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.     

Differential Item Functioning Effect Size From the Multigroup Confirmatory Factor Analysis for a Meta-Analysis: A Simulation Study

This study presents a new approach to synthesizing differential item functioning (DIF) effect size: First, using correlation matrices from each study, we perform a multigroup confirmatory factor analysis (MGCFA) that examines measurement invariance of a test item between two subgroups (i.e., focal and reference groups). Then we synthesize, across the studies, the differences in the estimated factor loadings between the two subgroups, resulting in a meta-analytic summary of the MGCFA effect sizes (MGCFA-ES). The performance of this new approach was examined using a Monte Carlo simulation, where we created 108 conditions by four factors: (1) three levels of item difficulty, (2) four magnitudes of DIF, (3) three levels of sample size, and (4) three types of correlation matrix (tetrachoric, adjusted Pearson, and Pearson). Results indicate that when MGCFA is fitted to tetrachoric correlation matrices, the meta-analytic summary of the MGCFA-ES performed best in terms of bias and mean square error values, 95% confidence interval coverages, empirical standard errors, Type I error rates, and statistical power; and reasonably well with adjusted Pearson correlation matrices. In addition, when tetrachoric correlation matrices are used, a meta-analytic summary of the MGCFA-ES performed well, particularly, under the condition that a high difficulty item with a large DIF was administered to a large sample size. Our result offers an option for synthesizing the magnitude of DIF on a flagged item across studies in practice.     

Using Log-Linear Smoothing to Improve Small-Sample DIF Estimation

This study examined the extent to which log-linear smoothing could improve the accuracy of differential item functioning (DIF) estimates in small samples of examinees. Examinee responses from a certification test were analyzed using White examinees in the reference group and African American examinees in the focal group. Using a simulation approach, separate DIF estimates for seven small-sample-size conditions were obtained using unsmoothed (U) and smoothed (S) score distributions. These small sample U and S DIF estimates were compared to a criterion (i.e., DIF estimates obtained using the unsmoothed total data) to assess their degree of variability (random error) and accuracy (bias). Results indicate that for most studied items smoothing the raw score distributions reduced random error and bias of the DIF estimates, especially in the small-sample-size conditions. Implications of these results for operational testing programs are discussed.     

Effect of Unequal Variances in Proficiency Distributions on Type-I Error of the Mantel-Haenszel Chi-square Test for Differential Item Functioning

Empirical studies demonstrated Type-I error (TIE) inflation (especially for highly discriminating easy items) of the Mantel-Haenszel chi-square test for differential item functioning (DIF), when data conformed to item response theory (IRT) models more complex than Rasch, and when IRT proficiency distributions differed only in means. However, no published study manipulated proficiency variance ratio (VR). Data were generated with the three-parameter logistic (3PL) IRT model. Proficiency VRs were 1, 2, 3, and 4. The present study suggests inflation may be greater, and may affect all highly discriminating items (low, moderate, and high difficulty), when IRT proficiency distributions of reference and focal groups differ also in variances. Inflation was greatest on the 21-item test (vs. 41) and 2,000 total sample size (vs. 1,000). Previous studies had not systematically examined sample size ratio. Sample size ratio of 1:1 produced greater TIE inflation than 3:1, but primarily for total sample size of 2,000.     

Power and Sample Size Calculations for Logistic Regression Tests for Differential Item Functioning

Logistic regression is a popular method for detecting uniform and nonuniform differential item functioning (DIF) effects. Theoretical formulas for the power and sample size calculations are derived for likelihood ratio tests and Wald tests based on the asymptotic distribution of the maximum likelihood estimators for the logistic regression model. The power is related to the item response function (IRF) for the studied item, the latent trait distributions, and the sample sizes for the reference and focal groups. Simulation studies show that the theoretical values calculated from the formulas derived in the article are close to what are observed in the simulated data when the assumptions are satisfied. The robustness of the power formulas are studied with simulations when the assumptions are violated.     

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.     

Using Logistic Regression and the Mantel-Haenszel With Multiple Ability Estimates to Detect Differential Item Functioning

Logistic regression has recently been advanced as a viable procedure for detecting differential item functioning (DIF). One of the advantages of this procedure is the considerable flexibility it offers in the specification of the regression equation. This article describes incorporating two ability estimates into a single regression analysis, with the result that substantially fewer items exhibit DIF. A comparable analysis is conducted using the Mantel-Haenszel with similar results. It is argued that by simultaneously conditioning on two relevant ability estimates, more accurate matching of examinees in the reference and focal groups is obtained, and thus multidimensional item impact is not mistakenly identified as DIF.     

DIF Detection and Interpretation in Large-Scale Science Assessments: Informing Item Writing Practices

Differential item functioning (DIF) analyses are a routine part of the development of large-scale assessments. Less common are studies to understand the potential sources of DIF. The goals of this study were (a) to identify gender DIF in a large-scale science assessment and (b) to look for trends in the DIF and non-DIF items due to content, cognitive demands, item type, item text, and visual-spatial or reference factors. To facilitate the analyses, DIF studies were conducted at 3 grade levels and for 2 randomly equivalent forms of the science assessment at each grade level (administered in different years). The DIF procedure itself was a variant of the "standardization procedure" of Dorans and Kulick (1986) and was applied to very large sets of data (6 sets of data, each involving 60,000 students). It has the advantages of being easy to understand and to explain to practitioners. Several findings emerged from the study that would be useful to pass on to test development committees. For example, when there was DIF in science items, MC items tended to favor male examinees and OR items tended to favor female examinees. Compiling DIF information across multiple grades and years increases the likelihood that important trends in the data will be identified and that item writing practices will be informed by more than anecdotal reports about DIF.     

DIF Detection and Effect Size Measures for Polytomously Scored Items

Data from a large-scale performance assessment ( N = 105,731) were analyzed with five differential item functioning (DIF) detection methods for polytomous items to examine the congruence among the DIF detection methods. Two different versions of the item response theory (IRT) model-based likelihood ratio test, the logistic regression likelihood ratio test, the Mantel test, and the generalized Mantel–Haenszel test were compared. Results indicated some agreement among the five DIF detection methods. Because statistical power is a function of the sample size, the DIF detection results from extremely large data sets are not practically useful. As alternatives to the DIF detection methods, four IRT model-based indices of standardized impact and four observed-score indices of standardized impact for polytomous items were obtained and compared with the R ² measures of logistic regression.     

第二语言学习者专业背景对HSK阅读成绩影响的项目功能差异检验

本文旨在考察HSK应试者的专业背景是否会对他们的阅读成绩产生影响。运用MH方法和SIBTEST方法对2009年HSK(初中等)考试阅读题目进行DIF筛查,把专业背景为自然科学的HSK考生设为目标组,专业背景为人文社会科学的HSK考生设为参照组。MH方法的结果是没有找到含有DIF的题目;SIBTEST方法的结果如下:第一轮DIF筛查检测到一个题目,第二轮DBF筛查检测到一组题目。这组题目有利于人文社会学科专业背景的被试。就检测DIF的方法而言,本研究认为SIBTEST方法更加敏感,DBF检验更加适合像阅读理解测验这样的一组或多组相互关联的题目。     

Differential Item Functioning for Accommodated Students with Disabilities: Effect of Differences in Proficiency Distributions

ABSTRACT

This study examined the effect of similar vs. dissimilar proficiency distributions on uniform DIF detection on a statewide eighth grade mathematics assessment. Results from the similar- and dissimilar-ability reference groups with an SWD focal group were compared for four models: logistic regression, hierarchical generalized linear model (HGLM), the Wald-1 IRT-based test, and the Mantel-Haenszel procedure. A DIF-free-then-DIF strategy was used. The rate of DIF detection was examined among all accommodated scores and common accommodation subcategories. No items were detected for DIF using the similar ability distribution reference group, regardless of method. With the dissimilar ability reference group, logistic regression and Mantel–Haenszel flagged 8–17%, and the Wald-1 and HGLM test flagged 23–38% of items for DIF. Forming focal groups by accommodation type did not alter the pattern of DIF detection. Creating a reference group to be similar in ability to the focal group may control the rate of erroneous DIF detection for SWD.     

Routing Strategies and Optimizing Design for Multistage Testing in International Large‐Scale Assessments

This study investigates the effect of several design and administration choices on item exposure and person/item parameter recovery under a multistage test (MST) design. In a simulation study, we examine whether number‐correct (NC) or item response theory (IRT) methods are differentially effective at routing students to the correct next stage(s) and whether routing choices (optimal versus suboptimal routing) have an impact on achievement precision. Additionally, we examine the impact of testlet length on both person and item recovery. Overall, our results suggest that no single approach works best across the studied conditions. With respect to the mean person parameter recovery, IRT scoring (via either Fisher information or preliminary EAP estimates) outperformed classical NC methods, although differences in bias and root mean squared error were generally small. Item exposure rates were found to be more evenly distributed when suboptimal routing methods were used, and item recovery (both difficulty and discrimination) was most precisely observed for items with moderate difficulties. Based on the results of the simulation study, we draw conclusions and discuss implications for practice in the context of international large‐scale assessments that recently introduced adaptive assessment in the form of MST. Future research directions are also discussed.     

Applying the Mantel-Haenszel Procedure to Complex Samples of Items

This Monte Carlo study examined the effect of complex sampling of items on the measurement of differential item functioning (DIF) using the Mantel-Haenszel procedure. Data were generated using a 3-parameter logistic item response theory model according to the balanced incomplete block (BIB) design used in the National Assessment of Educational Progress (NAEP). The length of each block of items and the number of DIF items in the matching variable were varied, as was the difficulty, discrimination, and presence of DIF in the studied item. Block, booklet, pooled booklet, and extra-information analyses were compared to a complete data analysis using the transformed log-odds on the delta scale. The pooled booklet approach is recommended for use when items are selected for examinees according to a BIB design. This study has implications for DIF analyses of other complex samples of items, such as computer administered testing or another complex assessment design.     

Effects of Population Heterogeneity on Accuracy of DIF Detection

Heterogeneity within English language learners (ELLs) groups has been documented. Previous research on differential item functioning (DIF) analyses suggests that accurate DIF detection rates are reduced greatly when groups are heterogeneous. In this simulation study, we investigated the effects of heterogeneity within linguistic (ELL) groups on the accuracy of DIF detection. Heterogeneity within such groups may occur for a myriad of reasons including differential lengths of time residing in English-speaking countries, degrees of exposure to English-speaking environments, and amounts of English instruction. Our findings revealed that at high levels of within-group heterogeneity, DIF detection is at the level of chance, implying that a large proportion of DIF items might remain undetected when assessing heterogeneous populations potentially leading to developing biased tests. Based on our findings, we urge test development organizations to consider heterogeneity within ELL and other heterogeneous focus groups in their routine DIF analyses.     

Lord's Wald Test for Detecting DIF in Multidimensional IRT Models: A Comparison of Two Estimation Approaches

Lord's Wald test for differential item functioning (DIF) has not been studied extensively in the context of the multidimensional item response theory (MIRT) framework. In this article, Lord's Wald test was implemented using two estimation approaches, marginal maximum likelihood estimation and Bayesian Markov chain Monte Carlo estimation, to detect uniform and nonuniform DIF under MIRT models. The Type I error and power rates for Lord's Wald test were investigated under various simulation conditions, including different DIF types and magnitudes, different means and correlations of two ability parameters, and different sample sizes. Furthermore, English usage data were analyzed to illustrate the use of Lord's Wald test with the two estimation approaches.     

Differential Item Functioning Assessment in Cognitive Diagnostic Modeling: Application of the Wald Test to Investigate DIF in the DINA Model

Analyzing examinees’ responses using cognitive diagnostic models (CDMs) has the advantage of providing diagnostic information. To ensure the validity of the results from these models, differential item functioning (DIF) in CDMs needs to be investigated. In this article, the Wald test is proposed to examine DIF in the context of CDMs. This study explored the effectiveness of the Wald test in detecting both uniform and nonuniform DIF in the DINA model through a simulation study. Results of this study suggest that for relatively discriminating items, the Wald test had Type I error rates close to the nominal level. Moreover, its viability was underscored by the medium to high power rates for most investigated DIF types when DIF size was large. Furthermore, the performance of the Wald test in detecting uniform DIF was compared to that of the traditional Mantel‐Haenszel (MH) and SIBTEST procedures. The results of the comparison study showed that the Wald test was comparable to or outperformed the MH and SIBTEST procedures. Finally, the strengths and limitations of the proposed method and suggestions for future studies are discussed.     

An Empirical Investigation Demonstrating the Multidimensional DIF Paradigm: A Cognitive Explanation for DIF

Differential Item Functioning (DIF) is traditionally used to identify different item performance patterns between intact groups, most commonly involving race or sex comparisons. This study advocates expanding the utility of DIF as a step in construct validation. Rather than grouping examinees based on cultural differences, the reference and focal groups are chosen from two extremes along a distinct cognitive dimension that is hypothesized to supplement the dominant latent trait being measured. Specifically, this study investigates DIF between proficient and non-proficient fourth- and seventh-grade writers on open-ended mathematics test items that require students to communicate about mathematics. It is suggested that the occurrence of DIF in this situation actually enhances, rather than detracts from, the construct validity of the test because, according to the National Council of Teachers of Mathematics (NCTM), mathematical communication is an important component of mathematical ability, the dominant construct being assessed. However, the presence of DIF influences the validity of inferences that can be made from test scores and suggests that two scores should be reported, one for general mathematical ability and one for mathematical communication. The fact that currently only one test score is reported, a simple composite of scores on multiple-choice and open-ended items, may lead to incorrect decisions being made about examinees.     

Influence of Prior Distributions on Detection of DIF

Detection of differential item functioning (DIF) on items intentionally constructed to favor one group over another was investigated on item parameter estimates obtained from two item response theory-based computer programs, LOGIST and BILOG. Signed- and unsigned-area measures based on joint maximum likelihood estimation, marginal maximum likelihood estimation, and two marginal maximum a posteriori estimation procedures were compared with each other to determine whether detection of DIF could be improved using prior distributions. Results indicated that item parameter estimates obtained using either prior condition were less deviant than when priors were not used. Differences in detection of DIF appeared to be related to item parameter estimation condition and to some extent to sample size.     

Comparing Methods of Assessing Differential Item Functioning in a Computerized Adaptive Testing Environment

Mantel-Haenszel and SIBTEST, which have known difficulty in detecting non-unidirectional differential item functioning (DIF), have been adapted with some success for computerized adaptive testing (CAT). This study adapts logistic regression (LR) and the item-response-theory-likelihood-ratio test (IRT-LRT), capable of detecting both unidirectional and non-unidirectional DIF, to the CAT environment in which pretest items are assumed to be seeded in CATs but not used for trait estimation. The proposed adaptation methods were evaluated with simulated data under different sample size ratios and impact conditions in terms of Type I error, power, and specificity in identifying the form of DIF. The adapted LR and IRT-LRT procedures are more powerful than the CAT version of SIBTEST for non-unidirectional DIF detection. The good Type I error control provided by IRT-LRT under extremely unequal sample sizes and large impact is encouraging. Implications of these and other findings are discussed.     

Simulated Tests of Differential Item Functioning Using SIBTEST With and Without Impact

Monte Carlo simulations with 20,000 replications are reported to estimate the probability of rejecting the null hypothesis regarding DIF using SIBTEST when there is DIF present and/or when impact is present due to differences on the primary dimension to be measured. Sample sizes are varied from 250 to 2000 and test lengths from 10 to 40 items. Results generally support previous findings for Type I error rates and power. Impact is inversely related to test length. The combination of DIF and impact, with the focal group having lower ability on both the primary and secondary dimensions, results in impact partially masking DIF so that items biased toward the reference group are less likely to be detected.