An Introduction to Network-driven Sampling Methods - Online

This online course introduces innovative sampling methods that are intuitive and feasible to apply. Rather than sampling from a predefined list, which is often not available, we sample from network connections instead, even if the connections are not known in advance. 

Whether you are a statistician, data scientist, biologist, or sociologist, you may find yourself studying data that has an inherent network structure. Have you ever opened a standard textbook on sampling, but could not find any methods that seem to help you? Or have you wondered how your data might be generalised to the real-world population it represents?

For example, you may be interested in social networks, but you are not sure how to collect the data in a way that capitalises on the links among the individuals. Or, during an epidemic, one knows one must track the people in contact with the confirmed cases, but you may have wondered what one needs to do to be able to estimate the case prevalence at the same time. For a third example, you may have heard that Google calculates PageRank based on randomly crawling the hyperlinks from one webpage to another, but you might want other measures of connectivity, such as small cliques of webpages with mutual hyperlinks to each other. Or perhaps you work with spatial data and are used to treat space continuously. Here, it is also possible to introduce a network structure among all the actual locations and represent the spatial data on a graph. 

If your problem resembles any of these examples, this is the course that will help you. Unlike the standard sampling methods, where units are selected from a list at each stage, we will introduce you to novel approaches where the sampling is based on the connections among the units, i.e. network-driven, even when the connections are not known in advance. You will be introduced to many such sampling methods as well as the unifying ideas that enable them, which have applications to various socio-economic, technical, biological or environmental phenomena.

Each method will be explained using small datasets that can be easily visualised so that the level of abstraction is kept to a minimum, and the accompanying computer workshops shall allow one to explore the underlying ideas directly through numerical results without having to labour with the relevant mathematical formulae and equations. 

The course covers the following topics, organised across 5 sessions.

• Multiplicity sampling, such as indirect sampling or line intercept sampling. Some examples: customer-to-product transaction data, sampling siblings via households for gene study, line-in-space transecting animal-tracks for the purpose of wildlife conservation.

• Adaptive network sampling, where a connection is sampled only if it satisfies a threshold condition. Some examples: persons-in-contact during an epidemic, contiguous cells-in-space in species abundance study.

• Probability snowball sampling, where the sample grows wave-by-wave by the new connections encountered. Some examples: sampling friendship networks, actor-actor co-casting for equitable professionalism. 

• Walk sampling by random, targeted and lagged walks. Some examples: Google PageRank, efficient sampling of spatial objects, dynamic monitoring of a specific subpopulation.

• Graph sampling as the unifying framework, where the interest may be any specific subgraphs such as nodes, edges, triangles, squares, K-stars. Generally applicable sampling strategies.

By the end of the course participants will:

• Have been introduced to a range of network-driven sampling methods
• Have obtained first-hand experience with various associated probability-based estimators 
• Be aware of the unifying ideas of graph sampling strategies applicable to a variety of situations

THIS COURSE IS BEING TAUGHT EVERY THURSDAY OVER 5 WEEKS FROM 18 FEBRUARY - 18 MARCH (10:00-12:00, hour break 13:00-14:00) AND EQUATES TO TWO TEACHING DAYS FOR PAYMENT PURPOSES.

Course Description

Pre-requisites

Basic knowledge of finite population sampling methods, such as simple random sampling, Poisson sampling would be ideal. Ideally also knowledge of two basic inverse-probability-weighted population total estimators, i.e. the Horvitz-Thompson estimator based on sampling without replacement, and the Hansen-Hurwitz estimator based on sampling with replacement. 

Basic skills with R and Rstudio - ideally knowledge of: 

• Using R to perform calculations or simulations, output or plot results.
• Use of given scripts in R to execute simple algorithms including loops. 
• Modify the given R scripts or write own codes to explore alternatives and different ideas (if interested).

It is possible to follow the course without attending the computer workshops, but they help to understand implementation and provide illustrations of the learned material.

Target Audience

Anyone working with data that has an inherent network structure. The course is suitable for people from all career stages with a quantitative or statistical background, including both senior and junior researchers, post-graduate and doctoral researchers. You may be a statistician by profession or a researcher from an applied area. You may be a data scientist, biologist, sociologist, psychologist, economist and so on, who uses sampling methods, or simply would like to find out whether network-driven sampling methods might help you to tackle the problems in your profession.