(by Yueqing Wang/YQnancy)
The Joint Statistical Meetings started with a very eventful day, with plenty of receptions and mixers to meet and mingle. That aside, I found both sessions I attended engaging and inspiring.
In the session on statistical methods for spatio-temporal data, Matthew J. Heaton from Duke University described the necessity and benefits of modeling lagged effects in both spatial and temporal dimensions. He then introduced a kernel-based predictor to incorporate different weight patterns in space and backwards in time. The plot of the estimated spatial kernel is informative in showing the different effective radius at different locations; similar results for the time dimension. Matthew also discussed the choices and consequences of using different kernels. After the session, Matthew told me his excitement to move forward in interpreting the results, as well as improving the algorithm’s computational efficiency. Wish him best luck!
In the same session, Veronica J. Berrocal from the University of Michigan introduced her method to combine point measurements and gridded simulation. It takes advantage of the accuracy of point measurements and the information about the spatial structure from the simulations.
In the session on analysis and inference for complex networks, Eric Kolaczyk from the Boston University presented a convincing and inspiring case, calling for statistical analysis to generalize the mapping, sampling, and characterizing complex networks. Particularly, he emphasized the importance to address the uncertainty in network summary statistics and model parameters. The notion of networks have evolved over the years, shifting from traffic and gene network analysis to social networks. In fact, Eric’s plot of number of JSM topics involving networks over time shows a peak at 2003 and a valley around 2005, with a rising tail. However, not much has been developed to shape the notions of properties and performance of network models over the years. Like Edoardo M. Airoldi from Harvard University pointed out, nowadays we actually have available data for networks that are 1) large-scale; 2) completely mapped; and 3) realistic. All of these exciting opportunities are inviting statisticians’ dedications.
Susan Martonosi from Harvey Mudd College showed us the possible application of network flow studies for terrorists network disruption. The centrality of a node is defined by the loss of paths in the network, if this particular node is removed from the network. One of the most pressing questions is which “end” node to disrupt, in order to force more flow going through the “root” node and make it more visible. Also her wish to improvement the communication and collaboration between statistics and operational research is shared.
(Yueqing Wang is a Ph.D. student at the University of California, Berkeley.)