I have a dataset that consists of 247 mosquito traps placed in fixed positions and sampled weekly for 23 weeks. I also have dengue disease incidence data (indexed to home address) during the same time period. I would like to see if there is a spatial and spatiotemporal correlation between mosquito catch rate at each trap location and dengue case incidence. I have already tried Global Moran's I and Local Moran's I in R and SaTScan Multivariate space-time permutation. None of these seem to give me what I want. Am I looking at the right place and just doing it wrong or are there better analyses I can run?

Thank you so much for the advice!! Best, Amy Green

  • "None of these seem to give me what I want." And what exactly would you like to achieve? – radek May 26 '13 at 14:00
  • 1
    Geographically Weighted Regression could be the way to go. You could run run per temporal period and compare the residuals. Logically the inputs would be volume of trapped insects as a predictor of disease as the starting point .... although, depending on the distribution of the traps, you might want to consider environmental settings as well – Andrew Tice May 28 '13 at 12:26
  • 1
    @Andrew Tice Not intending to be snarky but, out of curiosity, why would you jump to a GWR with no indication of nonstationarity, or even autocorrelation, in the data? This is a somewhat dubious approach that is only appropriate in very specific situations and should only really be used in an exploratory or experimental context. – Jeffrey Evans Nov 30 '15 at 17:02
  • @JeffreyEvans...good point...and with the benefit of two years hindsight I wouldn't! – Andrew Tice Nov 30 '15 at 20:40

I guess you're trying to estimate the effectiveness of trapping in reducing dengue contraction incidence. You may want to have a look at space-time interaction tests.

PySAL has a number of implementations in its spatial_dynamics module.


Are the mosquito data, counts? or what type of variables are you recollecting at these traps? I have seen a study where they use MaxEnt to predict leishmania in Mexico for the next 20 or so years using only the presence of the disease locations as data. I don't think it's very trustworthy but it's a starting point and there have been many developments. What are you "looking for" some sort of kriging would allow you to understand the spatial variability of your variables. You could even run some simulations if you get models that fit well. I'm pretty sure there are kriging techniques that work in a spatial-time context but I have never used them. A quick google search should throw some references. 23 weeks is not a lot of data points for time series analysis but you could try it out, the R package bfast allows you to detect abrupt breaks in the behavior of a time series, maybe you can detect these anomalies and associate them to dengue breakouts. I think you need to produce a better description of your problem and desired study outcomes.


Hard to say without seeing your data and working through some exploratory analysis. Some more details on hypothesis, sample design, and actual data collected, would be welcomed. When asking statistical methodology questions it is important that you state the hypothesis that you are testing. This can dictate statistical methodology and without knowing, we are shooting in the dark.

It is also not at all clear what the issue with the specified statistics is in relation to "not giving me what I want". I do not know what you expected with a univariate autocorrelation statistic indicating a bivariate spatial correlation. The family of SCAN statistics is quite variable with many defined distributions available. What distribution (model) did you define in SaTScan and do you actually have a hypothesis, and data, suitable for a point pattern analysis? Generally, a gridded, systematic, sample is not appropriate for a point pattern analysis.

A correlation would be very limiting from an inferential standpoint and it would seem that a regression type mode is in order here. At first blush I would think a mixed effects model with an AR-I term for time and a autocorrelation term for the spatial random effects would suit your needs. This would allow you to partition variation by time and normalize any influence autocorrelation would have on residual error and iid assumptions. Another option, if the data supported it, would be a Poisson point process model in an MCMC framework. If specified as a hierarchical model you could define time as a prior. With a kernel regression approach you could test multiple hypothesis of spatial diffusion processes or define a quadratic diffusion term. This type of model is commonly used in spatial epidemiology to get at rate of spread.

It is easy to get lost in "throwing your data against the wall" with spatial statistical approaches but, unless your sample design was intended to capture spatial process and you have a well formulated question regarding spatial effect, this can be a futile exercise.

Because of the easy availability of methodologies frequentists methods are often overlooked. There are regression models available that can readily deal with spatial data (spatial and conditional autoregressive, spatial regression, polynomial regression, mixed effect models, canonical regression, kernel regression, semi and non parametric regressions, ...) and if you are intending on inference these should be explored in relation to your hypothesis.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.