Matching Items (23)
Filtering by
- All Subjects: Statistics
- All Subjects: Item response theory
- Creators: Levy, Roy
Description
This work presents two complementary studies that propose heuristic methods to capture characteristics of data using the ensemble learning method of random forest. The first study is motivated by the problem in education of determining teacher effectiveness in student achievement. Value-added models (VAMs), constructed as linear mixed models, use students’ test scores as outcome variables and teachers’ contributions as random effects to ascribe changes in student performance to the teachers who have taught them. The VAMs teacher score is the empirical best linear unbiased predictor (EBLUP). This approach is limited by the adequacy of the assumed model specification with respect to the unknown underlying model. In that regard, this study proposes alternative ways to rank teacher effects that are not dependent on a given model by introducing two variable importance measures (VIMs), the node-proportion and the covariate-proportion. These VIMs are novel because they take into account the final configuration of the terminal nodes in the constitutive trees in a random forest. In a simulation study, under a variety of conditions, true rankings of teacher effects are compared with estimated rankings obtained using three sources: the newly proposed VIMs, existing VIMs, and EBLUPs from the assumed linear model specification. The newly proposed VIMs outperform all others in various scenarios where the model was misspecified. The second study develops two novel interaction measures. These measures could be used within but are not restricted to the VAM framework. The distribution-based measure is constructed to identify interactions in a general setting where a model specification is not assumed in advance. In turn, the mean-based measure is built to estimate interactions when the model specification is assumed to be linear. Both measures are unique in their construction; they take into account not only the outcome values, but also the internal structure of the trees in a random forest. In a separate simulation study, under a variety of conditions, the proposed measures are found to identify and estimate second-order interactions.
ContributorsValdivia, Arturo (Author) / Eubank, Randall (Thesis advisor) / Young, Dennis (Committee member) / Reiser, Mark R. (Committee member) / Kao, Ming-Hung (Committee member) / Broatch, Jennifer (Committee member) / Arizona State University (Publisher)
Created2013
Description
Dimensionality assessment is an important component of evaluating item response data. Existing approaches to evaluating common assumptions of unidimensionality, such as DIMTEST (Nandakumar & Stout, 1993; Stout, 1987; Stout, Froelich, & Gao, 2001), have been shown to work well under large-scale assessment conditions (e.g., large sample sizes and item pools; see e.g., Froelich & Habing, 2007). It remains to be seen how such procedures perform in the context of small-scale assessments characterized by relatively small sample sizes and/or short tests. The fact that some procedures come with minimum allowable values for characteristics of the data, such as the number of items, may even render them unusable for some small-scale assessments. Other measures designed to assess dimensionality do not come with such limitations and, as such, may perform better under conditions that do not lend themselves to evaluation via statistics that rely on asymptotic theory. The current work aimed to evaluate the performance of one such metric, the standardized generalized dimensionality discrepancy measure (SGDDM; Levy & Svetina, 2011; Levy, Xu, Yel, & Svetina, 2012), under both large- and small-scale testing conditions. A Monte Carlo study was conducted to compare the performance of DIMTEST and the SGDDM statistic in terms of evaluating assumptions of unidimensionality in item response data under a variety of conditions, with an emphasis on the examination of these procedures in small-scale assessments. Similar to previous research, increases in either test length or sample size resulted in increased power. The DIMTEST procedure appeared to be a conservative test of the null hypothesis of unidimensionality. The SGDDM statistic exhibited rejection rates near the nominal rate of .05 under unidimensional conditions, though the reliability of these results may have been less than optimal due to high sampling variability resulting from a relatively limited number of replications. Power values were at or near 1.0 for many of the multidimensional conditions. It was only when the sample size was reduced to N = 100 that the two approaches diverged in performance. Results suggested that both procedures may be appropriate for sample sizes as low as N = 250 and tests as short as J = 12 (SGDDM) or J = 19 (DIMTEST). When used as a diagnostic tool, SGDDM may be appropriate with as few as N = 100 cases combined with J = 12 items. The study was somewhat limited in that it did not include any complex factorial designs, nor were the strength of item discrimination parameters or correlation between factors manipulated. It is recommended that further research be conducted with the inclusion of these factors, as well as an increase in the number of replications when using the SGDDM procedure.
ContributorsReichenberg, Ray E (Author) / Levy, Roy (Thesis advisor) / Thompson, Marilyn S. (Thesis advisor) / Green, Samuel B. (Committee member) / Arizona State University (Publisher)
Created2013
Description
Parallel Monte Carlo applications require the pseudorandom numbers used on each processor to be independent in a probabilistic sense. The TestU01 software package is the standard testing suite for detecting stream dependence and other properties that make certain pseudorandom generators ineffective in parallel (as well as serial) settings. TestU01 employs two basic schemes for testing parallel generated streams. The first applies serial tests to the individual streams and then tests the resulting P-values for uniformity. The second turns all the parallel generated streams into one long vector and then applies serial tests to the resulting concatenated stream. Various forms of stream dependence can be missed by each approach because neither one fully addresses the multivariate nature of the accumulated data when generators are run in parallel. This dissertation identifies these potential faults in the parallel testing methodologies of TestU01 and investigates two different methods to better detect inter-stream dependencies: correlation motivated multivariate tests and vector time series based tests. These methods have been implemented in an extension to TestU01 built in C++ and the unique aspects of this extension are discussed. A variety of different generation scenarios are then examined using the TestU01 suite in concert with the extension. This enhanced software package is found to better detect certain forms of inter-stream dependencies than the original TestU01 suites of tests.
ContributorsIsmay, Chester (Author) / Eubank, Randall (Thesis advisor) / Young, Dennis (Committee member) / Kao, Ming-Hung (Committee member) / Lanchier, Nicolas (Committee member) / Reiser, Mark R. (Committee member) / Arizona State University (Publisher)
Created2013
Description
This simulation study compared the utility of various discrepancy measures within a posterior predictive model checking (PPMC) framework for detecting different types of data-model misfit in multidimensional Bayesian network (BN) models. The investigated conditions were motivated by an applied research program utilizing an operational complex performance assessment within a digital-simulation educational context grounded in theories of cognition and learning. BN models were manipulated along two factors: latent variable dependency structure and number of latent classes. Distributions of posterior predicted p-values (PPP-values) served as the primary outcome measure and were summarized in graphical presentations, by median values across replications, and by proportions of replications in which the PPP-values were extreme. An effect size measure for PPMC was introduced as a supplemental numerical summary to the PPP-value. Consistent with previous PPMC research, all investigated fit functions tended to perform conservatively, but Standardized Generalized Dimensionality Discrepancy Measure (SGDDM), Yen's Q3, and Hierarchy Consistency Index (HCI) only mildly so. Adequate power to detect at least some types of misfit was demonstrated by SGDDM, Q3, HCI, Item Consistency Index (ICI), and to a lesser extent Deviance, while proportion correct (PC), a chi-square-type item-fit measure, Ranked Probability Score (RPS), and Good's Logarithmic Scale (GLS) were powerless across all investigated factors. Bivariate SGDDM and Q3 were found to provide powerful and detailed feedback for all investigated types of misfit.
ContributorsCrawford, Aaron (Author) / Levy, Roy (Thesis advisor) / Green, Samuel (Committee member) / Thompson, Marilyn (Committee member) / Arizona State University (Publisher)
Created2014
Description
Missing data are common in psychology research and can lead to bias and reduced power if not properly handled. Multiple imputation is a state-of-the-art missing data method recommended by methodologists. Multiple imputation methods can generally be divided into two broad categories: joint model (JM) imputation and fully conditional specification (FCS) imputation. JM draws missing values simultaneously for all incomplete variables using a multivariate distribution (e.g., multivariate normal). FCS, on the other hand, imputes variables one at a time, drawing missing values from a series of univariate distributions. In the single-level context, these two approaches have been shown to be equivalent with multivariate normal data. However, less is known about the similarities and differences of these two approaches with multilevel data, and the methodological literature provides no insight into the situations under which the approaches would produce identical results. This document examined five multilevel multiple imputation approaches (three JM methods and two FCS methods) that have been proposed in the literature. An analytic section shows that only two of the methods (one JM method and one FCS method) used imputation models equivalent to a two-level joint population model that contained random intercepts and different associations across levels. The other three methods employed imputation models that differed from the population model primarily in their ability to preserve distinct level-1 and level-2 covariances. I verified the analytic work with computer simulations, and the simulation results also showed that imputation models that failed to preserve level-specific covariances produced biased estimates. The studies also highlighted conditions that exacerbated the amount of bias produced (e.g., bias was greater for conditions with small cluster sizes). The analytic work and simulations lead to a number of practical recommendations for researchers.
ContributorsMistler, Stephen (Author) / Enders, Craig K. (Thesis advisor) / Aiken, Leona (Committee member) / Levy, Roy (Committee member) / West, Stephen G. (Committee member) / Arizona State University (Publisher)
Created2015
Description
Many methodological approaches have been utilized to predict student retention and persistence over the years, yet few have utilized a Bayesian framework. It is believed this is due in part to the absence of an established process for guiding educational researchers reared in a frequentist perspective into the realms of Bayesian analysis and educational data mining. The current study aimed to address this by providing a model-building process for developing a Bayesian network (BN) that leveraged educational data mining, Bayesian analysis, and traditional iterative model-building techniques in order to predict whether community college students will stop out at the completion of each of their first six terms. The study utilized exploratory and confirmatory techniques to reduce an initial pool of more than 50 potential predictor variables to a parsimonious final BN with only four predictor variables. The average in-sample classification accuracy rate for the model was 80% (Cohen's κ = 53%). The model was shown to be generalizable across samples with an average out-of-sample classification accuracy rate of 78% (Cohen's κ = 49%). The classification rates for the BN were also found to be superior to the classification rates produced by an analog frequentist discrete-time survival analysis model.
ContributorsArcuria, Philip (Author) / Levy, Roy (Thesis advisor) / Green, Samuel B (Committee member) / Thompson, Marilyn S (Committee member) / Arizona State University (Publisher)
Created2015
Description
It is common in the analysis of data to provide a goodness-of-fit test to assess the performance of a model. In the analysis of contingency tables, goodness-of-fit statistics are frequently employed when modeling social science, educational or psychological data where the interest is often directed at investigating the association among multi-categorical variables. Pearson's chi-squared statistic is well-known in goodness-of-fit testing, but it is sometimes considered to produce an omnibus test as it gives little guidance to the source of poor fit once the null hypothesis is rejected. However, its components can provide powerful directional tests. In this dissertation, orthogonal components are used to develop goodness-of-fit tests for models fit to the counts obtained from the cross-classification of multi-category dependent variables. Ordinal categories are assumed. Orthogonal components defined on marginals are obtained when analyzing multi-dimensional contingency tables through the use of the QR decomposition. A subset of these orthogonal components can be used to construct limited-information tests that allow one to identify the source of lack-of-fit and provide an increase in power compared to Pearson's test. These tests can address the adverse effects presented when data are sparse. The tests rely on the set of first- and second-order marginals jointly, the set of second-order marginals only, and the random forest method, a popular algorithm for modeling large complex data sets. The performance of these tests is compared to the likelihood ratio test as well as to tests based on orthogonal polynomial components. The derived goodness-of-fit tests are evaluated with studies for detecting two- and three-way associations that are not accounted for by a categorical variable factor model with a single latent variable. In addition the tests are used to investigate the case when the model misspecification involves parameter constraints for large and sparse contingency tables. The methodology proposed here is applied to data from the 38th round of the State Survey conducted by the Institute for Public Policy and Michigan State University Social Research (2005) . The results illustrate the use of the proposed techniques in the context of a sparse data set.
ContributorsMilovanovic, Jelena (Author) / Young, Dennis (Thesis advisor) / Reiser, Mark R. (Thesis advisor) / Wilson, Jeffrey (Committee member) / Eubank, Randall (Committee member) / Yang, Yan (Committee member) / Arizona State University (Publisher)
Created2011
Description
The purpose of this study was to investigate the effect of complex structure on dimensionality assessment in compensatory and noncompensatory multidimensional item response models (MIRT) of assessment data using dimensionality assessment procedures based on conditional covariances (i.e., DETECT) and a factor analytical approach (i.e., NOHARM). The DETECT-based methods typically outperformed the NOHARM-based methods in both two- (2D) and three-dimensional (3D) compensatory MIRT conditions. The DETECT-based methods yielded high proportion correct, especially when correlations were .60 or smaller, data exhibited 30% or less complexity, and larger sample size. As the complexity increased and the sample size decreased, the performance typically diminished. As the complexity increased, it also became more difficult to label the resulting sets of items from DETECT in terms of the dimensions. DETECT was consistent in classification of simple items, but less consistent in classification of complex items. Out of the three NOHARM-based methods, χ2G/D and ALR generally outperformed RMSR. χ2G/D was more accurate when N = 500 and complexity levels were 30% or lower. As the number of items increased, ALR performance improved at correlation of .60 and 30% or less complexity. When the data followed a noncompensatory MIRT model, the NOHARM-based methods, specifically χ2G/D and ALR, were the most accurate of all five methods. The marginal proportions for labeling sets of items as dimension-like were typically low, suggesting that the methods generally failed to label two (three) sets of items as dimension-like in 2D (3D) noncompensatory situations. The DETECT-based methods were more consistent in classifying simple items across complexity levels, sample sizes, and correlations. However, as complexity and correlation levels increased the classification rates for all methods decreased. In most conditions, the DETECT-based methods classified complex items equally or more consistent than the NOHARM-based methods. In particular, as complexity, the number of items, and the true dimensionality increased, the DETECT-based methods were notably more consistent than any NOHARM-based method. Despite DETECT's consistency, when data follow a noncompensatory MIRT model, the NOHARM-based method should be preferred over the DETECT-based methods to assess dimensionality due to poor performance of DETECT in identifying the true dimensionality.
ContributorsSvetina, Dubravka (Author) / Levy, Roy (Thesis advisor) / Gorin, Joanna S. (Committee member) / Millsap, Roger (Committee member) / Arizona State University (Publisher)
Created2011
Description
ABSTRACT This study investigated the possibility of item parameter drift (IPD) in a calculus placement examination administered to approximately 3,000 students at a large university in the United States. A single form of the exam was administered continuously for a period of two years, possibly allowing later examinees to have prior knowledge of specific items on the exam. An analysis of IPD was conducted to explore evidence of possible item exposure. Two assumptions concerning items exposure were made: 1) item recall and item exposure are positively correlated, and 2) item exposure results in the items becoming easier over time. Special consideration was given to two contextual item characteristics: 1) item location within the test, specifically items at the beginning and end of the exam, and 2) the use of an associated diagram. The hypotheses stated that these item characteristics would make the items easier to recall and, therefore, more likely to be exposed, resulting in item drift. BILOG-MG 3 was used to calibrate the items and assess for IPD. No evidence was found to support the hypotheses that the items located at the beginning of the test or with an associated diagram drifted as a result of item exposure. Three items among the last ten on the exam drifted significantly and became easier, consistent with item exposure. However, in this study, the possible effects of item exposure could not be separated from the effects of other potential factors such as speededness, curriculum changes, better test preparation on the part of subsequent examinees, or guessing.
ContributorsKrause, Janet (Author) / Levy, Roy (Thesis advisor) / Thompson, Marilyn (Thesis advisor) / Gorin, Joanna (Committee member) / Arizona State University (Publisher)
Created2012
Description
The existing minima for sample size and test length recommendations for DIMTEST (750 examinees and 25 items) are tied to features of the procedure that are no longer in use. The current version of DIMTEST uses a bootstrapping procedure to remove bias from the test statistic and is packaged with a conditional covariance-based procedure called ATFIND for partitioning test items. Key factors such as sample size, test length, test structure, the correlation between dimensions, and strength of dependence were manipulated in a Monte Carlo study to assess the effectiveness of the current version of DIMTEST with fewer examinees and items. In addition, the DETECT program was also used to partition test items; a second feature of this study also compared the structure of test partitions obtained with ATFIND and DETECT in a number of ways. With some exceptions, the performance of DIMTEST was quite conservative in unidimensional conditions. The performance of DIMTEST in multidimensional conditions depended on each of the manipulated factors, and did suggest that the minima of sample size and test length can be made lower for some conditions. In terms of partitioning test items in unidimensional conditions, DETECT tended to produce longer assessment subtests than ATFIND in turn yielding different test partitions. In multidimensional conditions, test partitions became more similar and were more accurate with increased sample size, for factorially simple data, greater strength of dependence, and a decreased correlation between dimensions. Recommendations for sample size and test length minima are provided along with suggestions for future research.
ContributorsFay, Derek (Author) / Levy, Roy (Thesis advisor) / Green, Samuel (Committee member) / Gorin, Joanna (Committee member) / Arizona State University (Publisher)
Created2012