An Investigation of the Power of the Likelihood Ratio Goodness-of-Fit Statistic in Detecting Differential Item Functioning

The purpose of this study was to investigate the power and Type I error rate of the likelihood ratio goodness-of-fit (LR) statistic in detecting differential item functioning (DIF) under Samejima's (1969, 1972) graded response model. A multiple-replication Monte Carlo study was utilized in which DIF was modeled in simulated data sets which were then calibrated with MULTILOG (Thissen, 1991) using hierarchically nested item response models. In addition, the power and Type I error rate of the Mantel (1963) approach for detecting DIF in ordered response categories were investigated using the same simulated data, for comparative purposes. The power of both the Mantel and LR procedures was affected by sample size, as expected. The LR procedure lacked the power to consistently detect DIF when it existed in reference/focal groups with sample sizes as small as 500/500. The Mantel procedure maintained control of its Type I error rate and was more powerful than the LR procedure when the comparison group ability distributions were identical and there was a constant DIF pattern. On the other hand, the Mantel procedure lost control of its Type I error rate, whereas the LR procedure did not, when the comparison groups differed in mean ability; and the LR procedure demonstrated a profound power advantage over the Mantel procedure under conditions of balanced DIF in which the comparison group ability distributions were identical. The choice and subsequent use of any procedure requires a thorough understanding of the power and Type I error rates of the procedure under varying conditions of DIF pattern, comparison group ability distributions.–or as a surrogate, observed score distributions–and item characteristics.     

Testing Factorial Invariance in Multilevel Data: A Monte Carlo Study

Testing factorial invariance has recently gained more attention in different social science disciplines. Nevertheless, when examining factorial invariance, it is generally assumed that the observations are independent of each other, which might not be always true. In this study, we examined the impact of testing factorial invariance in multilevel data, especially when the dependency issue is not taken into account. We considered a set of design factors, including number of clusters, cluster size, and intraclass correlation (ICC) at different levels. The simulation results showed that the test of factorial invariance became more liberal (or had inflated Type I error rate) in terms of rejecting the null hypothesis of invariance held between groups when the dependency was not considered in the analysis. Additionally, the magnitude of the inflation in the Type I error rate was a function of both ICC and cluster size. Implications of the findings and limitations are discussed.     

Evaluating Type I Error and Power Rates Using an Effect Size Measure With the Logistic Regression Procedure for DIF Detection

The logistic regression (LR) procedure for differential item functioning (DIF) detection is a model-based approach designed to identify both uniform and nonuniform DIF. However, this procedure tends to produce inflated Type I errors. This outcome is problematic because it can result in the inefficient use of testing resources, and it may interfere with the study of the underlying causes of DIF. Recently, an effect size measure was developed for the LR DIF procedure and a classification method was proposed. However, the effect size measure and classification method have not been systematically investigated. In this study, we developed a new classification method based on those established for the Simultaneous Item Bias Test. A simulation study also was conducted to determine if the effect size measure affects the Type I error and power rates for the LR DIF procedure across sample sizes, ability distributions, and percentage of DIF items included on a test. The results indicate that the inclusion of the effect size measure can substantially reduce Type I error rates when large sample sizes are used, although there is also a reduction in power.     

Empirical Validation of DIMTEST on Nonnormal Ability Distributions

DIMTEST is a nonparametric statistical test procedure for assessing unidimensionality of binary item response data. The development of Stout's statistic, T, used in the DIMTEST procedure, does not require the assumption of a particular parametric form for the ability distributions or the item response functions. The purpose of the present study was to empirically investigate the performance of the statistic T with respect to different shapes of ability distributions. Several nonnormal distributions, both symmetric and nonsymmetric, were considered for this purpose. Other factors varied in the study were test length, sample size, and the level of correlation between abilities. The results of Type I error and power studies showed that the test statistic T exhibited consistently similar performance for all different shapes of ability distributions investigated in the study, which confirmed the nonparametric nature of the statistic T.     

A Comparison of Alternative Matching Strategies for DIF Detection in Tests That Are Multidimensional

Most currently accepted approaches for identifying differentially functioning test items compare performance across groups after first matching examinees on the ability of interest. The typical basis for this matching is the total test score. Previous research indicates that when the test is not approximately unidimensional, matching using the total test score may result in an inflated Type I error rate. This study compares the results of differential item functioning (DIF) analysis with matching based on the total test score, matching based on subtest scores, or multivariate matching using multiple subtest scores. Analysis of both actual and simulated data indicate that for the dimensionally complex test examined in this study, using the total test score as the matching criterion is inappropriate. The results suggest that matching on multiple subtest scores simultaneously may be superior to using either the total test score or individual relevant subtest scores.     

A Comparison of the Logistic Regression and Contingency Table Methods for Simultaneous Detection of Uniform and Nonuniform DIF

In this study, we investigate the logistic regression (LR), Mantel-Haenszel (MH), and Breslow-Day (BD) procedures for the simultaneous detection of both uniform and nonuniform differential item functioning (DIF). A simulation study was used to assess and compare the Type I error rate and power of a combined decision rule (CDR), which assesses DIF using a combination of the decisions made with BD and MH to those of LR. The results revealed that while the Type I error rate of CDR was consistently below the nominal alpha level, the Type I error rate of LR was high for the conditions having unequal ability distributions. In addition, the power of CDR was consistently higher than that of LR across all forms of DIF.     

Simulation Studies of the Effects of Small Sample Size and Studied Item Parameters on SIBTEST and Mantel-Haenszel Type I Error Performance

Two simulation studies investigated Type I error performance of two statistical procedures for detecting differential item functioning (DIF): SIBTEST and Mantel-Haenszel (MH). Because MH and SIBTEST are based on asymptotic distributions requiring "large" numbers of examinees, the first study examined Type 1 error for small sample sizes. No significant Type I error inflation occurred for either procedure. Because MH has the potential for Type I error inflation for non-Rasch models, the second study used a markedly non-Rasch test and systematically varied the shape and location of the studied item. When differences in distribution across examinee group of the measured ability were present, both procedures displayed inflated Type 1 error for certain items; MH displayed the greater inflation. Also, both procedures displayed statistically biased estimation of the zero DIF for certain items, though SIBTEST displayed much less than MH. When no latent distributional differences were present, both procedures performed satisfactorily under all conditions.     

Effect of Sample Size Ratio and Model Misfit When Using the Difficulty Parameter Differences Procedure to Detect DIF

This study examined the effect of sample size ratio and model misfit on the Type I error rates and power of the Difficulty Parameter Differences procedure using Winsteps. A unidimensional 30-item test with responses from 130,000 examinees was simulated and four independent variables were manipulated: sample size ratio (20/100/250/500/1000); model fit/misfit (1 PL and 3PLc =. 15 models); impact (no difference/mean differences/variance differences/mean and variance differences); and percentage of items with uniform and nonuniform DIF (0%/10%/20%). In general, the results indicate the importance of ensuring model fit to achieve greater control of Type I error and adequate statistical power. The manipulated variables produced inflated Type I error rates, which were well controlled when a measure of DIF magnitude was applied. Sample size ratio also had an effect on the power of the procedure. The paper discusses the practical implications of these results.     

Evaluating the Wald Test for Item‐Level Comparison of Saturated and Reduced Models in Cognitive Diagnosis

This article used the Wald test to evaluate the item‐level fit of a saturated cognitive diagnosis model (CDM) relative to the fits of the reduced models it subsumes. A simulation study was carried out to examine the Type I error and power of the Wald test in the context of the G‐DINA model. Results show that when the sample size is small and a larger number of attributes are required, the Type I error rate of the Wald test for the DINA and DINO models can be higher than the nominal significance levels, while the Type I error rate of the A‐CDM is closer to the nominal significance levels. However, with larger sample sizes, the Type I error rates for the three models are closer to the nominal significance levels. In addition, the Wald test has excellent statistical power to detect when the true underlying model is none of the reduced models examined even for relatively small sample sizes. The performance of the Wald test was also examined with real data. With an increasing number of CDMs from which to choose, this article provides an important contribution toward advancing the use of CDMs in practical educational settings.     

Examining type I error and power for detection of differential item and testlet functioning

In this study, the effectiveness of detection of differential item functioning (DIF) and testlet DIF using SIBTEST and Poly-SIBTEST were examined in tests composed of testlets. An example using data from a reading comprehension test showed that results from SIBTEST and Poly-SIBTEST were not completely consistent in the detection of DIF and testlet DIF. Results from a simulation study indicated that SIBTEST appeared to maintain type I error control for most conditions, except in some instances in which the magnitude of simulated DIF tended to increase. This same pattern was present for the Poly-SIBTEST results, although Poly-SIBTEST demonstrated markedly less control of type I errors. Type I error control with Poly-SIBTEST was lower for those conditions for which the ability was unmatched to test difficulty. The power results for SIBTEST were not adversely affected, when the size and percent of simulated DIF increased. Although Poly-SIBTEST failed to control type I errors in over 85% of the conditions simulated, in those conditions for which type I error control was maintained, Poly-SIBTEST demonstrated higher power than SIBTEST.     

The Use of Multiple Imputation for Missing Data in Uniform DIF Analysis: Power and Type I Error Rates

Methods of uniform differential item functioning (DIF) detection have been extensively studied in the complete data case. However, less work has been done examining the performance of these methods when missing item responses are present. Research that has been done in this regard appears to indicate that treating missing item responses as incorrect can lead to inflated Type I error rates (false detection of DIF). The current study builds on this prior research by investigating the utility of multiple imputation methods for missing item responses, in conjunction with standard DIF detection techniques. Results of the study support the use of multiple imputation for dealing with missing item responses. The article concludes with a discussion of these results for multiple imputation in conjunction with other research findings supporting its use in the context of item parameter estimation with missing data.     

The Effects of Type I Error Rate and Power of the ANCOVA F Test and Selected Alternatives Under Nonnormality and Variance Heterogeneity

The authors sought to identify through Monte Carlo simulations those conditions for which analysis of covariance (ANCOVA) does not maintain adequate Type I error rates and power. The conditions that were manipulated included assumptions of normality and variance homogeneity, sample size, number of treatment groups, and strength of the covariate-dependent variable relationship. Alternative tests studied were Quade's procedure, Puri and Sen's solution, Burnett and Barr's rank difference scores, Conover and Iman's rank transformation test, Hettmansperger's procedure, and the Puri-Sen-Harwell-Serlin test. For balanced designs, the ANCOVA F test was robust and was often the most powerful test through all sample-size designs and distributional configurations. With unbalanced designs, with variance heterogeneity, and when the largest treatment-group variance was matched with the largest group sample size, the nonparametric alternatives generally outperformed the ANCOVA test. When sample size and variance ratio were inversely coupled, all tests became very liberal; no test maintained adequate control over Type I error.     

Multiplicity Control in Structural Equation Modeling: Incorporating Parameter Dependencies

When structural equation modeling (SEM) analyses are conducted, significance tests for all important model relationships (parameters including factor loadings, covariances, etc.) are typically conducted at a specified nominal Type I error rate (α). Despite the fact that many significance tests are often conducted in SEM, rarely is multiplicity control applied. Cribbie (2000, 2007) demonstrated that without some form of adjustment, the familywise Type I error rate can become severely inflated. Cribbie also confirmed that the popular Bonferroni method was overly conservative due to the correlations among the parameters in the model. The purpose of this study was to compare the Type I error rates and per-parameter power of traditional multiplicity strategies with those of adjusted Bonferroni procedures that incorporate not only the number of tests in a family, but also the degree of correlation between parameters. The adjusted Bonferroni procedures were found to produce per-parameter power rates higher than the original Bonferroni procedure without inflating the familywise error rate.     

A Comparison of Chained Linear and Poststratification Linear Equating Under Different Testing Conditions

In this study I compared results of chained linear, Tucker, and Levine-observed score equatings under conditions where the new and old forms samples were similar in ability and also when they were different in ability. The length of the anchor test was also varied to examine its effect on the three different equating methods. The three equating methods were compared to a criterion equating to obtain estimates of random equating error, bias, and root mean squared error (RMSE). Results showed that, for most studied conditions, chained linear equating produced fairly good equating results in terms of low bias and RMSE. Levine equating also produced low bias and RMSE in some conditions. Although the Tucker method always produced the lowest random equating error, it produced a larger bias and RMSE than either of the other equating methods. As noted in the literature, these results also suggest that either chained linear or Levine equating be used when new and old form samples differ on ability and/or when the anchor-to-total correlation is not very high. Finally, by testing the missing data assumptions of the three equating methods, this study also shows empirically why an equating method is more or less accurate under certain conditions .     

Measurement Invariance Across Groups in Latent Growth Modeling

This Monte Carlo study investigated the impacts of measurement noninvariance across groups on major parameter estimates in latent growth modeling when researchers test group differences in initial status and latent growth. The average initial status and latent growth and the group effects on initial status and latent growth were investigated in terms of Type I error and bias. The location and magnitude of noninvariance across groups was related to the location and magnitude of bias and Type I error in the parameter estimates. That is, noninvariance in factor loadings and intercepts was associated with the Type I error inflation and bias in the parameter estimates of the slope factor (or latent growth) and the intercept factor (or initial status), respectively. As noninvariance became large, the degree of Type I error and bias also increased. On the other hand, a correctly specified second-order latent growth model yielded unbiased parameter estimates and correct statistical inferences. Other findings and implications on future studies were discussed.     

Mean and Mean-and-Variance Corrections With Big Data

Mean and mean-and-variance corrections are the 2 major principles to develop test statistics with violation of conditions. In structural equation modeling (SEM), mean-rescaled and mean-and-variance-adjusted test statistics have been recommended under different contexts. However, recent studies indicated that their Type I error rates vary from 0% to 100% as the number of variables p increases. Can we still trust the 2 principles and what alternative rules can be used to develop test statistics for SEM with “big data”? This article addresses the issues by a large-scale Monte Carlo study. Results indicate that empirical means and standard deviations of each statistic can differ from their expected values many times in standardized units when p is large. Thus, the problems in Type I error control with the 2 statistics are because they do not possess the properties to which they are entitled, not because of the wrongdoing of the mean and mean-and-variance corrections. However, the 2 principles need to be implemented using small sample methodology instead of asymptotics. Results also indicate that distributions other than chi-square might better describe the behavior of test statistics in SEM with big data.     

Estimation of MIMIC Model Parameters with Multilevel Data

The purpose of this simulation study was to assess the performance of latent variable models that take into account the complex sampling mechanism that often underlies data used in educational, psychological, and other social science research. Analyses were conducted using the multiple indicator multiple cause (MIMIC) model, which is a flexible and effective tool for relating observed and latent variables. The data were simulated in a hierarchical framework (e.g., individuals nested in schools) so that a multilevel modeling approach would be appropriate. Analyses were conducted accounting for and not accounting for the nested data to determine the impact of ignoring such multilevel data structures in full structural equation models. Results highlight the differences in modeling results when the analytic strategy is congruent with the data structure and what occurs when this congruency is absent. Type I error rates and power for the standard and multilevel methods were similar for within-cluster variables and for the multilevel model with between-cluster variables. However, Type I error rates were inflated for the standard approach when modeling between-cluster variables.     

Multiplicity Control in Structural Equation Modeling

Abstract

Researchers conducting structural equation modeling analyses rarely, if ever, control for the inflated probability of Type I errors when evaluating the statistical significance of multiple parameters in a model. In this study, the Type I error control, power and true model rates of famsilywise and false discovery rate controlling procedures were compared with rates when no multiplicity control was imposed. The results indicate that Type I error rates become severely inflated with no multiplicity control, but also that familywise error controlling procedures were extremely conservative and had very little power for detecting true relations. False discovery rate controlling procedures provided a compromise between no multiplicity control and strict familywise error control and with large sample sizes provided a high probability of making correct inferences regarding all the parameters in the model.     

Improving the Matching for DIF Analysis by Conditioning on Both Test Score and an Educational Background Variable

When tests are designed to measure dimensionally complex material, DIF analysis with matching based on the total test score may be inappropriate. Previous research has demonstrated that matching can be improved by using multiple internal or both internal and external measures to more completely account for the latent ability space. The present article extends this line of research by examining the potential to improve matching by conditioning simultaneously on test score and a categorical variable representing the educational background of the examinees. The responses of male and female examinees from a test of medical competence were analyzed using a logistic regression procedure. Results show a substantial reduction in the number of items identified as displaying significant DIF when conditioning is based on total test score and a variable representing educational background as opposed to total test score only.