type I error rates with normally distributed data at reduced sample sizes (even at their largest sample size of n = 10 yielding n:q =.6:1); (b) T B and T B1 performed extremely poorly under every condition T B yielded 100% model rejections while T B1 yielded 0% model rejections; (c) T SB1 was not competitive under the study conditions, yielding inflated Type I error rates ranging from 10-7%; and, (d) T B performed quite well over the range of experimental conditions. Based on these findings, Bentler and Yuan recommended that practitioners turn to the T B F-distributed statistic regardless of distributional form when sample sizes are at least n df + 1. The Current Study While the studies reviewed above yielded promising results, there is clearly the need to characterize more fully the nature of small sample test statistics in SEM. First, no studies have investigated the k-factor correction with latent variable systems (for which k > 0), and thus it is unknown if the improvement afforded to T ML for normal data and the improvements to T SB1 and T SB with nonnormal data using the k-factor multiplier will hold under such models. If an improvement with latent variable models is realized using this correction, then determining a lower bound on sample size requirements for the effective use of the k-factor correction could yield improved recommendations to practitioners. Second, much remains to be learned regarding the behavior of the ADF and residual-based adjusted test statistics, T ADF1, T ADF, and T B1, T B. Bentler and Yuan (1999) noted that much of the characterization of these test statistics focuses only on a specialized model (a three-factor confirmatory factor model) under a very narrow range of conditions. They called for continued research, investigating these test statistics under a greater variety of conditions, including variations in the number of variables in the system and the type of model examined. Third, for sample sizes in which ADF estimation is possible, no research exists that directly compares the relative performance of T ADF1 and T ADF against their respective residual-based analogs, T B1 and T B, leaving practitioners with no clear choice among the ADF-based and residual-based ML test statistics. This current investigation seeks to address these important issues, contributing to our understanding of small sample test statistics in SEM. 446 MULTIVARIATE BEHAVIORAL RESEARCH

9 Method J. Nevitt and G. Hancock Test Statistics Examined This study examined test statistics for evaluating global fit in SEM that, based on previous investigations, have shown promise for practitioner use under commonly encountered data conditions. The following statistics were investigated: T SB1, T SB, T ML k, T SB1 k, T SB k (a k suffix denotes the k-factor correction was applied), T ADF1, T ADF, T B1, and T B. The uncorrected T ML, T ADF, and T B test statistics were also analyzed under all conditions (given sample size constraints), providing benchmarks against which to judge their adjusted forms. Population Models Two population models were developed from which simulated samples of data were drawn. Population Model A (Figure 1) is a five-factor latent variable path (LVP) model with three indicators per factor; Population Model Figure 1 Latent Variable Path (LVP) Population Model A MULTIVARIATE BEHAVIORAL RESEARCH 447

10 B (Figure ) is a seven-factor confirmatory factor analysis (CFA) model, also with three indicators per factor. We chose models that would replicate previous methodological findings as well as extend knowledge to different model types. Model B includes features that have not been examined in previous methodological investigations a substantially larger model requiring large numbers of parameters to be estimated, and a model with correlated error terms, two aspects that are commonly found in applied modeling scenarios. We established non-zero error covariances among all residuals within three groups: among all factors first indicators, among all factors second indicators, among all factors third indicators. This yielded 6 non-zero error covariances, each set to a population value of 0.1. Three additional non-zero error covariances, also set to 0.1, were included between residuals 1 and, between 10 and 11, and between 19 and 0. Thus, a total of 66 non-zero error covariances were established in Population Model B. This population model reflects applied scenarios such as when constructs have indicators that are parallel questionnaire items differing only in content, such as those assessing self concept in multiple subject matter domains (e.g., mathematics, science, and reading). Figure Confirmatory Factor Analysis (CFA) Population Model B 448 MULTIVARIATE BEHAVIORAL RESEARCH

11 Model Specifications for Type I Error Analyses J. Nevitt and G. Hancock When fitting simulated samples of data drawn from each population, model specifications (all properly specified) were implemented that examined both constrained and unconstrained models. Again, we incorporated design elements that would replicate previous simulation studies, and examine previously unanalyzed model specifications found in applied modeling scenarios. Complete details for model specifications are given in Appendix A (using LISREL notation). Specifications A1 and B1 were unconstrained versions of Models A and B, respectively, which required larger numbers of parameters to be estimated. Specifications A and B had loading and error variance constraints imposed on the measurement portion of Models A and B, respectively, and thus required fewer parameters to be estimated (see Appendix A). Such constraints are similar to those found in longitudinal models in which the same constructs are measured at multiple time points. When the equality forced by these constraints holds true in the populations, as is the case in the current simulations, then these constrained models are properly specified as well (see, e.g., Yuan & Bentler, 1998). Table 1 summarizes, for each model specification, p, q, and df, and sample size information (discussed in the next section). Table 1 Simulation Conditions for Properly Specified Models Sample Size Condition Model p p* q df 1:1 :1 5:1 10:1 df + 1 p* A b 50 b 86 a 10 b A b 10 a 10 b B a 58 b 645 b 190 b 10 a 1 b B b 197 a 1 b Note. Sample size conditions reflect subject-to-estimated parameter ratios, with the exception of the last two columns which are the sample size minima for the residual-based and ADF test statistics, respectively. a indicates sample size sufficiently large to compute residual-based test statistics. b indicates sample size sufficiently large to compute residual-based and ADF test statistics. MULTIVARIATE BEHAVIORAL RESEARCH 449

12 Sample Size Conditions For all four models, ratios of n:q = 1:1, :1, 5:1, and 10:1 were examined. In addition each model was also inspected under a sample size condition at which the residual-based and ADF test statistics are at their respective minima. Thus, each model was examined under a total of six sample size conditions. Test statistics based on T ML are not limited to any sample size constraints per se and were examined across all six conditions. The ADF and residual-based test statistics were only examined under sample size conditions that were sufficiently large to support these methods. Table 1 presents each combination of model and sample size condition and the associated absolute sample size for that cell in the design of the experiment. Data Generation and Modeling; Computation and Verification of Test Statistics Three multivariate distributions were examined, each established through the manipulation of univariate skew and kurtosis of the measured variables. All manifest variables were drawn from the same univariate distribution for each data condition. Distribution 1 is multivariate normal with univariate skew and kurtosis both set equal to 0. (Note that normality is defined here, as is commonly done, by using a shifted kurtosis value of 0 rather than a value of.) Distribution represents an elliptical distribution data are nonnormal but symmetric with univariate skew of 0 and kurtosis of 6.0. Distribution is nonnormal and asymmetric with univariate skew of.0 and kurtosis of 1.0. Simulated data matrices were generated in GAUSS (Aptech Systems, 1996) using the programming described by Nevitt and Hancock (1999) that follows the algorithm developed by Vale and Maurelli (198). All modeling of simulated data was performed using EQS 5.7b (Bentler, 1996). The T ML and T ADF test statistics were captured directly from EQS, as were the Jacobian matrix of partial derivatives and the residual covariance matrix. All other test statistics were calculated using GAUSS. The duplication matrix requisite for calculating the SB test statistics was constructed as described by Magnus and Neudecker (1986, pp ); the algorithm for obtaining the Jacobian complement requisite for the residualbased test statistics followed Gill, Murray, and Wright (1981, pp. 7-40, 16-16). The fourth-order moment weight matrix, S Y, requisite for both the SB and the residual-based test statistics, was constructed from the unbiased sample variances and covariances of the Y i data (i.e., an n 1 divisor was used when computing the variances and covariances). 450 MULTIVARIATE BEHAVIORAL RESEARCH

13 Because most of the test statistics in the present investigation were constructed outside of EQS, careful steps were taken to verify the accuracy of intermediate matrices and test statistics. The (biased) weight matrix S Y (before computing the unbiased S Y with n 1 as divisor), and the T SB1 and T ADF1 test statistics were verified using EQS 5.7b (Bentler, 1996). Accuracy of the residual-based test statistic T B was verified using LISREL 8. (Jöreskog & Sörbom, 1996); T ADF statistics were matched with output from a pre-release of EQS 6.0. Design and Execution For the investigation of Type I error rates, the study fully crossed four model specifications with six sample size conditions with three distributional forms, yielding 7 between cells. For each cell, independent data sets were drawn from the associated population model, distributional form, and sample size, and fit to the appropriate model specification using EQS. The same simulated data sets were not used to fit different model specifications (i.e., across Models A1 and A, or B1 and B) even though they might have been drawn from the same population, distributional form, and sample size. Such an approach was adopted to prevent any potential dependencies among the results across Models A1 and A, or across Models B1 and B. The number of iterations to convergence was set at 00 for ML estimation; EQS 5.7b (Bentler, 1996) limits the maximum number of iterations for ADF estimation to 0. Simulated data samples were subjected to parameter estimation using ML and ADF to calculate the test statistics. Start values were assigned using the true population parameter values. For each data set all test statistics were computed (given adequate n). If a given estimation method failed, or a test statistic could not be computed (i.e., a matrix was not invertible), then only those test statistics associated with that sample of data that failed were discarded and replaced, not the sample of data itself. Other test statistics that could be constructed from that sample of data were used in the replication count. The SB and residual-based test statistics were flagged as failing (and discarded) whenever their necessary product matrices could not be inverted. Additionally, the residual-based test statistics were flagged and discarded whenever T B exceeded a clearly inappropriate value of 900,000. For each cell in the design we obtained 000 successful replications; this number of replications yields high power for detecting a departure from Bradley s (1978) liberal criterion of robustness at the =.05 level of significance (Robey & Barcikowski, 199), as described in the next section. MULTIVARIATE BEHAVIORAL RESEARCH 451

14 Summarizing Results and Defining Test Statistic Robustness Study results were analyzed in terms of empirical Type I error rates across each cell s 000 replicates. As Yuan and Bentler (1998, 1999) noted, although adherence to the referenced sampling distribution is important, the primary concern for hypothesis testing is with respect to the tail-behavior of the test statistic (i.e., in terms of model rejection rates). As such, the percentage of model rejections across each cell s replications is reported for each study condition and is used to gauge test statistic robustness. All hypothesis testing was performed at the.05 level. Type I error robustness is evaluated using both an expected 95% confidence interval (CI) about =.05 and Bradley s liberal criterion (Bradley, 1978). Given the 000 replications per condition in this investigation and =.05, the 95% CI is [ (.05.95/000) 1/ ] * 100% = (4.04%, 5.96%); the robustness interval corresponding to Bradley s liberal criterion is (.5, 1.5 ) * 100% = (.5%, 7.5%). Power Analysis While the Type I error propensity of the test statistics investigated is the central focus in our study, the power of a test statistic to reject an improperly specified model also is considered. This investigation takes a two-step approach to evaluating test statistic utility. In the first step, Type I error propensity is evaluated. Using Bradley s (1978) liberal criterion, test statistics under study conditions that yielded empirical Type I error rates above the upper bound of 7.5% were eliminated from subsequent power analyses under those same conditions. Such screening is deemed reasonable because liberal Type I error rates are indicative of inflated test statistics, which in turn would be expected under nonnull conditions to precipitate power estimates that are artificially inflated and thus not comparable with power estimates from test statistics that do maintain reasonable Type I error control. Test statistics with model rejection rates within the robustness interval, or below the lower bound of.5%, were examined in a follow-up power analysis. Test statistics that yielded empirical Type I error rates below the lower bound of robustness were retained for power analysis, rather than being eliminated, because these test statistics (albeit at a seeming disadvantage) could potentially yield acceptable power with respect to rejecting misspecified models. The objective of the power analyses is to provide secondary information regarding the relative performance of the test statistics in this study, allowing test statistics to be compared against one another. These analyses are not meant to estimate the power of the test statistics investigated in any absolute sense. 45 MULTIVARIATE BEHAVIORAL RESEARCH

15 For each test statistic that yielded an observed Type I error rate that was within or below the robustness interval, a new series of 1000 replicates was generated using the same population model for that cell in the study, and at the same distributional form and sample size condition. These new data sets were then fit to an improperly specified model. Sample data drawn from the fivefactor LVP population model were fit to a three-factor LVP model; sample data drawn from the seven-factor CFA population model were fit to a fourfactor CFA model. Complete model (mis)specifications are given in Appendix A. These model misspecifications are reflective of situations in which models are under-parameterized, an aspect common in applied modeling scenarios. The relative degree of noncentrality across all misspecified models (under multivariate normality) was fairly consistent, as indicated by the population root mean-square error of approximation (ε) values (see, e.g., Browne & Cudeck, 199). Specifically, fitting misspecified models to the population covariance matrices yielded a nonzero F ML for each model, from which ε = (F ML /df) 1/ is determined: ε A1 =.144, ε A =.145, ε B1 =.117, and ε B =.1. Because the misspecified models did not perfectly match in model df against their companion properly specified models it was not possible to maintain the exact same absolute sample sizes and n:q ratios in the cells of the power analysis as compared to the Type I error analysis. Thus, the absolute sample sizes examined in the Type I error analysis were preserved in the power analysis (rather than the n:q ratios). Results Non-convergence, Improper Solutions, and Test Statistic Failure Rates Rates of non-convergence and improper solutions for the ML and ADF estimation methods, as well as failure rates for the SB and residual-based test statistics, were tracked for all cells in the Type I error portion of the study. Improper solution and test statistic failure rates are based only on those data sets that yielded converged solutions. Non-convergence, improper solutions, and test statistic failure rates are summarized here briefly; complete tables are given in Appendix B. Overall, for the properly specified models, non-convergence and improper solutions for ML and ADF were a function of distributional form and sample size, with increasing departure from multivariate normality and decreasing sample size systematically leading to increasing rates of non-convergence and improper solutions. While population ε values may be used to determine power to reject misspecified models (e.g., MacCallum, Browne, & Sugawara, 1996), such power determinations are not reported here as they only apply under multivariate normality and with normal theory estimators. MULTIVARIATE BEHAVIORAL RESEARCH 45

16 For ML estimation, non-convergence was rare with the highest rate at %. With respect to improper solutions, ML was problematic only at n:q 1:1, with rates as high as 6%. At all larger n:q ratios, improper solution rates for ML estimation were usually not more than 0%. Unlike ML, ADF estimation frequently yielded high rates of model non-convergence, most particularly at the sample size lower bound n = p* where non-convergence rates ranged from 0-40%. At n > p* ADF yielded non-convergence rates around 1-15%. ADF was also more problematic than ML estimation, with rates of improper solutions that ranged from 1-87% at n = p*, were as high as % at n:q = 5:1, and were up to 1% for some cells at n:q = 10:1. With respect to test statistics, the SB test statistics never failed, delivering 0% failure rates across the board (i.e., the triple-product matrix requisite for computing the SB-based test statistics was successfully inverted every time). The T B test statistic also never failed with respect to inversion of its necessary product matrix; however, there were replicates at the lower bound n = df + 1 sample size in which computed values of T B exceeded a cut-off value of 900,000 and were discarded and replaced. Such failure rates ranged from 7-44% and appeared to be somewhat model-dependent, with unconstrained models showing minimal failure (7-9%) while constrained models tended to yield aberrant values for T B more frequently (10-44%). Type I Error Rates Results for the unconstrained LVP model (A1) are given in Table. With normal data at small sample size conditions T ML delivered inflated model rejection rates that monotonically decreased with increasing sample size, as expected. Using Bradley s (1978) liberal criterion, T ML was robust only at n:q = 10:1. The Bartlett k-factor correction to T ML showed improvement with normal data with Type I error rates within the robustness interval at sample size conditions of n:q 5:1. With nonnormal data, one finds in Table (as expected) that Type I error rates associated with T ML and T ML k are inflated, although the k-factor correction to T ML did provide some improvement when data were symmetric nonnormal. The first SB test statistic, T SB1, provided some control over Type I error rates, demonstrating performance with asymmetric nonnormal data that was noticeably superior to T ML. However, Type I error rates fell within the robustness range only at the largest sample size condition examined. Applying the k-factor correction to T SB1 yielded a test statistic that was robust at all combinations of sample sizes and distributional forms investigated, with model rejection rates even within the 95% CI for most sample size ratios above n:q = :1. In contrast to the inflated rejection rates 454 MULTIVARIATE BEHAVIORAL RESEARCH

17 Table Type I Error Rates (%) for Model A1 (85 df) J. Nevitt and G. Hancock condition: 1:1 :1 5:1 10:1 df + 1 p* n: D a T ML b.9 a.9 a.1 a T ML k b a 5.5 b 6.0 b a T SB a a.8 a 5.1 b 4. b 4. b.7 a T SB1 k.7 a 4.8 b 4.8 b 4.4 b 4.8 b 4.6 b 5.1 b 6.8 a 6.0 b 5.4 b 6.7 a 6.8 a a 4.5 b.7 a.8 a.9 a T SB 5. b a a T SB k T B b T B a T B 6. a 5.7 b a.9 a.0 a a T ADF b 0.0 T ADF a a T ADF b Note. Type I error rates are computed across 000 successful replications at the.05 level of significance. D = distributional form; D = 1, multivariate normal; D =, elliptically symmetric nonnormal; D =, asymmetric nonnormal. a indicates rejection rate within the robustness interval. b indicates rejection rate within the 95% CI. MULTIVARIATE BEHAVIORAL RESEARCH 455

18 associated with T SB1, model rejection rates for the second SB test statistic, T SB, tended to be attenuated. Applying the k-factor correction to T SB pushed model rejection rates even lower to near zero under most of the conditions examined. For the residual-based and ADF test statistics in Table, performance of the unadjusted T B and T ADF mirrored one another closely, with near 100% Type I error rates at all sample sizes and distributional forms investigated. The adjustments T B1 and T ADF1 also were close in performance to one another with near-zero rejection rates at n:q < 10:1 and showing monotonically decreasing Type I error rates with increasing departure from multivariate normality. The F-distributed adjustments, T B and T ADF, mirrored one another in performance at the test statistics minimum sample sizes with Type I error rates at or near zero; at larger sample size conditions T B generally maintained robustness while T ADF yielded suppressed model rejection rates. Type I error rates under the constrained LVP model (A) are shown in Table. Overall test statistic performance was similar to that seen under the unconstrained model, with the caveat that performance associated with all of the test statistics examined under the constrained model deteriorated to some extent. Applying the Bartlett k-factor correction to these test statistics led to improved Type I error performance in the constrained model, although at the smallest sample size conditions T ML k and T SB1 k exerted too much correction yielding Type I error rates below the.5% lower bound of the robustness criterion. The behavior of T SB and T SB k was again similar to the unconstrained model; the correction pushed Type I error rates to zero or near-zero values (in particular with nonnormal distributional forms). The T B and T ADF test statistics, and their adjusted forms, paralleled one another with respect to Type I error propensities in the constrained model. Table shows 100% model rejection rates for T B and T ADF, with zero or near-zero Type I error rates for all of the adjusted forms of T B and T ADF at their minimum sample size conditions. At other sample size conditions, model rejection rates were less than % for the -distributed T B1 and T ADF1 ; Type I error rates for the F-distributed T B and T ADF were within or slightly above the robustness interval. Table 4 presents Type I error results for the unconstrained CFA model (B1). With multivariate normal data, T ML was robust at n:q = 5:1 and 10:1; at smaller n:q ratios model rejection rates were inflated. The k-factor corrected form T ML k maintained Type I error rates (with normal data) to within the robustness interval at sample size conditions of n:q = 1:1 and larger. T ML and T ML k became inflated with nonnormal distributional forms with Type I error rates climbing as high as 98%. 456 MULTIVARIATE BEHAVIORAL RESEARCH

19 Table Type I Error Rates (%) for Model A (10 df) J. Nevitt and G. Hancock condition: 1:1 :1 5:1 10:1 df + 1 p* n: D T ML a 4.5 b.6 a. a T ML k T SB a.7 a 5.5 b 4.7 b 4.5 b T SB1 k b 5.7 b 4.8 b 5.6 b. 4.4 b 6.7 a a.6 a.6 a. a.8 a T SB T SB k T B T B T B T ADF T ADF T ADF 7. a b 0.7 Note: Type I error rates are computed across 000 successful replications at the.05 level of significance. D = distributional form; D = 1, multivariate normal; D =, elliptically symmetric nonnormal; D =, asymmetric nonnormal. a indicates rejection rate within the robustness interval. b indicates rejection rate within the 95% CI. MULTIVARIATE BEHAVIORAL RESEARCH 457

20 Table 4 Type I Error Rates (%) for Model B1 (10 df) condition: 1:1 :1 5:1 10:1 df + 1 p* n: D a 6. a T ML a 4. b 4.7 b 5.4 b a T ML k 7.0 a b a 6. a T SB a 5.6 b a a 4.8 b 4.8 b 5.6 b.4.9 a T SB1 k 5.1 b 4. b 5.1 b 4.4 b.6 a 5.4 b a 6.0 b b 5.4 b 4.8 b 5.6 b 4.1 b 4.4 b T SB a.8 a a 4.7 b T SB k a T B b 4.6 b T B a.4 a a 6.4 a 5.8 b 5.5 b a T B 4.1 b 4. b 4.1 b.7 a b 6.5 a.9 a. a.1 a a T ADF b 4.5 b 0.5 T ADF a a 1.4 a 5.8 b 5.5 b.1 T ADF a.5 a Note. Type I error rates are computed across 000 successful replications at the.05 level of significance. D = distributional form; D = 1, multivariate normal; D =, elliptically symmetric nonnormal; D =, asymmetric nonnormal. a indicates rejection rate within the robustness interval. b indicates rejection rate within the 95% CI. 458 MULTIVARIATE BEHAVIORAL RESEARCH

21 T SB1 showed robustness with most distributional forms at n:q = 5:1 and 10:1 (with the exception of n:q = 5:1 and asymmetric nonnormal data), but yielded inflated Type I error rates at smaller n:q ratios. The k-factor corrected form, T SB1 k, maintained Type I error rates to within the robustness interval at sample size conditions of n:q = 1:1 and larger, although error rates exceeded the upper bound of the robustness interval with asymmetric nonnormal data and n:q < 5:1. The T SB test statistic showed robustness with normal data, but then yielded attenuated model rejection rates with nonnormal distributional forms; the k-factor adjusted T SB k over-corrected, delivering Type I error rates near zero. Model rejection rates for T B and T ADF, T B1 and T ADF1, and T B and T ADF again mostly mirrored one another, with some local distinctions. The unadjusted test statistics yielded inflated Type I error rates that monotonically decreased with increasing sample size; note that even at the n:q = 10:1 ratio model rejection percentages associated with T B and T ADF were at 18%. The adjusted forms of T B and T ADF performed optimally under several of the sample size conditions investigated but suppressed model rejections with increasing nonnormality and decreasing sample size. The residual-based F- distributed statistic, T B, maintained robustness under all conditions for this model, with the exception of the n = df + 1 sample size minimum for this test statistic in which model rejection rates were at zero. Results for the constrained CFA model (B) are presented in Table 5. T ML produced inflated Type I error rates at every sample size and distributional form combination examined (even at the largest sample size condition and multivariate normal data). T SB1 yielded inflated Type I error rates that decreased with increasing sample size, approaching the 7.5% upper bound of the robustness criterion at the largest sample size conditions. Applying the k-factor correction to these test statistics improved Type I error control for T ML k with normal data and for T SB1 k with all distributional forms, albeit evidencing some degree of over-correction at the smallest sample size conditions. The T SB statistic controlled Type I error rates with normal data at all but the smallest sample size condition; error rates with nonnormal distributional forms were mostly attenuated. Applying the Bartlett k-factor correction, T SB k yielded near-zero model rejection rates for all conditions investigated. Results for T B and T ADF in Table 5 are similar to the patterns seen under the constrained LVP model. The unadjusted forms of the test statistics yielded 100% model rejection rates; the -adjusted T B1 and T ADF1 were robust only at the largest sample size condition and with normal data; the F-distributed statistics T B and T ADF yielded inflated Type I error rates ranging from % across the conditions examined here. MULTIVARIATE BEHAVIORAL RESEARCH 459

22 Table 5 Type I Error Rates (%) for Model B (196 df) condition: 1:1 :1 5:1 10:1 df + 1 p* n: D T ML a. a.9 a T ML k T SB a 4.5 b.9 a.7 a T SB1 k a 5.1 b 4.1 b 5.0 b b 5.6 b 5.6 b 5.5 b b.6 a.1 a.1 a. T SB T SB k T B a T B T B T ADF a 0.0 T ADF T ADF Note. Type I error rates are computed across 000 successful replications at the.05 level of significance. D = distributional form; D = 1, multivariate normal; D =, elliptically symmetric nonnormal; D =, asymmetric nonnormal. a indicates rejection rate within the robustness interval. b indicates rejection rate within the 95% CI. 460 MULTIVARIATE BEHAVIORAL RESEARCH

23 Empirical Power J. Nevitt and G. Hancock Results for follow-up power analyses are given in Tables 6-9, with each table paralleling one of the four models investigated in the Type I error analysis. As noted earlier, empirical power estimates presented in these tables are not considered absolute, but rather are relative estimates for the purpose of comparing performance of the test statistics investigated. Power was not analyzed for the residual-based and ADF test statistics at their respective sample size minima. The high frequency of non-convergence, improper solutions, and test statistic failure rates and the poor Type I error performance associated with these test statistics at their sample size lower bounds logically precluded them from further investigation. A general pattern is evidenced across all of the power tables, whereby power systematically increased with increasing sample size (as would be expected) and uniformly decreased with increasing departure from multivariate normality. Table 6 presents empirical power estimates for the misspecified unconstrained LVP Model. Note that for n = 50 the power estimate associated with every test statistic under all distributional forms was 100%, and as such is not useful for discriminating test statistic performance. At n < 50, however, test statistic performance with respect to power diverged. The Bartlett k-factor correction to T ML, T ML k, yielded empirical power at % with multivariate normal data. The SB test statistics and their k-factorcorrected forms generally yielded power at or above 80%. Of note is the power associated with T SB1 k which delivered model rejection rates near 90% at n = 5. In contrast T SB and T SB k delivered low power at the smallest sample size conditions power fell to about 5% for T SB k at the n = 5 sample size. Empirical power estimates for the residual-based and ADF test statistics in Table 6 suggest some noteworthy features. The residual-based test statistics generally exhibited greater power than the ADF test statistics; the F-distributed forms of T B and T ADF (T B and T ADF ) delivered greater power as compared to the -corrected forms of these test statistics (T B1 and T ADF1 ). Under the misspecified constrained LVP model, Table 7 reports empirical power estimates that are noticeably lower than those found in Table 6. For the T ML k test statistic empirical power was reported at 0% at the smallest sample size condition with multivariate normal data; at all larger sample sizes power was at or near 100%. At n = 18 (with all distributional forms) empirical power associated with T SB1 k dropped off, and went to near-zero for T SB k. Notice the 0% model rejection rate with all distributional forms for T B1 at n = 10. Again, the F-distributed correction to ADF yielded greater power than the - corrected form (e.g., 14% versus 80% with asymmetric nonnormal data). MULTIVARIATE BEHAVIORAL RESEARCH 461

Recovery of weak factor loadings in confirmatory factor analysis under conditions of model misspecification

Recovery of weak factor loadings in confirmatory factor analysis under conditions of model misspecification Behavior Research Methods 29, 41 (4), 138-152 doi:1.3758/brm.41.4.138 Recovery of weak factor loadings in confirmatory factor analysis under conditions of model misspecification CARMEN XIMÉNEZ Autonoma

More information

Misspecification in Nonrecursive SEMs 1. Nonrecursive Latent Variable Models under Misspecification

Misspecification in Nonrecursive SEMs 1. Nonrecursive Latent Variable Models under Misspecification Misspecification in Nonrecursive SEMs 1 Nonrecursive Latent Variable Models under Misspecification Misspecification in Nonrecursive SEMs 2 Abstract A problem central to structural equation modeling is

More information

Testing Structural Equation Models: The Effect of Kurtosis

Testing Structural Equation Models: The Effect of Kurtosis Testing Structural Equation Models: The Effect of Kurtosis Tron Foss, Karl G Jöreskog & Ulf H Olsson Norwegian School of Management October 18, 2006 Abstract Various chi-square statistics are used for

More information

ABSTRACT. Between-Subjects Design under Variance. Heterogeneity and Nonnormality. Evaluation

ABSTRACT. Between-Subjects Design under Variance. Heterogeneity and Nonnormality. Evaluation ABSTRACT Title of dissertation: Robust Means Modeling: An Alternative to Hypothesis Testing Of Mean Equality in the Between-Subjects Design under Variance Heterogeneity and Nonnormality Weihua Fan, Doctor

More information

Robust Means Modeling vs Traditional Robust Tests 1

Robust Means Modeling vs Traditional Robust Tests 1 Robust Means Modeling vs Traditional Robust Tests 1 Comparing Means under Heteroscedasticity and Nonnormality: Further Exploring Robust Means Modeling Alyssa Counsell Department of Psychology Ryerson University

More information

Scaled and adjusted restricted tests in. multi-sample analysis of moment structures. Albert Satorra. Universitat Pompeu Fabra.

Scaled and adjusted restricted tests in. multi-sample analysis of moment structures. Albert Satorra. Universitat Pompeu Fabra. Scaled and adjusted restricted tests in multi-sample analysis of moment structures Albert Satorra Universitat Pompeu Fabra July 15, 1999 The author is grateful to Peter Bentler and Bengt Muthen for their

More information

Extending the Robust Means Modeling Framework. Alyssa Counsell, Phil Chalmers, Matt Sigal, Rob Cribbie

Extending the Robust Means Modeling Framework. Alyssa Counsell, Phil Chalmers, Matt Sigal, Rob Cribbie Extending the Robust Means Modeling Framework Alyssa Counsell, Phil Chalmers, Matt Sigal, Rob Cribbie One-way Independent Subjects Design Model: Y ij = µ + τ j + ε ij, j = 1,, J Y ij = score of the ith

More information

Fit Indices Versus Test Statistics

Fit Indices Versus Test Statistics MULTIVARIATE BEHAVIORAL RESEARCH, 40(1), 115 148 Copyright 2005, Lawrence Erlbaum Associates, Inc. Fit Indices Versus Test Statistics Ke-Hai Yuan University of Notre Dame Model evaluation is one of the

More information

Department of Psychology, Beihang University, China 2. Department of Psychology, University of Notre Dame, USA 3

Department of Psychology, Beihang University, China 2. Department of Psychology, University of Notre Dame, USA 3 Chapter 03 Structural Equation Modeling With Many Variables: A Systematic Review of Issues and Developments Lifang Deng 1*, Miao Yang 2 and Katerina M Marcoulides 3 1 Department of Psychology, Beihang

More information

EVALUATION OF STRUCTURAL EQUATION MODELS

EVALUATION OF STRUCTURAL EQUATION MODELS 1 EVALUATION OF STRUCTURAL EQUATION MODELS I. Issues related to the initial specification of theoretical models of interest 1. Model specification: a. Measurement model: (i) EFA vs. CFA (ii) reflective

More information

Model Estimation Example

Model Estimation Example Ronald H. Heck 1 EDEP 606: Multivariate Methods (S2013) April 7, 2013 Model Estimation Example As we have moved through the course this semester, we have encountered the concept of model estimation. Discussions

More information

Model fit evaluation in multilevel structural equation models

Model fit evaluation in multilevel structural equation models Model fit evaluation in multilevel structural equation models Ehri Ryu Journal Name: Frontiers in Psychology ISSN: 1664-1078 Article type: Review Article Received on: 0 Sep 013 Accepted on: 1 Jan 014 Provisional

More information

A note on structured means analysis for a single group. André Beauducel 1. October 3 rd, 2015

A note on structured means analysis for a single group. André Beauducel 1. October 3 rd, 2015 Structured means analysis for a single group 1 A note on structured means analysis for a single group André Beauducel 1 October 3 rd, 2015 Abstract The calculation of common factor means in structured

More information

A simulation study to investigate the use of cutoff values for assessing model fit in covariance structure models

A simulation study to investigate the use of cutoff values for assessing model fit in covariance structure models Journal of Business Research 58 (2005) 935 943 A simulation study to investigate the use of cutoff values for assessing model fit in covariance structure models Subhash Sharma a, *, Soumen Mukherjee b,

More information

sempower Manual Morten Moshagen

sempower Manual Morten Moshagen sempower Manual Morten Moshagen 2018-03-22 Power Analysis for Structural Equation Models Contact: morten.moshagen@uni-ulm.de Introduction sempower provides a collection of functions to perform power analyses

More information

FIT CRITERIA PERFORMANCE AND PARAMETER ESTIMATE BIAS IN LATENT GROWTH MODELS WITH SMALL SAMPLES

FIT CRITERIA PERFORMANCE AND PARAMETER ESTIMATE BIAS IN LATENT GROWTH MODELS WITH SMALL SAMPLES FIT CRITERIA PERFORMANCE AND PARAMETER ESTIMATE BIAS IN LATENT GROWTH MODELS WITH SMALL SAMPLES Daniel M. McNeish Measurement, Statistics, and Evaluation University of Maryland, College Park Background

More information

Evaluating the Sensitivity of Goodness-of-Fit Indices to Data Perturbation: An Integrated MC-SGR Approach

Evaluating the Sensitivity of Goodness-of-Fit Indices to Data Perturbation: An Integrated MC-SGR Approach Evaluating the Sensitivity of Goodness-of-Fit Indices to Data Perturbation: An Integrated MC-SGR Approach Massimiliano Pastore 1 and Luigi Lombardi 2 1 Department of Psychology University of Cagliari Via

More information

Chapter 8. Models with Structural and Measurement Components. Overview. Characteristics of SR models. Analysis of SR models. Estimation of SR models

Chapter 8. Models with Structural and Measurement Components. Overview. Characteristics of SR models. Analysis of SR models. Estimation of SR models Chapter 8 Models with Structural and Measurement Components Good people are good because they've come to wisdom through failure. Overview William Saroyan Characteristics of SR models Estimation of SR models

More information

Introduction to Structural Equation Modeling

Introduction to Structural Equation Modeling Introduction to Structural Equation Modeling Notes Prepared by: Lisa Lix, PhD Manitoba Centre for Health Policy Topics Section I: Introduction Section II: Review of Statistical Concepts and Regression

More information

TESTING FOR NORMALITY IN THE LINEAR REGRESSION MODEL: AN EMPIRICAL LIKELIHOOD RATIO TEST

TESTING FOR NORMALITY IN THE LINEAR REGRESSION MODEL: AN EMPIRICAL LIKELIHOOD RATIO TEST Econometrics Working Paper EWP0402 ISSN 1485-6441 Department of Economics TESTING FOR NORMALITY IN THE LINEAR REGRESSION MODEL: AN EMPIRICAL LIKELIHOOD RATIO TEST Lauren Bin Dong & David E. A. Giles Department

More information

A Cautionary Note on the Use of LISREL s Automatic Start Values in Confirmatory Factor Analysis Studies R. L. Brown University of Wisconsin

A Cautionary Note on the Use of LISREL s Automatic Start Values in Confirmatory Factor Analysis Studies R. L. Brown University of Wisconsin A Cautionary Note on the Use of LISREL s Automatic Start Values in Confirmatory Factor Analysis Studies R. L. Brown University of Wisconsin The accuracy of parameter estimates provided by the major computer

More information

Joint Estimation of Risk Preferences and Technology: Further Discussion

Joint Estimation of Risk Preferences and Technology: Further Discussion Joint Estimation of Risk Preferences and Technology: Further Discussion Feng Wu Research Associate Gulf Coast Research and Education Center University of Florida Zhengfei Guan Assistant Professor Gulf

More information

Nesting and Equivalence Testing

Nesting and Equivalence Testing Nesting and Equivalence Testing Tihomir Asparouhov and Bengt Muthén August 13, 2018 Abstract In this note, we discuss the nesting and equivalence testing (NET) methodology developed in Bentler and Satorra

More information

Structural Equation Modeling

Structural Equation Modeling CHAPTER 23 Structural Equation Modeling JODIE B. ULLMAN AND PETER M. BENTLER A FOUR-STAGE GENERAL PROCESS OF MODELING 663 MODEL ESTIMATION TECHNIQUES AND TEST STATISTICS 667 MODEL EVALUATION 671 MODEL

More information

Uppsala University and Norwegian School of Management, b Uppsala University, Online publication date: 08 July 2010

Uppsala University and Norwegian School of Management, b Uppsala University, Online publication date: 08 July 2010 This article was downloaded by: [UAM University Autonoma de Madrid] On: 28 April 20 Access details: Access Details: [subscription number 93384845] Publisher Psychology Press Informa Ltd Registered in England

More information

Least Absolute Value vs. Least Squares Estimation and Inference Procedures in Regression Models with Asymmetric Error Distributions

Least Absolute Value vs. Least Squares Estimation and Inference Procedures in Regression Models with Asymmetric Error Distributions Journal of Modern Applied Statistical Methods Volume 8 Issue 1 Article 13 5-1-2009 Least Absolute Value vs. Least Squares Estimation and Inference Procedures in Regression Models with Asymmetric Error

More information

STRUCTURAL EQUATION MODELING. Khaled Bedair Statistics Department Virginia Tech LISA, Summer 2013

STRUCTURAL EQUATION MODELING. Khaled Bedair Statistics Department Virginia Tech LISA, Summer 2013 STRUCTURAL EQUATION MODELING Khaled Bedair Statistics Department Virginia Tech LISA, Summer 2013 Introduction: Path analysis Path Analysis is used to estimate a system of equations in which all of the

More information

Mixture Modeling. Identifying the Correct Number of Classes in a Growth Mixture Model. Davood Tofighi Craig Enders Arizona State University

Mixture Modeling. Identifying the Correct Number of Classes in a Growth Mixture Model. Davood Tofighi Craig Enders Arizona State University Identifying the Correct Number of Classes in a Growth Mixture Model Davood Tofighi Craig Enders Arizona State University Mixture Modeling Heterogeneity exists such that the data are comprised of two or

More information

SRMR in Mplus. Tihomir Asparouhov and Bengt Muthén. May 2, 2018

SRMR in Mplus. Tihomir Asparouhov and Bengt Muthén. May 2, 2018 SRMR in Mplus Tihomir Asparouhov and Bengt Muthén May 2, 2018 1 Introduction In this note we describe the Mplus implementation of the SRMR standardized root mean squared residual) fit index for the models

More information

A Study of Statistical Power and Type I Errors in Testing a Factor Analytic. Model for Group Differences in Regression Intercepts

A Study of Statistical Power and Type I Errors in Testing a Factor Analytic. Model for Group Differences in Regression Intercepts A Study of Statistical Power and Type I Errors in Testing a Factor Analytic Model for Group Differences in Regression Intercepts by Margarita Olivera Aguilar A Thesis Presented in Partial Fulfillment of

More information

1 The Robustness of LISREL Modeling Revisited

1 The Robustness of LISREL Modeling Revisited 1 The Robustness of LISREL Modeling Revisited Anne Boomsma 1 and Jeffrey J. Hoogland 2 This is page 1 Printer: Opaque this January 10, 2001 ABSTRACT Some robustness questions in structural equation modeling

More information

The Impact of Varying the Number of Measurement Invariance Constraints on. the Assessment of Between-Group Differences of Latent Means.

The Impact of Varying the Number of Measurement Invariance Constraints on. the Assessment of Between-Group Differences of Latent Means. The Impact of Varying the Number of Measurement on the Assessment of Between-Group Differences of Latent Means by Yuning Xu A Thesis Presented in Partial Fulfillment of the Requirements for the Degree

More information

Factor analysis. George Balabanis

Factor analysis. George Balabanis Factor analysis George Balabanis Key Concepts and Terms Deviation. A deviation is a value minus its mean: x - mean x Variance is a measure of how spread out a distribution is. It is computed as the average

More information

Simulating Uniform- and Triangular- Based Double Power Method Distributions

Simulating Uniform- and Triangular- Based Double Power Method Distributions Journal of Statistical and Econometric Methods, vol.6, no.1, 2017, 1-44 ISSN: 1792-6602 (print), 1792-6939 (online) Scienpress Ltd, 2017 Simulating Uniform- and Triangular- Based Double Power Method Distributions

More information

A Monte Carlo Simulation of the Robust Rank- Order Test Under Various Population Symmetry Conditions

A Monte Carlo Simulation of the Robust Rank- Order Test Under Various Population Symmetry Conditions Journal of Modern Applied Statistical Methods Volume 12 Issue 1 Article 7 5-1-2013 A Monte Carlo Simulation of the Robust Rank- Order Test Under Various Population Symmetry Conditions William T. Mickelson

More information

Structural Equation Modeling and Confirmatory Factor Analysis. Types of Variables

Structural Equation Modeling and Confirmatory Factor Analysis. Types of Variables /4/04 Structural Equation Modeling and Confirmatory Factor Analysis Advanced Statistics for Researchers Session 3 Dr. Chris Rakes Website: http://csrakes.yolasite.com Email: Rakes@umbc.edu Twitter: @RakesChris

More information

Multiple Comparison Procedures, Trimmed Means and Transformed Statistics. Rhonda K. Kowalchuk Southern Illinois University Carbondale

Multiple Comparison Procedures, Trimmed Means and Transformed Statistics. Rhonda K. Kowalchuk Southern Illinois University Carbondale Multiple Comparison Procedures 1 Multiple Comparison Procedures, Trimmed Means and Transformed Statistics Rhonda K. Kowalchuk Southern Illinois University Carbondale H. J. Keselman University of Manitoba

More information

3/10/03 Gregory Carey Cholesky Problems - 1. Cholesky Problems

3/10/03 Gregory Carey Cholesky Problems - 1. Cholesky Problems 3/10/03 Gregory Carey Cholesky Problems - 1 Cholesky Problems Gregory Carey Department of Psychology and Institute for Behavioral Genetics University of Colorado Boulder CO 80309-0345 Email: gregory.carey@colorado.edu

More information

Methodology Review: Applications of Distribution Theory in Studies of. Population Validity and Cross Validity. James Algina. University of Florida

Methodology Review: Applications of Distribution Theory in Studies of. Population Validity and Cross Validity. James Algina. University of Florida Distribution Theory 1 Methodology eview: Applications of Distribution Theory in Studies of Population Validity and Cross Validity by James Algina University of Florida and H. J. Keselman University of

More information

SC705: Advanced Statistics Instructor: Natasha Sarkisian Class notes: Introduction to Structural Equation Modeling (SEM)

SC705: Advanced Statistics Instructor: Natasha Sarkisian Class notes: Introduction to Structural Equation Modeling (SEM) SC705: Advanced Statistics Instructor: Natasha Sarkisian Class notes: Introduction to Structural Equation Modeling (SEM) SEM is a family of statistical techniques which builds upon multiple regression,

More information

AN EMPIRICAL LIKELIHOOD RATIO TEST FOR NORMALITY

AN EMPIRICAL LIKELIHOOD RATIO TEST FOR NORMALITY Econometrics Working Paper EWP0401 ISSN 1485-6441 Department of Economics AN EMPIRICAL LIKELIHOOD RATIO TEST FOR NORMALITY Lauren Bin Dong & David E. A. Giles Department of Economics, University of Victoria

More information

Testing structural equation models: the effect of kurtosis. Tron Foss BI Norwegian Business School. Karl G. Jøreskog BI Norwegian Business School

Testing structural equation models: the effect of kurtosis. Tron Foss BI Norwegian Business School. Karl G. Jøreskog BI Norwegian Business School This file was downloaded from the institutional repository BI Brage - http://brage.bibsys.no/bi (Open Access) Testing structural equation models: the effect of kurtosis Tron Foss BI Norwegian Business

More information

Accounting for Population Uncertainty in Covariance Structure Analysis

Accounting for Population Uncertainty in Covariance Structure Analysis Accounting for Population Uncertainty in Structure Analysis Boston College May 21, 2013 Joint work with: Michael W. Browne The Ohio State University matrix among observed variables are usually implied

More information

Psychology 282 Lecture #4 Outline Inferences in SLR

Psychology 282 Lecture #4 Outline Inferences in SLR Psychology 282 Lecture #4 Outline Inferences in SLR Assumptions To this point we have not had to make any distributional assumptions. Principle of least squares requires no assumptions. Can use correlations

More information

Assessing the relation between language comprehension and performance in general chemistry. Appendices

Assessing the relation between language comprehension and performance in general chemistry. Appendices Assessing the relation between language comprehension and performance in general chemistry Daniel T. Pyburn a, Samuel Pazicni* a, Victor A. Benassi b, and Elizabeth E. Tappin c a Department of Chemistry,

More information

On Selecting Tests for Equality of Two Normal Mean Vectors

On Selecting Tests for Equality of Two Normal Mean Vectors MULTIVARIATE BEHAVIORAL RESEARCH, 41(4), 533 548 Copyright 006, Lawrence Erlbaum Associates, Inc. On Selecting Tests for Equality of Two Normal Mean Vectors K. Krishnamoorthy and Yanping Xia Department

More information

Journal of Multivariate Analysis. Use of prior information in the consistent estimation of regression coefficients in measurement error models

Journal of Multivariate Analysis. Use of prior information in the consistent estimation of regression coefficients in measurement error models Journal of Multivariate Analysis 00 (2009) 498 520 Contents lists available at ScienceDirect Journal of Multivariate Analysis journal homepage: www.elsevier.com/locate/jmva Use of prior information in

More information

Confirmatory Factor Analysis. Psych 818 DeShon

Confirmatory Factor Analysis. Psych 818 DeShon Confirmatory Factor Analysis Psych 818 DeShon Purpose Takes factor analysis a few steps further. Impose theoretically interesting constraints on the model and examine the resulting fit of the model with

More information

Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful?

Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful? Journal of Modern Applied Statistical Methods Volume 10 Issue Article 13 11-1-011 Estimation and Hypothesis Testing in LAV Regression with Autocorrelated Errors: Is Correction for Autocorrelation Helpful?

More information

Can Variances of Latent Variables be Scaled in Such a Way That They Correspond to Eigenvalues?

Can Variances of Latent Variables be Scaled in Such a Way That They Correspond to Eigenvalues? International Journal of Statistics and Probability; Vol. 6, No. 6; November 07 ISSN 97-703 E-ISSN 97-7040 Published by Canadian Center of Science and Education Can Variances of Latent Variables be Scaled

More information

Computationally Efficient Estimation of Multilevel High-Dimensional Latent Variable Models

Computationally Efficient Estimation of Multilevel High-Dimensional Latent Variable Models Computationally Efficient Estimation of Multilevel High-Dimensional Latent Variable Models Tihomir Asparouhov 1, Bengt Muthen 2 Muthen & Muthen 1 UCLA 2 Abstract Multilevel analysis often leads to modeling

More information

Probability and Statistics

Probability and Statistics Probability and Statistics Kristel Van Steen, PhD 2 Montefiore Institute - Systems and Modeling GIGA - Bioinformatics ULg kristel.vansteen@ulg.ac.be CHAPTER 4: IT IS ALL ABOUT DATA 4a - 1 CHAPTER 4: IT

More information

Bayesian Analysis of Latent Variable Models using Mplus

Bayesian Analysis of Latent Variable Models using Mplus Bayesian Analysis of Latent Variable Models using Mplus Tihomir Asparouhov and Bengt Muthén Version 2 June 29, 2010 1 1 Introduction In this paper we describe some of the modeling possibilities that are

More information

An Alternative to Cronbach s Alpha: A L-Moment Based Measure of Internal-consistency Reliability

An Alternative to Cronbach s Alpha: A L-Moment Based Measure of Internal-consistency Reliability Southern Illinois University Carbondale OpenSIUC Book Chapters Educational Psychology and Special Education 013 An Alternative to Cronbach s Alpha: A L-Moment Based Measure of Internal-consistency Reliability

More information

THE 'IMPROVED' BROWN AND FORSYTHE TEST FOR MEAN EQUALITY: SOME THINGS CAN'T BE FIXED

THE 'IMPROVED' BROWN AND FORSYTHE TEST FOR MEAN EQUALITY: SOME THINGS CAN'T BE FIXED THE 'IMPROVED' BROWN AND FORSYTHE TEST FOR MEAN EQUALITY: SOME THINGS CAN'T BE FIXED H. J. Keselman Rand R. Wilcox University of Manitoba University of Southern California Winnipeg, Manitoba Los Angeles,

More information

Research Article A Nonparametric Two-Sample Wald Test of Equality of Variances

Research Article A Nonparametric Two-Sample Wald Test of Equality of Variances Advances in Decision Sciences Volume 211, Article ID 74858, 8 pages doi:1.1155/211/74858 Research Article A Nonparametric Two-Sample Wald Test of Equality of Variances David Allingham 1 andj.c.w.rayner

More information

Size and Power of the RESET Test as Applied to Systems of Equations: A Bootstrap Approach

Size and Power of the RESET Test as Applied to Systems of Equations: A Bootstrap Approach Size and Power of the RESET Test as Applied to Systems of Equations: A Bootstrap Approach Ghazi Shukur Panagiotis Mantalos International Business School Department of Statistics Jönköping University Lund

More information

Improper Solutions in Exploratory Factor Analysis: Causes and Treatments

Improper Solutions in Exploratory Factor Analysis: Causes and Treatments Improper Solutions in Exploratory Factor Analysis: Causes and Treatments Yutaka Kano Faculty of Human Sciences, Osaka University Suita, Osaka 565, Japan. email: kano@hus.osaka-u.ac.jp Abstract: There are

More information

Introduction to Matrix Algebra and the Multivariate Normal Distribution

Introduction to Matrix Algebra and the Multivariate Normal Distribution Introduction to Matrix Algebra and the Multivariate Normal Distribution Introduction to Structural Equation Modeling Lecture #2 January 18, 2012 ERSH 8750: Lecture 2 Motivation for Learning the Multivariate

More information

A Threshold-Free Approach to the Study of the Structure of Binary Data

A Threshold-Free Approach to the Study of the Structure of Binary Data International Journal of Statistics and Probability; Vol. 2, No. 2; 2013 ISSN 1927-7032 E-ISSN 1927-7040 Published by Canadian Center of Science and Education A Threshold-Free Approach to the Study of

More information

ABSTRACT. Chair, Dr. Gregory R. Hancock, Department of. interactions as a function of the size of the interaction effect, sample size, the loadings of

ABSTRACT. Chair, Dr. Gregory R. Hancock, Department of. interactions as a function of the size of the interaction effect, sample size, the loadings of ABSTRACT Title of Document: A COMPARISON OF METHODS FOR TESTING FOR INTERACTION EFFECTS IN STRUCTURAL EQUATION MODELING Brandi A. Weiss, Doctor of Philosophy, 00 Directed By: Chair, Dr. Gregory R. Hancock,

More information

Interpreting Regression Results

Interpreting Regression Results Interpreting Regression Results Carlo Favero Favero () Interpreting Regression Results 1 / 42 Interpreting Regression Results Interpreting regression results is not a simple exercise. We propose to split

More information

How to use Stata s sem with small samples? New corrections for the L. R. χ 2 statistics and fit indices

How to use Stata s sem with small samples? New corrections for the L. R. χ 2 statistics and fit indices How to use Stata s sem with small samples? New corrections for the L. R. χ 2 statistics and fit indices Meeting of the German Stata User Group at the Konstanz University, June 22nd, 218?All models are

More information

CHAPTER 17 CHI-SQUARE AND OTHER NONPARAMETRIC TESTS FROM: PAGANO, R. R. (2007)

CHAPTER 17 CHI-SQUARE AND OTHER NONPARAMETRIC TESTS FROM: PAGANO, R. R. (2007) FROM: PAGANO, R. R. (007) I. INTRODUCTION: DISTINCTION BETWEEN PARAMETRIC AND NON-PARAMETRIC TESTS Statistical inference tests are often classified as to whether they are parametric or nonparametric Parameter

More information

Assessing Factorial Invariance in Ordered-Categorical Measures

Assessing Factorial Invariance in Ordered-Categorical Measures Multivariate Behavioral Research, 39 (3), 479-515 Copyright 2004, Lawrence Erlbaum Associates, Inc. Assessing Factorial Invariance in Ordered-Categorical Measures Roger E. Millsap and Jenn Yun-Tein Arizona

More information

Causal Inference Using Nonnormality Yutaka Kano and Shohei Shimizu 1

Causal Inference Using Nonnormality Yutaka Kano and Shohei Shimizu 1 Causal Inference Using Nonnormality Yutaka Kano and Shohei Shimizu 1 Path analysis, often applied to observational data to study causal structures, describes causal relationship between observed variables.

More information

PLEASE SCROLL DOWN FOR ARTICLE. Full terms and conditions of use:

PLEASE SCROLL DOWN FOR ARTICLE. Full terms and conditions of use: This article was downloaded by: [Howell, Roy][Texas Tech University] On: 15 December 2009 Access details: Access Details: [subscription number 907003254] Publisher Psychology Press Informa Ltd Registered

More information

What is in the Book: Outline

What is in the Book: Outline Estimating and Testing Latent Interactions: Advancements in Theories and Practical Applications Herbert W Marsh Oford University Zhonglin Wen South China Normal University Hong Kong Eaminations Authority

More information

MODEL IMPLIED INSTRUMENTAL VARIABLE ESTIMATION FOR MULTILEVEL CONFIRMATORY FACTOR ANALYSIS. Michael L. Giordano

MODEL IMPLIED INSTRUMENTAL VARIABLE ESTIMATION FOR MULTILEVEL CONFIRMATORY FACTOR ANALYSIS. Michael L. Giordano MODEL IMPLIED INSTRUMENTAL VARIABLE ESTIMATION FOR MULTILEVEL CONFIRMATORY FACTOR ANALYSIS Michael L. Giordano A thesis submitted to the faculty at the University of North Carolina at Chapel Hill in partial

More information

A nonparametric two-sample wald test of equality of variances

A nonparametric two-sample wald test of equality of variances University of Wollongong Research Online Faculty of Informatics - Papers (Archive) Faculty of Engineering and Information Sciences 211 A nonparametric two-sample wald test of equality of variances David

More information

An Introduction to Path Analysis

An Introduction to Path Analysis An Introduction to Path Analysis PRE 905: Multivariate Analysis Lecture 10: April 15, 2014 PRE 905: Lecture 10 Path Analysis Today s Lecture Path analysis starting with multivariate regression then arriving

More information

Streamlining Missing Data Analysis by Aggregating Multiple Imputations at the Data Level

Streamlining Missing Data Analysis by Aggregating Multiple Imputations at the Data Level Streamlining Missing Data Analysis by Aggregating Multiple Imputations at the Data Level A Monte Carlo Simulation to Test the Tenability of the SuperMatrix Approach Kyle M Lang Quantitative Psychology

More information

Chapter 5. Introduction to Path Analysis. Overview. Correlation and causation. Specification of path models. Types of path models

Chapter 5. Introduction to Path Analysis. Overview. Correlation and causation. Specification of path models. Types of path models Chapter 5 Introduction to Path Analysis Put simply, the basic dilemma in all sciences is that of how much to oversimplify reality. Overview H. M. Blalock Correlation and causation Specification of path

More information

CONFIRMATORY FACTOR ANALYSIS

CONFIRMATORY FACTOR ANALYSIS 1 CONFIRMATORY FACTOR ANALYSIS The purpose of confirmatory factor analysis (CFA) is to explain the pattern of associations among a set of observed variables in terms of a smaller number of underlying latent

More information

Inference with Heywood cases

Inference with Heywood cases Inference with Joint work with Kenneth Bollen (UNC) and Victoria Savalei (UBC) NSF support from SES-0617193 with funds from SSA September 18, 2009 What is a case? (1931) considered characterization of

More information

MA 575 Linear Models: Cedric E. Ginestet, Boston University Non-parametric Inference, Polynomial Regression Week 9, Lecture 2

MA 575 Linear Models: Cedric E. Ginestet, Boston University Non-parametric Inference, Polynomial Regression Week 9, Lecture 2 MA 575 Linear Models: Cedric E. Ginestet, Boston University Non-parametric Inference, Polynomial Regression Week 9, Lecture 2 1 Bootstrapped Bias and CIs Given a multiple regression model with mean and

More information

General structural model Part 2: Categorical variables and beyond. Psychology 588: Covariance structure and factor models

General structural model Part 2: Categorical variables and beyond. Psychology 588: Covariance structure and factor models General structural model Part 2: Categorical variables and beyond Psychology 588: Covariance structure and factor models Categorical variables 2 Conventional (linear) SEM assumes continuous observed variables

More information

ABSTRACT. Phillip Edward Gagné. priori information about population membership. There have, however, been

ABSTRACT. Phillip Edward Gagné. priori information about population membership. There have, however, been ABSTRACT Title of dissertation: GENERALIZED CONFIRMATORY FACTOR MIXTURE MODELS: A TOOL FOR ASSESSING FACTORIAL INVARIANCE ACROSS UNSPECIFIED POPULATIONS Phillip Edward Gagné Dissertation directed by: Professor

More information

A Multivariate Two-Sample Mean Test for Small Sample Size and Missing Data

A Multivariate Two-Sample Mean Test for Small Sample Size and Missing Data A Multivariate Two-Sample Mean Test for Small Sample Size and Missing Data Yujun Wu, Marc G. Genton, 1 and Leonard A. Stefanski 2 Department of Biostatistics, School of Public Health, University of Medicine

More information

What s New in Econometrics. Lecture 13

What s New in Econometrics. Lecture 13 What s New in Econometrics Lecture 13 Weak Instruments and Many Instruments Guido Imbens NBER Summer Institute, 2007 Outline 1. Introduction 2. Motivation 3. Weak Instruments 4. Many Weak) Instruments

More information

Conventional And Robust Paired And Independent-Samples t Tests: Type I Error And Power Rates

Conventional And Robust Paired And Independent-Samples t Tests: Type I Error And Power Rates Journal of Modern Applied Statistical Methods Volume Issue Article --3 Conventional And And Independent-Samples t Tests: Type I Error And Power Rates Katherine Fradette University of Manitoba, umfradet@cc.umanitoba.ca

More information

Evaluation of structural equation models. Hans Baumgartner Penn State University

Evaluation of structural equation models. Hans Baumgartner Penn State University Evaluation of structural equation models Hans Baumgartner Penn State University Issues related to the initial specification of theoretical models of interest Model specification: Measurement model: EFA

More information

Advising on Research Methods: A consultant's companion. Herman J. Ader Gideon J. Mellenbergh with contributions by David J. Hand

Advising on Research Methods: A consultant's companion. Herman J. Ader Gideon J. Mellenbergh with contributions by David J. Hand Advising on Research Methods: A consultant's companion Herman J. Ader Gideon J. Mellenbergh with contributions by David J. Hand Contents Preface 13 I Preliminaries 19 1 Giving advice on research methods

More information

Appendix from L. J. Revell, On the Analysis of Evolutionary Change along Single Branches in a Phylogeny

Appendix from L. J. Revell, On the Analysis of Evolutionary Change along Single Branches in a Phylogeny 008 by The University of Chicago. All rights reserved.doi: 10.1086/588078 Appendix from L. J. Revell, On the Analysis of Evolutionary Change along Single Branches in a Phylogeny (Am. Nat., vol. 17, no.

More information

Robust Confidence Intervals for Effects Sizes in Multiple Linear Regression

Robust Confidence Intervals for Effects Sizes in Multiple Linear Regression Robust Confidence Intervals for Effects Sizes in Multiple Linear Regression Paul Dudgeon Melbourne School of Psychological Sciences The University of Melbourne. Vic. 3010 AUSTRALIA dudgeon@unimelb.edu.au

More information

UNIVERSITY OF CALGARY. The Influence of Model Components and Misspecification Type on the Performance of the

UNIVERSITY OF CALGARY. The Influence of Model Components and Misspecification Type on the Performance of the UNIVERSITY OF CALGARY The Influence of Model Components and Misspecification Type on the Performance of the Comparative Fit Index (CFI) and the Root Mean Square Error of Approximation (RMSEA) in Structural

More information

An Investigation of the Accuracy of Parallel Analysis for Determining the Number of Factors in a Factor Analysis

An Investigation of the Accuracy of Parallel Analysis for Determining the Number of Factors in a Factor Analysis Western Kentucky University TopSCHOLAR Honors College Capstone Experience/Thesis Projects Honors College at WKU 6-28-2017 An Investigation of the Accuracy of Parallel Analysis for Determining the Number

More information

INTRODUCTION TO STRUCTURAL EQUATION MODELS

INTRODUCTION TO STRUCTURAL EQUATION MODELS I. Description of the course. INTRODUCTION TO STRUCTURAL EQUATION MODELS A. Objectives and scope of the course. B. Logistics of enrollment, auditing, requirements, distribution of notes, access to programs.

More information

An Introduction to Mplus and Path Analysis

An Introduction to Mplus and Path Analysis An Introduction to Mplus and Path Analysis PSYC 943: Fundamentals of Multivariate Modeling Lecture 10: October 30, 2013 PSYC 943: Lecture 10 Today s Lecture Path analysis starting with multivariate regression

More information

Detection and quantification capabilities

Detection and quantification capabilities 18.4.3.7 Detection and quantification capabilities Among the most important Performance Characteristics of the Chemical Measurement Process (CMP) are those that can serve as measures of the underlying

More information

POWER AND TYPE I ERROR RATE COMPARISON OF MULTIVARIATE ANALYSIS OF VARIANCE

POWER AND TYPE I ERROR RATE COMPARISON OF MULTIVARIATE ANALYSIS OF VARIANCE POWER AND TYPE I ERROR RATE COMPARISON OF MULTIVARIATE ANALYSIS OF VARIANCE Supported by Patrick Adebayo 1 and Ahmed Ibrahim 1 Department of Statistics, University of Ilorin, Kwara State, Nigeria Department

More information

Ruth E. Mathiowetz. Chapel Hill 2010

Ruth E. Mathiowetz. Chapel Hill 2010 Evaluating Latent Variable Interactions with Structural Equation Mixture Models Ruth E. Mathiowetz A thesis submitted to the faculty of the University of North Carolina at Chapel Hill in partial fulfillment

More information

Wooldridge, Introductory Econometrics, 3d ed. Chapter 9: More on specification and data problems

Wooldridge, Introductory Econometrics, 3d ed. Chapter 9: More on specification and data problems Wooldridge, Introductory Econometrics, 3d ed. Chapter 9: More on specification and data problems Functional form misspecification We may have a model that is correctly specified, in terms of including

More information

Condition 9 and 10 Tests of Model Confirmation with SEM Techniques

Condition 9 and 10 Tests of Model Confirmation with SEM Techniques Condition 9 and 10 Tests of Model Confirmation with SEM Techniques Dr. Larry J. Williams CARMA Director Donald and Shirley Clifton Chair of Survey Science Professor of Management University of Nebraska

More information

Controlling for latent confounding by confirmatory factor analysis (CFA) Blinded Blinded

Controlling for latent confounding by confirmatory factor analysis (CFA) Blinded Blinded Controlling for latent confounding by confirmatory factor analysis (CFA) Blinded Blinded 1 Background Latent confounder is common in social and behavioral science in which most of cases the selection mechanism

More information

Strati cation in Multivariate Modeling

Strati cation in Multivariate Modeling Strati cation in Multivariate Modeling Tihomir Asparouhov Muthen & Muthen Mplus Web Notes: No. 9 Version 2, December 16, 2004 1 The author is thankful to Bengt Muthen for his guidance, to Linda Muthen

More information

Orthogonal, Planned and Unplanned Comparisons

Orthogonal, Planned and Unplanned Comparisons This is a chapter excerpt from Guilford Publications. Data Analysis for Experimental Design, by Richard Gonzalez Copyright 2008. 8 Orthogonal, Planned and Unplanned Comparisons 8.1 Introduction In this

More information

Lectures 5 & 6: Hypothesis Testing

Lectures 5 & 6: Hypothesis Testing Lectures 5 & 6: Hypothesis Testing in which you learn to apply the concept of statistical significance to OLS estimates, learn the concept of t values, how to use them in regression work and come across

More information

HANDBOOK OF APPLICABLE MATHEMATICS

HANDBOOK OF APPLICABLE MATHEMATICS HANDBOOK OF APPLICABLE MATHEMATICS Chief Editor: Walter Ledermann Volume VI: Statistics PART A Edited by Emlyn Lloyd University of Lancaster A Wiley-Interscience Publication JOHN WILEY & SONS Chichester

More information

A Practitioner s Guide to Cluster-Robust Inference

A Practitioner s Guide to Cluster-Robust Inference A Practitioner s Guide to Cluster-Robust Inference A. C. Cameron and D. L. Miller presented by Federico Curci March 4, 2015 Cameron Miller Cluster Clinic II March 4, 2015 1 / 20 In the previous episode

More information