An improved cumulative sum-based procedure for prospective disease surveillance for count data in multiple regions
Dassanayake, S., & French, J. P. (2016). An improved cumulative sum-based procedure for prospective disease surveillance for count data in multiple regions. Statistics in Medicine. https://doi.org/10.1002/sim.6887
- Cumulative counts are created from neighbouring areas for monitoring/detecting diseases outbreaks
- prospective surveillance
- decisions made at regular intervals about incidence, c.f. analysing the whole dataset retrospectively
- so possible multiple comparisons issues?
- For CUSUM-based methods, the in-control distribution can be
estimated via bootstrap
- p-values for exceeding CUSUM threshold can then be calculated
- why p-values?
- is this another case for going Bayesian to get posterior probabilities
- Provides discussion on standard Poisson CUSUM
- parameters controlling the process are calculated via average run length (ARL)
- ARL: desired average time between two false alarms when the process is \‘in control\’
- this seems to be a difficult concept to set a limit for
- spatial CUSUM modifies counts and control parameters by a weighted
average of all counts
- weights are based on the distance between areas
- Justification for the procedure is that p-values are preferred over
- why? surely there\’s a one-to-one correspondence?
- false discovery rate is used for error control
- this seems fine, but I\’m not sure I understand the procedure fully
- FDR = 1 - PPV, which makes that point a bit clearer
- In the application, the authors use the first two years of data to
estimate the in-control distribution
- this is because \‘… there were no unusually high disease counts reported from any of the states during this period\’
- this is clearly problematic: is this the eyeball test
- the results will clearly depend on this choice, so a sensitivity analysis would be important
- I found this somewhat difficult to read overall
- this could be due to my lack of knowledge in the area though!