Thursday, Aug 7: 8:30 AM - 10:20 AM
4206
Contributed Papers
Music City Center
Room: CC-105B
Main Sponsor
Survey Research Methods Section
Presentations
Missing data issues are highly prevalent in High Throughput Studies (HTS). Missing patterns in such studies are rarely missing at random. We describe varying percentages of missingness and quantify the amount of missingness in a clinical study. Acute Myeloid Leukemia (AML) is a type of cancer of the myeloid line of blood cells in the bone marrow and blood. This is one of most lethal cancer types. We have gene expression data on AML for thousands of genes. There are three different subtypes of AML (Normal, CK type, and CBF type) that we plan to compare. Usually, gene expression data have many genes with zero counts. Missing value imputation methods are versatile techniques to deal with missingness. The imputation methods facilitate analysis by keeping almost majority of the dataset for further analysis. Here, our goal is to compare a robust imputation technique (an MLE based approach) to the conventional imputation techniques namely mean imputation, KNN imputation, and EM algorithm and choose the best one for our AML study.
Keywords
Acute Myeloid Leukemia
High Throughput Data
Imputation
Maximum Likelihood approach
Mean, KNN, EM algorithm
Decision limits are crucial in laboratory medicine for guiding diagnostic and decision-making processes. While reference ranges offer general guidelines, decision limits, often one-sided upper limits, serve as diagnostic criteria for specific conditions. In complex diagnoses which require the use of several analytes, computing separate univariate decision limits increases the number of false positives. As an alternative to this, it is recommended to construct multivariate decision limits that account for the cross-correlations among analytes. Moreover, appropriate decision limits may also be needed for specific values of covariates (e.g., age and sex). For this reason, this study proposes an approach to compute regression-based multivariate statistical decision limits within the multivariate normal framework. The criterion used in obtaining the decision limits is related to Bayesian tolerance regions. Simulation results show that the proposed statistical decision limits have highly satisfactory frequentist properties. Finally, the approach used in this study controls the desired false positive rate at a prespecified level of confidence.
Keywords
decision limits
Bayesian multivariate regression
tolerance interval
laboratory medicine
Logistic linear mixed models are used in small area estimation to construct unit level model-based estimators for binary outcomes. Instead of assuming common regression coefficients for all small domains in the traditional model, a new model with clustered coefficients is proposed with consideration of random effects, which allows different regression coefficients or intercepts in different clusters of domains. To achieve the goal, an optimization problem based on penalized quasi-likelihood (PQL) and pairwise penalties is considered. A new algorithm based on the linearized alternating direction method of multipliers (ADMM) algorithm is developed to find clusters and estimate parameters simultaneously. Simulations are used to compare the proposed approach and traditional approaches to show the advantages of the proposed estimator.
Keywords
Logistic linear mixed models
Clustering
ADMM algorithm
Small area estimation
First Author
Xin Wang, San Diego State University
Presenting Author
Xin Wang, San Diego State University
Large language model (LLM)-generated responses to open-ended questions have become increasingly common in online surveys. Unfortunately, this potentially compromises survey data quality and increases the cost of data collection and review. To tackle this challenge, we have developed a machine learning classifier that detects AI-generated responses to open-ended survey questions. This presentation will highlight the ways in which off-the-shelf LLMs do not respond like typical survey respondents and how key differences helped drive feature selection for the classifier. To create training data, we generated responses from LLMs (e.g., GPT, Llama, and Claude) to compare with responses from survey respondents to multiple open-ended questions. Performance is excellent, with precision and recall as high as 99% on held out training data and in the low 90% for unseen observations from subsequent surveys using different types of questions and subject matter domains. We will conclude this presentation with a discussion of bias and equity considerations, noting how performance varies across groups and suggesting equitable approaches to handling responses labeled a potentially from AI.
Keywords
AI
Large language models
Survey data quality
Machine learning
Text analysis
Natural language processing
Advancements in cancer therapies have significantly improved long-term survival. However, survivors face higher risks of long-term morbidities and mortality. Those treated with cardiotoxic therapy (anthracyclines or chest radiation) are at greater risk of cardiovascular diseases that include Afterload (AF) as a continuous variable and Fractional Shortening (FS) as a binary variable. FS is classified as abnormal (FS < 0.28) or normal (FS ≥ 0.28). Hudson et al. (2007) evaluated risk factors for these outcomes independently. This manuscript presents a likelihood-based approach for jointly analyzing these mixed endpoints. We illustrate this by assessing the effect of risk group (AR vs. NAR) on the cardiovascular outcomes of AF and FS using Hudson et al.'s (2007) data. First, we analyze AF and FS separately using a linear regression model and a probit model. Then, we apply joint modeling to account for FS-AF correlation and compare results with the independent approach of analysis. Additionally, we conduct simulations to assess the performance of the joint modeling approach under different sample sizes and correlation values to evaluate the operating characteristics.
Keywords
Frequentist Approach
Mixed Endpoint
Probit Model
Linear Regression
Hotelling’s T-Squared Test
Simulation
Data from the Hawaii Marine Recreational Fishing Survey (HMRFS) were used to obtain non-commercial catch. Non-commercial catch was estimated as the product of the catch rate and fishing effort, and adjustments were made to both components to exclude fishing trips that were not entirely recreational from the HMRFS data. By estimating only non-commercial catch, values could be combined with mandatory commercial fishing reports to estimate total catch for the stock assessment. Two adjustments were made during the catch rate estimation: 1) catch claimed as sold in HMRFS was excluded; and 2) catch claimed as non-sold by expense fishers (who sometimes sell fish to cover fishing expenses) and part-time commercial fishers was excluded, as these fishers must report all catch in the commercial reporting system, even if only a portion is sold. Fishing effort estimates (derived from telephone and mail surveys) were also adjusted to exclude trips from expense and part-time commercial fishers. The non-commercial catch from this study in combination with the catch in commercial reports may better define total fish removal for Deep7 bottomfish stock assessments in the main Hawaiian Islands (MHI).
Keywords
Hawaii Marine Recreational Fishing Survey (HMRFS)
Deep7 bottomfish
non-commercial catch
fishing effort
catch rate
stock assessment
In economics and the social and health sciences, longitudinal sample surveys often exhibit complex sampling design features such as unequal selection probabilities, stratification and clustering of individuals. For data collected from some large-scale surveys, or from surveys linked to administrative data files, special methods are required for inference when exploring relationships between outcome variables and covariates.
Under the semiparametric modeling approach, the within-cluster correlation is unspecified. The quadratic inference function approach provides consistent and asymptotically normal estimators of model parameters when their number is finite. In this paper, we consider the case when the number of covariates grows to infinity as the number of clusters increases and we illustrate how the rate of divergence in the number of parameters affects the convergence rate of penalized estimators. The procedure simultaneously estimates parameters and selects important variables, accounting for both within-cluster correlation and the complex survey design features.
Keywords
Complex sampling design
Diverging number of parameters
Longitudinal data
Model selection
Oracle property
Quadratic inference functions