Though the likelihood is a useful tool for obtaining estimates of regression parameters, it is not readily available in the fit of hierarchical binary data models. The correlated observations negate the opportunity to have a joint likelihood when fitting hierarchical logistic regression models. Through conditional likelihood, inferences for the regression and covariance parameters as well as the intraclass correlation coefficients are usually obtained. In those cases, I have resorted to use of Laplace approximation and large sample theory approach for point and interval estimates such as Wald-type confidence intervals and profile likelihood confidence intervals. These methods rely on distributional assumptions and large sample theory. However, when dealing with small hierarchical datasets they often result in severe bias or non-convergence. I present a generalized quasi-likelihood approach and a generalized method of moments approach; both do not rely on any distributional assumptions but only moments of response. As an alternative to the typical large sample theory approach, I present bootstrapping hierarchical logistic regression models which provides more accurate interval estimates for small binary hierarchical data. These models substitute computations as an alternative to the traditional Wald-type and profile likelihood confidence intervals. I use a latent variable approach with a new split bootstrap method for estimating intraclass correlation coefficients when analyzing binary data obtained from a three-level hierarchical structure. It is especially useful with small sample size and easily expanded to multilevel. Comparisons are made to existing approaches through both theoretical justification and simulation studies. Further, I demonstrate my findings through an analysis of three numerical examples, one based on cancer in remission data, one related to the China’s antibiotic abuse study, and a third related to teacher effectiveness in schools from a state of southwest US.

Responsible test use requires validation \u2014 the process of collecting evidence to support the inferences drawn from test scores. In high-stakes testing contexts, the need for validation is especially great; the far-reaching nature of high-stakes testing affects the educational, professional, and financial futures of stakeholders. The Standards for Educational and Psychological Measurement (AERA et al., 2014) offers specific guidance in developing and implementing tests. Still, concerns exist over the extent to which test developers and users of high-stakes tests are making valid inferences from test scores. This paper explores the current state of high-stakes educational testing and the validity issues surrounding it. Drawing on measurement theory literature, educational literature, and professional standards of test development and use, I assess the significance of these concerns and their potential implications for the stakeholders of high-stakes testing programs.

Problems related to alcohol consumption cause not only extra economic expenses, but are an expense to the health of both drinkers and non-drinkers due to the harm directly and indirectly caused by alcohol consumption. Investigating predictors and reasons for alcohol-related problems is of importance, as alcohol-related problems could be prevented by quitting or limiting consumption of alcohol. We were interested in predicting alcohol-related problems using multiple linear regression and regression trees, and then comparing the regressions to the tree. Impaired control, anxiety sensitivity, mother permissiveness, father permissiveness, gender, and age were included as predictors. The data used was comprised of participants (n=835) sampled from students at Arizona State University. A multiple linear regression without interactions, multiple linear regression with two-way interactions and squares, and a regression tree were used and compared. The regression and the tree had similar results. Multiple interactions of variables predicted alcohol-related problems. Overall, the tree was easier to interpret than the regressions, however, the regressions provided specific predicted alcohol-related problems scores, whereas the tree formed large groups and had a predicted alcohol-related problems score for each group. Nevertheless, the tree still predicted alcohol-related problems nearly as well, if not better than the regressions.

This article proposes a new information-based subdata selection (IBOSS) algorithm, Squared Scaled Distance Algorithm (SSDA). It is based on the invariance of the determinant of the information matrix under orthogonal transformations, especially rotations. Extensive simulation results show that the new IBOSS algorithm retains nice asymptotic properties of IBOSS and gives a larger determinant of the subdata information matrix. It has the same order of time complexity as the D-optimal IBOSS algorithm. However, it exploits the advantages of vectorized calculation avoiding for loops and is approximately 6 times as fast as the D-optimal IBOSS algorithm in R. The robustness of SSDA is studied from three aspects: nonorthogonality, including interaction terms and variable misspecification. A new accurate variable selection algorithm is proposed to help the implementation of IBOSS algorithms when a large number of variables are present with sparse important variables among them. Aggregating random subsample results, this variable selection algorithm is much more accurate than the LASSO method using full data. Since the time complexity is associated with the number of variables only, it is also very computationally efficient if the number of variables is fixed as n increases and not massively large. More importantly, using subsamples it solves the problem that full data cannot be stored in the memory when a data set is too large.

The Pearson and likelihood ratio statistics are commonly used to test goodness-of-fit for models applied to data from a multinomial distribution. When data are from a table formed by cross-classification of a large number of variables, the common statistics may have low power and inaccurate Type I error level due to sparseness in the cells of the table. The GFfit statistic can be used to examine model fit in subtables. It is proposed to assess model fit by using a new version of GFfit statistic based on orthogonal components of Pearson chi-square as a diagnostic to examine the fit on two-way subtables. However, due to variables with a large number of categories and small sample size, even the GFfit statistic may have low power and inaccurate Type I error level due to sparseness in the two-way subtable. In this dissertation, the theoretical power and empirical power of the GFfit statistic are studied. A method based on subsets of orthogonal components for the GFfit statistic on the subtables is developed to improve the performance of the GFfit statistic. Simulation results for power and type I error rate for several different cases along with comparisons to other diagnostics are presented.

Statistical mediation analysis allows researchers to identify the most important the mediating constructs in the causal process studied. Information about the mediating processes can be used to make interventions more powerful by enhancing successful program components and by not implementing components that did not significantly change the outcome. Identifying mediators is especially relevant when the hypothesized mediating construct consists of multiple related facets. The general definition of the construct and its facets might relate differently to external criteria. However, current methods do not allow researchers to study the relationships between general and specific aspects of a construct to an external criterion simultaneously. This study proposes a bifactor measurement model for the mediating construct as a way to represent the general aspect and specific facets of a construct simultaneously. Monte Carlo simulation results are presented to help to determine under what conditions researchers can detect the mediated effect when one of the facets of the mediating construct is the true mediator, but the mediator is treated as unidimensional. Results indicate that parameter bias and detection of the mediated effect depends on the facet variance represented in the mediation model. This study contributes to the largely unexplored area of measurement issues in statistical mediation analysis.