|
Department of Biostatistics Neurostatistics Working Group 2011 - 2012 |
ABSTRACT: This seminar will describe a longitudinal extension of the Multiple Indicators Multiple Causes (MIMIC) model to characterize associations between cognitive decline and findings of Alzheimer’s Disease or Cerebral Infarction at death. The data come from the Religious Orders Study, a longitudinal study of priests, monks, and nuns who agreed prospectively to autopsy.The speakers will describe statistical methods for identifying a specific neuropsychological profile characteristic of emerging Alzheimer’s disease (AD) and cerebral infarction. They hypothesized that specific neuropsychological functions are preferentially impaired in the presence of AD and vascular neuropathology. The seminar will cover three topics, (1) Background, (2) an extension of the MIMIC (multiple indicator, multiple cause) model to the longitudinal setting and its implemention in Mplus, and (3) results.
The study used data from the Religious Orders Study (ROS), a large prospective study of cognitive aging and neuropathology. The sample included 502 ROS participants followed from enrollment to death with an annual neuropsychological battery and brain autopsy. The analytic approach involved the use of Mplus software to estimate a measurement model for neuropsychological performance assessed with 17 neuropsychological tests, extended to accommodate repeated assessments over 10 years. Preliminary results will be presented describing the general pattern of cognitive decline and impairments specific to individual tests in the presence of AD neuropathology or cerebrovascular infarction.
ABSTRACT: In older adults without clinical or behavioral evidence of neurodegenerative disease, substantial neuropathological markers are often evident. These markers include amyloid deposition (related to the development of Alzheimer's disease) and white matter hyperintensities (related to cerebrovascular disease). Such "silent" neuropathology may have early influences on cognitive function and contribute to the developmental trajectory of cognitive decline during advanced aging. This talk will explore our efforts to combine methods using PET and MRI to measure age-related neuropathological markers, examine their relation to cognition, and probe their influence on underlying neurocognitive systems measured with functional connectivity and task-based functional MRI.
ABSTRACT: In this work we provide a new class of frailty-based competing risks models for clustered failure times data. This class is based on expanding the competing risks model of Prentice et al. (1978) to incorporate frailty variates, with the use of cause-specific proportional hazards frailty models for all the causes. Parametric and Nonparametric maximum likelihood estimators are proposed. The main advantages of the proposed class of models, in contrast to the existing models, are: (1) the inclusion of covariates; (2) the flexible structure of the dependency among the various types of failure times within a cluster; and (3) the unspecified within-subject dependency structure. The proposed estimation procedures produce the most efficient parametric and semiparametric estimators and are easy to implement. Simulation studies show that the proposed methods perform very well in practical situations.
ABSTRACT: In studies that compare several diagnostic groups, subjects may not only be measured on a certain set of response or feature variables, but also matched on a number of demographic characteristics and measured on additional covariates. The results from multiple studies done on the same groups of subjects can be integrated using statistical discrimination techniques to identify which feature variables best distinguish among different diagnostic or treatment groups, while accounting for the dependencies among the feature variables within a subject. However, subject matching and the use of covariates appear not to have been taken into consideration when implementing these methods. This talk will present a series of modified approaches to two discrimination procedures, linear discriminant analysis and classification trees via CART, that account for the effects that both subject matching and covariates may have on the feature variables of interest. The proposed methodology is then applied to a series of post-mortem tissue studies comparing several neurobiological characteristics of schizophrenia subjects and normal controls, and to a post-mortem tissue study comparing brain biomarker measurements of monkeys across three treatment groups. The observed results obtained from these applications demonstrate the improvements the proposed methodology can achieve compared to traditional discrimination methods.
ABSTRACT: Tests of higher cognitive processing such as IQ tests and memory tests have been used to identify neurological deficits resulting from heavy metal toxicity. While the data from these tests have aided in understanding the etiology of neurotoxicant poisoning, they measure processes emerging from highly integrated neuronal circuits that may not show measurable deficits until after prolonged exposure. Assays that better access affected neuronal circuits might reveal signs of dysfunction before clinical symptoms become apparent and could provide better insight into the mechanisms that underlie these deficits. To this end, I have explored the use of a neurophysiological test of the acoustic startle reflex (ASR), which is defensive response involving the contraction of facial muscles that occurs in response to a loud and unexpected stimulus . The magnitude of the ASR is measured using electromyography (EMG), and can be modulated by experimental manipulations or psychopathological conditions. This talk will present findings from a study on the effects of lead exposure on classical conditioning of the ASR -- a simple form of learning in the brain-- in a population of elderly men. The statistical methods used in analyzing acoustic startle response data will also be discussed.
ABSTRACT: In analyzing high-dimensional datasets, there is often interest in uncovering cluster structures and identifying variables associated with the clusters. I will present some Bayesian methods we have proposed to address such questions in a unified manner. The first problem I will discuss is concerned with discovering homogeneous subgroups of samples and identifying variables that discriminate across the subgroups. We use mixture models with an unknown number of components to uncover the cluster structures and build a stochastic search variable selection method into the model to identify discriminating variables. The second problem is concerned with relating data sets from various high-throughput technologies by uncovering cluster structures in the data and identifying groups of associated markers across the data sets. We use a stochastic partitioning method that combines ideas of mixtures of regression models and variable selection methods to search for sets of covariates associated with sets of correlated outcomes. I will illustrate the methods with applications to genomic data sets.
ABSTRACT: Functional data collection has become increasingly common and in many studies they are obtained longitudinally. For example, MRS produces a spectrum composed of a mixture of pure metabolite spectra, instrument noise and baseline profile from an in-vivo MRI scan. Such data can be used to study the associations of the neurological and inflammatory factors with the neurocognitive impairment (NCI) of HIV-infected patients. Most common statistical methods applied to study such associations follows a two-step approach: extraction of the metabolite concentrations and modeling of their relationship with an outcome.
We propose to use the full spectrum via a recently established functional linear model methodology using partially empirical eigenvectors for regression (PEER). This method provides a principled way of using a priori knowledge via an informative penalty operator in the regression function estimation process. In this work, we extend PEER to the longitudinal setting with continuous outcome and a longitudinal functional (MRS) covariate. We establish properties of the regression function estimator and present a simulation study as well as preliminary analysis of the longitudinal MRS data from the HIV Neuroimaging Consortium study of neurocognitive impairment of HIV-infected patients.
Joint work with Madan Kundu and Timothy Randolph.
ABSTRACT: Consider a clinical trial of a potentially lethal disease where subjects are assigned to two treatments and the response is recorded at baseline (prior to treatment) and after a pre-specified follow-up period. The treatment effect is then assessed on the change in response from baseline to the end of follow-up time. Unfortunately, for some patients, death may preclude the measure of the post-treatment response. Estimates of the treatment effect based solely on the subset of truly observed measures might be biased if death selectively removes patients who had "poor" response.
To account for all randomized patients in the analysis of the data, one common approach is to use worst-rank composite outcome. Patients with missing observations are assigned the worst ranks and survivors are ranked according to their outcome measures.
For this talk, we will explore the statistical properties of the two-sample Wilcoxon test on worst-rank composite outcome and present closed-form formulas for power and sample size calculations. In addition, we will present the weighted Wilcoxon test and show that it is more powerful than the ordinary Wilcoxon test. Finally, the weighted Wilcoxon will be applied to a data set from a randomized clinical trial on acute ischemic stroke.
ABSTRACT: Epidemiologic research on the association between hypertension and cognition or dementia collectively suggests that the association is dependent on the age at which hypertension status is assessed. Several potential explanations for this pattern are possible and include the influence of selection bias, an independent effect of duration of hypertension, an independent effect of age at onset of hypertension, and reverse causation. I will present two analyses that attempt to determine whether these factors, or a combination of these factors, can explain the age-dependent association that has been observed in the literature. The first analysis reproduces the age-dependent association observed across studies within a single dataset including longitudinal data on hypertension over the course of 30 years, using a penalized cubic spline to model betas across multiple point estimate studies nested within this larger dataset. We then explore the influence of selection bias using inverse probability weights for censoring and the influence of duration through stratification based on duration of hypertension prior to cognitive testing. The second analysis focuses more narrowly on how age at onset and duration of hypertension modify the association between hypertension and cognition. I will present a linear marginal structural model, using inverse probability weights for censoring and confounding, that explores this issue.
ABSTRACT: Efforts at controlling the spread ofHIV or other infectious agents can be aided by an understanding ofthe transmission networks along which infection spreads. However, itcan be challenging to identify the most valuable network features toestimate for this purpose and to obtain sufficiently reliableestimates of them. This is particularly true in the setting ofsexually transmitted infections as people are often reluctant toreveal their sexual connections. These challenges increase with thecomplexity of the network feature of interest; hence, most efforts tounderstand sexual networks make use of only egocentric data. Thoughcollection of additional network properties may be difficult, it isimportant to assess the incremental gain in accuracy of predictingthe size of an epidemic by knowing network properties that extendbeyond ego-centric properties. Concurrency has been identified as apossible important ego-centric factor in the spread of HIV insub-Saharan Africa and elsewhere. In this presentation, we willexplore a method to construct dynamic networks from estimated networkproperties, including degree distribution, concurrency, and mixingpatterns. This method allows us to assess the added value ofcollecting network-level data that permits assessment of featureslike assortativity, that cannot be estimated from ego-centric dataalone. The methods will be investigated using a data characterizingthe sexual network in Likoma Island, Malawi as well as from Mochudi,a large village in Botswana. I am also interested in discussing how these techniques can be applied to Neuroscience data.
ABSTRACT: All currently available functional neuroimaging techniques face the intrinsic limitation that the contrast being imaged is based on metabolic or vascular changes accompanying neuronal activity, rather than on the neuroelectric activity itself. These metabolic/vascular changes are slow, spatially coarse, and are likely to miss much of the neuronal information encoding. We have developed a mode of imaging, using fast gradient-echo MRI combined with EEG, to attempt to more directly image human brain discharges. We focus on epilepsy patients who have large-amplitude interictal spikes. Our approach and early results will be described in this talk, as will potential clinical and basic research utility of the technique.
ABSTRACT: Background: Survival bias is the phenomenon by which individuals are excluded from analysis of a trait because of mortality related to the expression of that trait. In genetic association studies, variants increasing risk for disease onset as well as risk of disease-related mortality (lethality) could be difficult to detect in genetic association case-control designs, possibly leading to underestimation of a variant's effect on disease risk.
Methods and Results: We modeled cohorts for three diseases of high lethality (intracerebral hemorrhage, ischemic stroke, and myocardial infarction) using existing longitudinal data. Based on these models, we simulated case-control genetic association studies for genetic risk factors of varying effect sizes, lethality, and minor allele frequencies (MAF). For each disease, erosion of detected effect size was larger for case-control studies of individuals of advanced age (age > 75 years) and/or variants with very high event-associated lethality (Genotype Relative Risk for event-related death > 2.0). We found that survival bias results in no more than 20% effect size erosion for cohorts with mean age < 75 years, even for variants that double lethality risk. Furthermore, we found that increasing effect size erosion was accompanied by depletion of MAF in the case population, yielding a "signature" of the presence of survival bias.
Conclusion: Our simulation provides formulas to allow estimation of effect size erosion given a variant's odds-ratio (OR) of disease, OR of lethality, and MAF. These formulas will add precision to power calculation and replication efforts for case-control genetic studies. Our approach requires validation using prospective data.
ABSTRACT: All currently available functional neuroimaging techniques face the intrinsic limitation that the contrast being imaged is based on metabolic or vascular changes accompanying neuronal activity, rather than on the neuroelectric activity itself. These metabolic/vascular changes are slow, spatially coarse, and are likely to miss much of the neuronal information encoding. We have developed a mode of imaging, using fast gradient-echo MRI combined with EEG, to attempt to more directly image human brain discharges. We focus on epilepsy patients who have large-amplitude interictal spikes. Our approach and early results will be described in this talk, as will potential clinical and basic research utility of the technique.
ABSTRACT: Adaptive designs have been proposed as a promising new approach that may help improve the efficiency of the drug discovery process. Many recent papers have focused on the potential gains in flexibility and efficiency with the use of adaptive hypothesis testing methods that control the type I error. However, confirmatory phase III clinical trials also need to produce results that are interpretable, in that sufficiently reliable and precise estimates of treatment effect can be computed at the end of the study. This helps ensure that regulatory decisions are based on reliable evidence of benefit to risk, new drugs are appropriately labeled, and clinicians can effectively practice evidence-based medicine. We introduce and evaluate different methods to compute point and interval estimates after a pre-specified adaptive design allowing interim modifications to the sampling plan based on the unblinded estimate of treatment effect.
ABSTRACT: Data from a large prospective longitudinal study was used to estimate propensity scores (PS) of dementia risk for older adults clinically diagnosed as cognitively normal at baseline. PS were estimated as a function of age, education, gender, APOEe4, blood pressure, diabetes, BMI, physical activity, smoking, and depression. PS, estimated by maximum likelihood logistic regression, and demographic factors were further modeled as predictors of a composite score measuring global cognition. Growth mixture models were used to estimate latent classes with distinct cognitive trajectories over time. Two latent classes emerged for global cognition - one with a subtle decline and one with a faster decline over time. Older age, higher propensity scores, and gender were significantly associated with group membership in the faster decline group; while higher education level was associated with membership in the class with subtle decline. The study provides support for the protective role of education in mitigating cognitive decline for a group with a favorable cardiovascular disease (CVD) risk profile. Level of education, however, did not protect the latent group with a higher prevalence of CVD.
| Back to HSPH Biostatistics |
Maintained by the
Biostatistics Webmaster
Last Update: May 7, 2012 |