Risk assessment models in cancer

Dysphagia is a common adverse effect resulting from radiation therapy for cancer. Since clinicians can vary the radiation dose on a fine scale, it is natural to ask how the risk for adverse events quantitatively changes as the radiation dose increases. Logistic regression is a widely used tool for relating exposures to risks. When working with a continuous exposure such as radiation dose, it is possible to combine model-based analysis with logistic regression, and descriptive summary statistics, to assess whether the exposure affects the risk as specified by the model. CSCAR Research Scientist Myra Kim recently collaborated with researchers from the UM Medical School to examine how the risk for complications is related to the radiation dose in patients being treated for oropharyngeal cancer.

Glycoprotein screening for cancer biomarkers

To enable early cancer detection, researchers have long sought to identify biochemical markers that discriminate people with early stage cancer from risk-matched controls. High throughput screening allows large collections of candidate biomarkers to be assessed for potential use as cancer biomarkers. Working with researchers from the University of Michigan Department of Surgery, CSCAR Director Kerby Shedden developed new screening statistics and visualization tools that help identify the most promising candidates for follow-up study, while excluding certain types of false positives that are peculiar to the assay system that was employed.

Likelihood-based correlation analysis for censored data

Censored data arise when a value of interest, such as the concentration of a chemical in a specimen, lies within a known set of values, but the specific value is not known. For example, assays used in environmental toxicology typically have a lower limit of detection. The measured concentration of the compound of interest will be censored if the true concentration lies below the detection limit. Most standard statistical tools cannot accommodate all possible types of censored data. Motivated by joint work with toxicology researchers in the Netherlands, CSCAR graduate student research assistant Yanming Li, along with CSCAR Associate Director Brenda Gillespie and Director Kerby Shedden are developing an R package that will perform likelihood-based correlation analysis on all types of censored data. Likelihood-based inference for such problems has long been known to perform well in terms of its statistical properties, but only recently became computationally feasible for routine use.

Functional regression analysis of engine efficiency

Modern internal combustion engines utilize sophisticated control units to precisely time the ignition of fuel during each engine cycle. Ignition timing has a major influence on critical engine performance characteristics such as efficiency and emissions. Researchers aiming to improve engine performance are able to collect detailed data on hundreds of engine cycles using instruments that capture multiple parameters at high temporal resolution. Working with researchers from the University of Michigan Department of Mechanical Engineering, CSCAR Director Kerby Shedden used functional regression techniques to quantify the uncertainty in the relationship between crank angle and the oxygen concentration in the combustion chamber. Modern functional data analysis techniques have been developed by statisticians over the past twenty years to improve the analysis of data whose conditional means or quantiles follow continuous functional relationships. Techniques involving penalized regression in function spaces have largely supplanted unwieldy classical methods for analyzing functional data such as parametric nonlinear regression.

Longitudinal analysis of hormone levels

Many research investigations focus on changes within subjects over time, with an emphasis on how the pattern of such longitudinal changes varies among individuals. State of the art longitudinal modeling techniques allow these effects to be assessed efficiently, while accommodating common data complications such as uneven frequencies of observation among subjects, uneven duration of observation among subjects, feedback and autocorrelation, selection bias, and measurement errors. Working with a multidisciplinary team of researchers, CSCAR consultant Brady West used longitudinal modeling techniques to identify changes in hormone levels that were specifically associated with early pregnancy loss.

Analysis of neuroimaging data

Magnetoencephalography (MEG) is a form of neuroimaging that can be used to assess activity in various brain regions. Working with researchers from the University of Michigan Department of Psychiatry, CSCAR consultant Dave Childers used spectral coherence analysis to examine connections between pairs of brain regions. Since there are over 50 brain regions, and 1400 pairs of regions, a direct analysis would produce an overwhelming set of results that would be difficult to interpret. Modern methods for statistical error control allow false positive associations to be limited, while still retaining power to detect potentially important differences between the study groups.

Path modeling

Diabetes, hypertension, and chronic kidney disease (CKD) commonly co-occur within individuals. A team of University of Michigan and external researchers including CSCAR consultant Brady West recently demonstrated that periodontal disease may also play an important role in this comorbid system (Fisher et al., Kidney International 78, Feb 2011). Using Structural Equation Modeling (SEM), the team found evidence that periodontal disease may increase an individual’s risk for CKD, even when taking into account the roles of diabetes and hypertension. SEM is a powerful technique for establishing the plausibility of causal pathways. It allows competing explanations for a relationship among several variables to be objectively contrasted. This work establishes a strong basis for future longitudinal studies exploring causal pathways in CKD.

Geographic analysis of drug compliance

Long term drug treatment with drugs such as ACE inhibitors and statins has been shown to protect patients with coronary artery problems from subsequent events. But poor adherence to these treatments is common. To better understand this phenomenon, CSCAR staff consultant Giselle Kolenic collaborated with researchers from the UM Schools of Medicine and Pharmacy to conduct a geographical analysis of treatment compliance (Hoang et al., Pharmacotherapy; Oct, 2011). Using home addresses and compliance information from a patient registry, the researchers carried out a hot-spot analysis to identify geographic areas where compliance rates were unusually high or low. One potentially important trend that emerged from the team’s findings is that compliance is lower for patients living further from the hospital where they were treated. This study illustrates that geographic statistical analysis can be an important tool for understanding health behavior.