Program: Montana Chapter of ASA 2006 Fall Meeting

Butte-Highlands Room, Student Union Building
Montana Tech, Butte, Montana

September 19, 2006

Time	Topic	Speaker	Affiliation
8:45 - 9:30	*Coffee and Doughnuts*
9:30 - 9:50	Improving General Education Math?	Chip Todd	Montana Tech
9:55 - 10:15	Uncertainty Analysis of Environmental Models	Suzan Gazioglu	Montana Tech
10:20 - 10:40	Ensemble Modeling and Hierarchical Mixtures of Experts: an Application to Watershed Hydrology	Lucy Marshall	Montana State University
10:45 - 11:05	Classifying Proteins from Pull-down Assays Using a False Discovery Rate	Julia Sharp	Montana State University
11:10 - 11:30	Testing Homogeneity of Variances for Nonnormal Models Using Entropy	Solomon Harrar	University of Montana
11:35 - 1:15	*No-host Lunch*
1:15 - 1:35	A Comparison of Full Cross-validation and Cross-validation Based Measures for Linear Regression Model Selection	Mark Greenwood	Montana State University
1:40 - 2:00	New Classes in Life Testing with a Guaranteed Survival Time	Moe Sepehrifar	Montana State University
2:00 - ?	*MT Chapter ASA Business Meeting*	Suzan Gazioglu, President	Minutes

ABSTRACTS

“Improving General Education Math?”
- Based, in part, on recommendations derived from a statistical analysis, we will examine whether math general education course changes have helped the general education program at a university.
“Uncertainty Analysis of Environmental Models”
- Uncertainty in predictions of models arises from, among other sources, the model structure representing the real world system, the scenario reflecting different views of the future and the uncertainty in the moodel input factors. In this study we have looked at scenario uncertainty, model uncertainty and input factor uncertainty within the framework of the three global carbon cycle models. Within a model and a scenario, the prediction uncertainty is partitioned between input factors, using methods like stepwise regression. Prediction uncertainty due to input factors uncertainty has been investigated. The effects of sampling technique, sample size and distributional assumptions for the input factors on model predictions have also been assessed. In the scenario uncertainty, we explored the amount of uncertainty in the predictions due to three emission scenarios both with and without input factor uncertainty included. The results showed that for the more complex (25-compartment) model the uncertainty in the predictions were not determined by the scenario uncertainty, but for the simpler (8-compartment) model, the uncertainty in the predictions was highly influenced by scenario uncertainty. Uncertainty in the predictions due to model was found to be much higher with the 8-compartment models.
  
  In order to reduce the uncertainty in model predictions, we need to know which source of uncertainty is determining the overall uncertainty. First, focusing on uncertainty in the predictions resulting from scenario and input factor uncertainties within each model, we have partitioned the prediction uncertainty in y_Atm(t = 2100) between these two sources of uncertainty. The results of this analysis confirmed what we have found out from the analysis of scenario uncertainty.
“Ensemble Modeling and Hierarchical Mixtures of Experts: an Application to Watershed Hydrology”
- We explore an approach for combining a group of hydrological models that utilizes a “divide and conquer” strategy to allocate models to the available data. The approach, known as Hierarchical Mixtures of Expert (HME), is particularly useful for modeling the dynamics of hydrologic systems where dominant physical processes are known to vary depending on the watershed antecedent conditions. More recently in hydrology, multi-model formulations have become a popular way to try to represent uncertainty in the modeling process and better assess the risk associated with differeent outcomes. However, the complexity of many hydrological models has limited widespread applicaion of multi-model ensembles due to computational constraints. We illustrate the utility of HME for a range of different hydrological models, the associated computational algorithms for specifying the mode, and the predictive performance of the ensemble simulation when specified in a Bayesian framework.
“Classifying Proteins from Pull-down Assays Using a False Discovery Rate”
- Background: One method to infer protein-protein associations in through a “bait-prey pull-down” assay using a protein affinity agent and an LC-MS (liquid chromatography-mass spectrometry)-based protein indentification method. False positive and negative protein indentifications are not uncommon, however, leading to incorrect inferences.
  
  Methods: A pull-down experiment generates a protein association matrix wherein each column represents a sample from one bait protein, each row represents one prey protein and each cell contains a presence/absence association indicator. This method evaluates the presence/absence pattern across a prey protein (row) with a Likelihood Ratio Test (LRT). A Monte Carlo simulation of binomial random variates disqualified the large sample χ² test, therefore, simulated LR test statistic distributions were used to estimate the p-values. A pull-down experiment often involves hundreds of tests so we apply the false discovery rate method to control the false positive rate. Based on the p-value, each prey protein is assigned a category (specific or non-specific) and appraised with respect to the pull-down experiment's goal and design.
  
  Results: The method is illustrated using a pull-down experiment investigating the protein complexes of Shewanella oneidensis MR-1 at the Proteomics Facility of Pacific Northwest National Laboratory. Of the156 prey proteins for the 12 baits in the experiment, 68 were categorized as "Specific" while 88 were categorized as "Non-Specific". A further exploratory follow-up analysis categorized two of the "Non-Specific" prey proteins as "Associated" and remaining 86 as "Non-Associated".
  
  Conclusions: The Monte Carlo simulated LRT p-values objectively reveal specific and ubiquitous prey, as well as potential systematic errors. The example analysis shows the results to be biologically sensible and more realistic than the adhoc screening methods previously utilized. The method presented is informative for screening for protein-protein associations.
“Testing Homogeneity of Variances for Nonnormal Models Using Entropy”
- This talk deals with testing equality of variances of observations in different treatment groups assuming treatment effects are fixed. We derive a test based on Shannon's entropy for a distribution function. Then we will derive the asymptotic expansion for the null distribution of the test statistic under normality and non-normality. We examine the finite sample performance of the statistic in comparison with the modified likelihood ratio criteria by generating data from normal and some non-normal populations. We provide an example to illustrate the application.
“A Comparison of Full Cross-validation and Cross-validation Based Measures for Linear Regression Model Selection”
- Full cross-validation (FCV) based estimates of the Mean Squared Error of Prediction (MSEP) have been recommended as potential model selection criteria (Bunke, Droge and Polzehl, 1999). They have shown FCV can be used to find less variable estimators of MSEP than typical cross-validation (CV) measures. Davies, Neath and Cavanaugh (2005) suggest a model selection criterion based on cross-validation called the Predictive Discrepancy Criterion (PDC) which is an estimate of the Kullback-Leibler discrepancy. It provides improved performance over typical CV-based criteria and the AIC for model selection. The development of a FCV-based analogue of the PDC is described and explored in a simulation study.
“New Classes in Life Testing with a Guaranteed Survival Time”
- Knowing the class to which a life distribution belongs gives us an idea about the aging of the device or system the life distribution represents. These classes play an important role in the study of reliablility theory, survival analysis, maintenance policies, economics, actuarial sciences and many other applied areas. These classes also enable us to compare the aging properties of different systems. The purpose of this talk is to introduce new nonparametric classes of life distributions defined by the concept of inactivity time of a unit with a guaranteed survival time. These classes provide a meaningful framework for modeling and studying lifetime data for most physical systems. We also study the characterizations of these classes along with some preservation results under typical reliabliity operations such as convolution and mixture operations.