Bryan Keller

A Note on Cluster Effects in Latent Class Analysis

This article examines the effects of clustering in latent class analysis. A comprehensive simulation study is conducted, which begins by specifying a true multilevel latent class model with varying within- and between-cluster sample sizes, varying latent class proportions, and varying intraclass correlations. These models are then estimated under the assumption of a single-level latent class model. The outcomes of interest are measures of bias in the Bayesian Information Criterion (BIC) and the entropy R 2 statistic relative to accounting for the multilevel structure of the data. The results indicate that the size of the intraclass correlation as well as between- and within-cluster sizes are the most prominent factors in determining the amount of bias in these outcome measures, with increasing intraclass correlations combined with small between-cluster sizes resulting in increased bias. Bias is particularly noticeable in the BIC. In addition, there is evidence that class separation interacts with the size of the intraclass correlations and cluster sizes in producing bias in these measures.

Probability of Vitamin D Deficiency by Body Weight and Race/Ethnicity

Introduction: While most physicians recognize that vitamin D status varies by skin color because darker skin requires more light to synthesize vitamin D than lighter skin, the importance of body weight to vitamin D status is a newer, less recognized, finding. The purpose of this study was to use nationally representative US data to determine the probability of vitamin D deficiency by body weight and skin color. Methods: Using data for individuals age ≥6 years from the 2001 to 2010 cycles of the US National Health and Nutrition Examination Survey, we calculated the effect of skin color, body weight, and age on vitamin D status. We determined the probability of deficiency within the normal range of body weight for 3 race/ethnicity groups at 3 target levels of 25-hydroxyvitamin D. Results: Darker skin colors and heavier body weights are independently and significantly associated with poorer vitamin D status. We report graphically the probability of vitamin D deficiency by body weight and skin color at vitamin D targets of 20 and 30 ng/mL. Conclusion: The effects of skin color and body weight on vitamin D status are large both statistically and clinically. Knowledge of these effects may facilitate diagnosis of vitamin D deficiency.

Neural Networks for Propensity Score Estimation: Simulation Results and Recommendations

Neural networks have been noted as promising for propensity score estimation because they algorithmically handle nonlinear relationships and interactions. We examine the performance neural networks as compared with main-effects logistic regression for propensity score estimation via simulation study. When the main-effects logistic propensity score model is correctly specified, the two approaches yield almost identical mean square error. When the logistic propensity score model is misspecified due to the addition of quadratic terms and interactions to the data-generating propensity score model, neural networks perform better in terms of bias and mean square error. We link the performance results to balance on observed covariates and demonstrate that our results underscore the importance of checking balance on higher-order covariate terms.

Propensity Score Analysis in R: A Software Review.

In this article, we review four software packages for implementing propensity score analysis in R: Matching, MatchIt, PSAgraphics, and twang. After briefly discussing essential elements for propensity score analysis, we apply each package to a data set from the Early Childhood Longitudinal Study in order to estimate the average effect of elementary school special education services on math achievement in fifth grade. In the context of this real data example, we evaluate documentation and support resources, built-in quantitative and graphical diagnostic features, and methods available for estimating a causal effect. We conclude by making some recommendations aimed at helping researchers decide which package to turn to based upon their familiarity with propensity score methods, programming in R, and the type of analysis being conducted.

An Online Drug Abuse Prevention Program for Adolescent Girls: Posttest and 1-Year Outcomes

Early adolescent girls’ rates of drug use have matched, and in some instances, surpassed boys’ rates. Though girls and boys share risk factors for drug use, girls also have gender-specific risks. Tailored interventions to prevent girls’ drug use are warranted. This study developed and tested a web-based, drug abuse prevention program for adolescent girls. The nationwide sample of 13- and 14-year-old girls (N = 788) was recruited via Facebook ads. Enrolled girls were randomly assigned to the intervention or control condition. All girls completed pretest measures online. Following pretest, intervention girls interacted with the 9-session, gender-specific prevention program online. The program aimed to reduce girls’ drug use and associated risk factors by improving their cognitive and behavioral skills around such areas as coping with stress, managing mood, maintaining a healthy body image, and refusing drug use offers. Girls in both conditions again completed measures at posttest and 1-year follow-up. At posttest, and compared to girls in the control condition, girls who received the intervention smoked fewer cigarettes and reported higher self-esteem, goal setting, media literacy, and self-efficacy. At 1-year follow-up, and compared to girls in the control condition, girls who received the intervention reported engaging in less binge drinking and cigarette smoking; girls assigned to the intervention condition also had higher alcohol, cigarette, and marijuana refusal skills, coping skills, and media literacy and lower rates of peer drug use. This study’s findings support the use of tailored, online drug abuse prevention programming for early adolescent girls.

Weight and skin colour as predictors of vitamin D status: results of an epidemiological investigation using nationally representative data.

OBJECTIVE Current US dietary recommendations for vitamin D vary by age. Recent research suggests that body weight and skin colour are also major determinants of vitamin D status. The objective of the present epidemiological investigation was to clarify the role of age as a predictor of vitamin D status, while accounting for body weight and skin colour, among a nationally representative sample. DESIGN We calculated the mean serum 25-hydroxyvitamin D levels for the US population by age and weight, as well as by weight and race/ethnicity group. Multiple regression analyses were utilized to evaluate age and weight as predictors of vitamin D status: serum 25-hydroxyvitamin D levels with age alone, age and body weight, and age, body weight and their two-way interaction were modelled for the entire sample and each age subgroup. Graphical data were developed using B-spline non-linear regression. SETTING National Health and Nutrition Examination Survey (31 934 unweighted cases). SUBJECTS Individuals aged 1 year and older. RESULTS There were highly significant differences in mean vitamin D status among US residents by weight and skin colour, with those having darker skin colour or higher body weight having worse vitamin D status. Although a significant factor, the impact of age on vitamin D status was notably less than the impact of body weight. CONCLUSIONS Vitamin D status varied predominantly by body weight and skin colour. Recommendations by nutritionists for diet and supplementation needs should take this into account if vitamin D-related health disparities are to be meaningfully reduced across the USA.

Analysis of accelerometer-derived interpersonal spatial proximities: A calibration, simulation, and validation study

ABSTRACT The purpose of this study was to estimate distances from accelerometer-derived Bluetooth signals as a measure of interpersonal spatial proximity. Accelerometer-derived proximity data were collected indoors and outdoors over a 10m range to calibrate simulation models. Proximity data were simulated over 20m (indoor) and 50m (outdoor) ranges. Competing statistical and machine learning models were used to predict simulated distances; the Root-Mean-Square-Error (RMSE) was calculated. Simulation estimates were validated under conditions wherein a single beacon-receiver (SBR) and multiple beacons-receivers (MBR) collected proximity data indoors and outdoors within a ≤10m range. Simulation data showed that a Random Forest (RF) model performed optimally. The validated RF RMSE was ≤2.7 for SBR, and ≥90% of predicted distances were accurately classified as ≤10m. For MBR, ≥67% of predicted distances were accurately classified as ≤10m. Simulation and validation data suggest that distances can be estimated from accelerometer-derived proximity data within a 20m range using a SBR.

Note-taking and science inquiry in an open-ended learning environment

Highlights • Usage of digital notepad is related to performance in science inquiry tasks in OELE.• Both taking and reaccessing notes facilitate science inquiry performance.• Elaborative and reproductive notes’ relationship with success is content dependent.

Abstract: Data Mining Alternatives to Logistic Regression for Propensity Score Estimation: Neural Networks and Support Vector Machines

Data Mining Alternatives to Logistic Regression for Propensity Score Estimation: Neural Networks and Support Vector Machines Bryan S. B. Keller, Jee-Seon Kim, and Peter M. Steiner Department of Educational Psychology, University of Wisconsin-Madison Logistic regression (LR) has traditionally been the most frequently used method for modeling selection in propensity score (PS) analysis. The dominance of a single method is not for a lack of alternatives; rather, there is a perception among practitioners and methodologists that the extant research on alternatives to LR has not yet made a strong enough case for considering an alternative method. There are, however, circumstances under which logistic regression may not perform well. If the response surface is not a hyperplane, the LR selection model requires polynomial and interaction terms. When there are many covariates, the number of terms to consider can be overwhelmingly large. In addition, when the ratio of the number of covariates to the sample size is high, the estimates produced by LR might be unstable. Data mining methods such as the neural networks (NN) and support vector machines (SVM) are designed to deal with high-dimensional data and algorithmically handle nonlinearities in the selection surface, thus avoiding the need for iterative model respecifications. The performance of the data mining methods relative to LR is examined via simulation. Design factors in the study include the PS estimation method (LR with main effects only, NN with 8 hidden nodes, SVM with radial basis), the true selection model (linear and additive logistic vs. not), and the true outcome model (linear and additive vs. not). Performance outcomes include absolute bias; mean squared error; and standard error of the treatment effect estimate and covariate balance on linear, quadratic, and 2-way interaction terms. Our results show that NN uniformly outperforms SVM on all outcomes under all scenarios. NN and LR perform similarly when the LR model is correctly specified. When the LR is misspecified, it is outperformed by NN. In the most extreme case, with quadratic and interaction terms in both the true PS and outcome models, the absolute bias averaged over 1,000 replications was 145% for LR and less than 6% for NN. The results underscore the importance of checking balance on higher order terms. We recommend estimating PSs via NN in conjunction with LR. If NN achieves better balance, either respecify the LR model or use the estimates from the NN. Bryan S. B. Keller is grateful to his SMEP sponsor, David Kaplan. Correspondence concerning this abstract should be addressed to Bryan S. B. Keller, Department of Educational Psychology, University of WisconsinMadison, 859 Education Sciences, 1025 W. Johnson St., Madison, WI 53706. E-mail: bskeller@wisc.edu