Abstract

Publication selection bias, i.e. the selective reporting of statistically significant results, is a serious problem for any empirical research. As consequence, the estimated effects in the literature can appear considerably larger/smaller than they are and more specifically meta-estimated effect size can be severely biased. Recently, several methods of publication correction have been proposed. None of these methods, however, works reasonably well in general. In particular, PET performs well if the true eﬀect is zero, PEESE only works well for large true eﬀects, and Top 10 also works well for large true, but is ad-hoc and ineﬃcient. We propose a novel publication bias method that works well across a wide range of eﬀect sizes, publication bias incidences, and meta-sample sizes. It also outperforms a combination of PET and PEESE in a variety of scenarios. Extensive simulations illustrate the gains in terms of efficiency as well as bias.

]]>Abstract

Functional magnetic resonance imaging (fMRI) is a functional neuroimaging method to measure brain activity by detecting changes of blood flow in the brain. The analysis of fMRI time series during real-world experience is methodological challenging because of the human brain processing a variety of features simultaneously in multiple-uncontrolled and -dynamic stimuli. Previous studies have used block designs to localize the activated brain regions. Those experimentally defined control-states hardly exist in the real world. The reproducibility analysis is an approach to investigating event-free spontaneous brain activity. Here we designed an experiment engaging long-term auditory stimulation reflecting a real world experience. The reproducible brain activity is estimated between multiple experimental runs within a participant and across a group of participants were calculated using the proposed intraclass correlation (ICC) statistic, which is directly applicable to pre-processed fMRI time series. An agreement between within- and between-participant analysis results elucidates the possible inter-individual variability in BOLD responses during the experimental stimulation. The results showed that the auditory network in the temporal cortex reflects the experimental stimulation along with other networks such as those associated with imitation, memory, and emotions. We suggested that the proposed ICC statistic as well as agreement between within- and between-participant analyses can be used to investigate functional networks in real-world fMRI experiments.

]]>Abstract

A regression tree method for count data with excess zeros is proposed. At each node, a Poisson regression model which accommodates excess zeros is fitted. A likelihood-base procedure is proposed to select split variables and split sets. Node deviance is then used in the tree pruning process to avoid overfitting. Our method is free of variable selection bias. It is demonstrated to be effective in simulation and real data studies.

]]>Abstract

The integer low-rank approximation of integer matrices (ILA) has received attention recently due to its capacity of naturally representing parts of integer data sets. Different from the general low-rank approximations, the integer approximation is naturally discrete, therefore, the conventional techniques for matrix approximation, such as SVD and non-negative matrix approximation, are inappropriate and unable to solve this problem. To the best of our knowledge, a numerical method for finding a low-rank integer approximation of an integer matrix has not been proposed in the literature earlier.

In this talk, we want to propose a block coordinate descent method to obtain the integer low-rank approximation of integer matrices. This method consists of recursively finding integer solutions of integer least square problems. Applications on the real world problems such as the market basket transactions, association rule mining, cluster analysis, and pattern extraction will be given. Numerically, we show that our ILA method can find a more accurate solution than any other existing methods designed for continuous data sets.

]]>