Neurostatistics Working Group

This working group provides a forum for presentation and discussion of completed, ongoing, or planned statistical analyses of neurological data. Such data include, for example, in vivo human brain images (anatomic, functional and spectroscopic magnetic resonance imaging), gene expression studies of human and non-human animal brain tissue (brightfield and immunofluorescence microscopy, DNA microarrays, laser micro-dissection), in vivo micro-dialysis, clinical trials data for a variety of neurologic diseases, and genetic data from family studies. Non-statistical presentations of neurological, psychiatric and technological background material will also be included. Through this seminar, statisticians will gain exposure to the statistical issues that arise in the broad field of neurology and brain imaging psychiatry and to the diverse ongoing research in this area throughout Harvard and the world. A main goal of the seminar is to stimulate statistical interest in neuroscience and neurology and to develop strategies for collaboration within these fields.

Elena Verbruggen, Ph.D.
Master's Degree Candidate, Department of Statistics, Hasselt University

"Analysis of Diagnostic Tests in the Case of an Imperfect Gold Standard: Application to Acute Kidney Injury"

ABSTRACT: None Given.

Eric Macklin, Ph.D.
Instructor in Medicine, Massachusetts General Hospital and Harvard Medical School

"Stratified Randomization with Gender-Balancing in the SURE-PD Trial"

ABSTRACT: The rationale and implementation of the randomization design for the multi-center Safety of URate Elevation in Parkinson's Disease (SURE-PD) Trial is described. Stratification by site is combined with balancing by gender using suggestions from Zelen (1974). Performance of the chosen randomization design and of stratified designs that do not employ balancing are contrasted with respect to power and logistics of drug supply.

Caterina Stamoulis, Ph.D.
Instructor in Neurology, Harvard Medical School and Beth Israel Deaconess Medical Center

"Signal Processing Methods for the Detection of Copy-number Changes in Array Comparative Genomic Hybridization Data"

ABSTRACT: Our genomes are dynamic and heterogeneous, at least in terms of small-scale genetic variability. Allelic aberrations, such as copy number changes (CNVs) have been identified in thousands of regions of the healthy human genome. However, in addition to this common genetic variability, CNVs have also been associated with a wide number of diseases. Array Comparative Genomic Hybridization (aCGH) is a powerful technology which is used to measure the entire genome and enables the simultaneous detection of a potentially large number o allelic aberrations. However, aCGH data are noisy and small-scale CNVs are often undetectable. To improve CNV detection, signal processing methods typically used to increase signal-to-noise ratio in time series contaminated by different types of noise may be adapted to analyze genomic data. The robustness of these methods to detect already documented common CNVs in healthy humans, and their application to detect disease-related CNVs in DNA of patients with menigiomas will be discussed.

J. Alex Becker, Ph.D.
Research Fellow, Massachusetts General Hospital & Harvard Medical School

"Amyloid Deposition and Brain Volume across the Continuum of Aging and AD"

ABSTRACT: A clear relationship between amyloid plaque burden and neuronal loss or regional atrophy has not been demonstrated in autopsy studies. To examine this relationship in vivo we tested the hypothesis that amyloid deposition, as measured by 11C Pittsburgh Compound B (PiB), is associated with reduced brain volume.
C11-PIB PET and 3T MRI-MPRAGE data were acquired in 117 subjects, 82 who were CDR 0, 19 CDR 0.5, and 16 CDR 1 (CDR: Clinical Dementia Rating). Each MPRAGE dataset was processed by FreeSurfer to extract gray/white segments, to parcellate the gray matter of the cortical mantle into a set of regions of interest, and to generate surface models of the gray-white and gray-CSF interfaces. The gray and white surfaces and segmentations were used to derive whole brain volume (excluding ventricles and other CSF spaces) and cortical thickness at each vertex, total cortical gray matter volume, and hippocampal volume. Volumes were adjusted for total intracranial volume. PIB retention at vertices and in ROIs was expressed as the DVR calculated using the Logan graphical method with cerebellar cortex as reference tissue.
Across the full sample of subjects, linear regression analysis revealed that increased precuneus PiB retention was associated with significantly reduced total gray matter and hippocampal volumes, and this linear relationship was maintained within the subset of CDR0.5 and CDR1 subjects. Within the subset of CDR 0 subjects, precuneus PiB retention showed no asssociation with hippocampal volume; however, a vertex level analysis revealed evidence of amyloid-associated cortical gray matter thinning in precuneus and lateral parietal regions.
This data suggests that amyloid deposition is related to brain volume loss over the continuum from normal aging to AD and that the strongest relation is observed at later stages as clinical impairment emerges. More sensitive techniques may be required to detect amyloid - related volume loss at presymptomatic stages.

Christopher Corcoran, Sc.D.
Associate Professor, Department of Mathematics and Statistics, Utah State University

"Mitochondrial DNA Copy Number, Cognitive Decline in Old Age, and Risk of Alzheimer's Disease"

ABSTRACT: Mitochondria convert calories within cells into energy through a process that yields oxidative stress. As mitochondria additionally play a key role on the signalling pathway that regulates cell death, emerging hypotheses in the study of age-related diseases are focusing on the role of mitochondrial function. A novel assay that measures average mitochondrial DNA (mtDNA) amounts in human cells has recently been used on samples from the Cache County Memory Study, and initial results indicate strong associations between mtDNA intensity with both AD onset and cognitive decline. Because of linkage between members of the CCMS with the comprehensive geneological records in the Utah Population Database, mtDNA sequencing of a relatively small proportion of the CCMS cohort (about 200 individuals) will allow imputation of mtDNA sequences for nearly all 5092 CCMS participants. This will provide a rich resource for further exploring associations between dementia risk and cognitive change with specific mtDNA mutations and haplogroups.

J. Alex Becker, Ph.D.
Research Fellow, Massachusetts General Hospital & Harvard Medical School

"Amyloid Deposition and Brain Volume across the Continuum of Aging and AD - Part II"

ABSTRACT: A clear relationship between amyloid plaque burden and neuronal loss or regional atrophy has not been demonstrated in autopsy studies. To examine this relationship in vivo we tested the hypothesis that amyloid deposition, as measured by 11C Pittsburgh Compound B (PiB), is associated with reduced brain volume.
C11-PIB PET and 3T MRI-MPRAGE data were acquired in 117 subjects, 82 who were CDR 0, 19 CDR 0.5, and 16 CDR 1 (CDR: Clinical Dementia Rating). Each MPRAGE dataset was processed by FreeSurfer to extract gray/white segments, to parcellate the gray matter of the cortical mantle into a set of regions of interest, and to generate surface models of the gray-white and gray-CSF interfaces. The gray and white surfaces and segmentations were used to derive whole brain volume (excluding ventricles and other CSF spaces) and cortical thickness at each vertex, total cortical gray matter volume, and hippocampal volume. Volumes were adjusted for total intracranial volume. PIB retention at vertices and in ROIs was expressed as the DVR calculated using the Logan graphical method with cerebellar cortex as reference tissue.
Across the full sample of subjects, linear regression analysis revealed that increased precuneus PiB retention was associated with significantly reduced total gray matter and hippocampal volumes, and this linear relationship was maintained within the subset of CDR0.5 and CDR1 subjects. Within the subset of CDR 0 subjects, precuneus PiB retention showed no asssociation with hippocampal volume; however, a vertex level analysis revealed evidence of amyloid-associated cortical gray matter thinning in precuneus and lateral parietal regions.
This data suggests that amyloid deposition is related to brain volume loss over the continuum from normal aging to AD and that the strongest relation is observed at later stages as clinical impairment emerges. More sensitive techniques may be required to detect amyloid - related volume loss at presymptomatic stages.

Armin Schwartzman, Ph.D.
Assistant Professor, Department of Biostatistics, Harvard School of Public Health and Dana-Farber Cancer Institute

"Voxel-Based Group Tests of Diffusion Tensor Brain Images"

ABSTRACT: Diffusion Tensor Imaging (DTI) data differ fundamentally from most brain imaging data in that values at each voxel are not scalars but 3-by-3 positive definite matrices, also called diffusion tensors (DTs). Typically, analyses of DTI data are based on scalar summaries of the DT, which do not capture all the information available in the data, especially directional information. In this talk I present several tools I have developed over the last few years for testing voxel-based differences between two groups of independent DTIs, first in terms of their principal diffusion direction, and then in terms of DT's full set of eigenvalues and full frame of eigenvectors. The test statistics in the latter case are likelihood ratio test statistics derived from a Gaussian model for symmetric matrices. The null distribution is approximated for an arbitrary covariance structure between the DT’s entries. The methods are illustrated on a DTI comparison study of boys and girls.

Nazem Atassi, M.D.
Instructor in Neurology, Massachusetts General Hospital and Harvard Medical School

"Retrospective Analysis Comparing Fast and Slow ALS Progressors and Studying Functional Decline after Initiating PEG and BiPAP"

ABSTRACT: Background: ALS Functional Rating Scale-Revised (ALSFRS-R) is a commonly used outcome measure that may predict survival in ALS clinical trials. Determining which of the 12 ALSFRS-R questions best predicts survival is valuable information for both trials and clinical practice. Bi-level Positive Airway Pressure (BiPAP) and Percutaneous Endoscopic Gastrostomy (PEG) improve ALS survival but their impact on patients' functional status is unclear.

Objectives: To determine patient characteristics and early ALSFRS-R answers that best predict future disease progression, and to compare the rate of functional (ALSFRS-R) and respiratory (FVC) decline before and after BiPAP and PEG placement.

Methods: This is a retrospective study of (519) ALS patients who participated in one of the following trials: celebrex (n=300), creatine (n=104), arimoclomol (n=84), and CoQ10 (n=31). A linear mixed effects model was used to distinguish fast progressing (FP) and slow progressing (SP) patients compared to mean ALSFRS-R decline over time. A t-test was used to compare continuous variables and a Chi-Squared test was used to compare categorical baseline variables. A Cochran-Armitage test was used to test for trends of baseline ALSFRS-R questions. A stepwise-Cox regression model was used to investigate the relationship between ALSFRS-R answers and survival. A linear mixed effects model was used to compare ALSFRS-R and FVC slopes before and after starting BiPAP and PEG. Only subjects with at least two ALSFRS-R and FVC measurements before and after BiPAP or PEG were included in this analysis.

Results: There were 41% fast progressors (FP) and 59% slow progressors (SP). Site of onset and time from symptom onset to diagnosis were different between the two groups (p=0.03) and (p=0.004), respectively. There was no significant difference in other baseline characteristics including age, gender, family history, baseline weight, or riluzole intake. Answers to ALSFRS-R questions at baseline visit were all significantly different between SP and FP except for handwriting (p=0.41) and respiratory insufficiency (p=0.17). Trial dropout rate of FP (41%) was 3.2 times more than SP (13%). Out of the 12 ALSFRS-R questions, climbing (p=0008), writing (p=0.04), and salivation (p=0.04) were the only predictors of survival. ALSFRS-R rate of decline was 0.35 units/month faster after BiPAP onset (p<0.0001) and 0.32 units/month faster after PEG placement (p<0.0004). These results remained significant after omitting ALSFRS-R "Swallowing" and "Respiratory insufficiency" questions for PEG (p=0.01) and BiPAP (p=0.001) analysis. The rate of FVC decline was not significantly different after BiPAP (p=0.09) or PEG (p=0.94).

Discussion and Conclusion: Slow ALS progressors are more likely to enroll in and complete clinical trials. Most ALSFRS-R questions are good early predictors of disease progression. Climbing, writing, and salivation ALSFRS-R questions are good predictors of survival and none of the "respiratory" questions are. Functional decline measured by ALSFRS-R is faster after PEG or BiPAP use.

Joint work with Kerrie Nelson Ph.D., Yuvika Paliwal M.S., Merit Cudkowicz M.D., MSc.

Jing Qian, Ph.D.
Research Fellow, Department of Biostatistics, Harvard School of Public Health

"Semiparametric Inference for Successive Durations"

ABSTRACT: For many chronic diseases, a bi-state progressive process provides a useful model for the disease progression before reaching death. For example, after surgery, colon cancer patients progress through cancer-free and caner-recurrence states. Often, scientific interests lie in the successive durations. For the one-sample problem with incomplete follow-up data, recent investigations have focused on nonparametric inference. However, in many practical situations, the distribution of the second duration is nonparametrically nowhere identifiable. Furthermore, most existing approaches require a rather restrictive censoring mechanism and have difficulty in predicting the process with given history. To address these issues, we suggest a semiparametric model that postulates normal copula for the association between the two durations, while leaving the marginals unspecified. Motivated by the colon cancer data example, we allow our model to accommodate the situation where the second duration has a probability mass at zero. We propose an inference procedure for estimation. The Finite sample performance of the proposed method is evaluated via the simulation studies and illustrated with the data from a colon cancer study.