A Spatially Discrete Approximation to Log-Gaussian Cox Processes for Modelling Aggregated Disease Count Data

Olatunji Johnson

2018-08-13

Introduction

This article presents a simple tutorial code from SDALGCP package to make inference on spatially aggregated disease count data when one assume that the disease risk is spatially continious. There are two main functions provided by the package, for parameter estimation and for prediction.

Model

Our goal is to analyse of diease count data, more specifically when disease cases are aggregated over a partition, say \((\mathcal{R}_{1}, \ldots, \mathcal{R}_{n})\), of the area of interest, \(A\), which can be written mathematically as \[\begin{eqnarray} \label{eq:data} \mathcal{D} = \left\{(y_{i}, d_{i}, \mathcal{R}_{i}): i=1,\ldots,n\right\} \end{eqnarray}\]

where \(y_{i}\) and \(d_{i}\) are the number of reported cases and a vector of explanatory variables associated with \(i\)-th region \(\mathcal{R}_{i}\), respectively. Hence, we model \(y_{i}\) conditional on the stochastic process \(S(X)\) as poission distribution with mean \(\lambda_i= m_{i} \exp\{d_{i}\beta^* + S_{i}^*\}\). Then we assume that \(S^* \sim MVN(0, \Sigma)\), where \[\Sigma_{ij} = \sigma^2 \int_{\mathcal{R}_{i}} \int_{\mathcal{R}_{j}} w_i(x) w_j(x') \: \rho(\|x-x'\|; \phi) \: dx \: dx'\], where \(w(x)\) is population density weight. There are two classes of models in this package; one is when we approximate \[S_i^* = \int_{\mathcal{R}_{i}} w_i(x) S^*(x) \: dx \] and the other is \[S_i^* = \frac{1}{\mathcal{R}_{i}} \int_{\mathcal{R}_{i}} S^*(x) \: dx. \] ## Inference We used Monte Carlo Maximum Likelihood for inference. The likelihood function for this class of model is usually intractible, hence we approximate the likelihood function as \[\frac{1}{N}~ \sum_{j=1}^N~\frac{f(\eta_{(j)}; \psi)}{f(\eta_{(j)}; \psi_0)}.\], where \(\psi\) is the vector of the parameters.

Tutorial

This part illustrates how to fit an SDALGCP model to spatially aggregated data. We used the example dataset that is supplied in the package.

load the package

require(SDALGCP)

load the data

data("PBCshp")

extract the dataframe containing data from the object loaded

data <- as.data.frame(PBCshp@data)

load the population density raster

data("pop_den")

set any population density that is NA to zero

pop_den[is.na(pop_den[])] <- 0

write a formula of the model you want to fit

FORM <- X ~ propmale + Income + Employment + Education + Barriers + Crime + 
  Environment +  offset(log(pop))

Now to proceed to fitting the model, note that there two types of model that can be fitted. One is when approximate the intensity of LGCP by taking the population weighted average and the other is by taking the simple average. We shall consider both cases in this tutorial, starting with population weighted since we have population density on a raster grid of 300m by 300m.

SDALGCP I (population weighted)

Here we estimate the parameters of the model

Discretise the value of scale parameter \(\phi\)

phi <- seq(500, 1700, length.out = 20)

estimate the parameter using MCML

my_est <- SDALGCPMCML(data=data, formula=FORM, my_shp=PBCshp, delta=200, phi=phi, method=1, pop_shp=pop_den, 
                      weighted=TRUE, par0=NULL, control.mcmc=NULL)

To print the summary of the parameter estimates as well as the confidence interval, use;

summary(my_est)
#and for confidence interval use
confint(my_est)

We create a function to compute the confidence interval of the scale parameter using the deviance method. It also provides the deviance plot.

phiCI(my_est, coverage = 0.95, plot = TRUE)

Having estimated the parameters of the model, one might be interested in area-level inference or spatially continuous inference.

  1. If interested in STRICTLY area-level inference use the code below. This can either give either region-specific covariate-adjusted relative risk or region-specific incidence. This is achieved by simply setting in the function.
Dis_pred <- SDALGCPPred(para_est=my_est,  continuous=FALSE)

From this discrete inference one can map either the region-specific incidence or the covariate adjusted relative risk.

#to map the incidence
plot(Dis_pred, type="incidence", continuous = FALSE)
#and its standard error
plot(Dis_pred, type="SEincidence", continuous = FALSE)
#to map the covariate adjusted relative risk
plot(Dis_pred, type="CovAdjRelRisk", continuous = FALSE)
#and its standard error
plot(Dis_pred, type="SECovAdjRelRisk", continuous = FALSE)
#to map the exceedance probability that the incidence is greter than a particular threshold
plot(Dis_pred, type="incidence", continuous = FALSE, thresholds=0.0015)
  1. If interested in spatially continuous prediction of the covariate adjusted relative risk. This is achieved by simply setting in the function.
Con_pred <- SDALGCPPred(para_est=my_est, cellsize = 300, continuous=TRUE)

Then we map the spatially continuous covariate adjusted relative risk.

#to map the covariate adjusted relative risk
plot(Con_pred, type="relrisk")
#and its standard error
plot(Con_pred, type="SErelrisk")
#to map the exceedance probability that the relative risk is greter than a particular threshold
plot(Con_pred, type="relrisk", thresholds=2)

SDALGCP II (Unweighted)

As for the unweighted which is typically by taking the simple average of the intensity an LGCP model, the entire code in the weighted can be used by just setting in the line below.

my_est <- SDALGCPMCML(data=data, formula=FORM, my_shp=PBCshp, delta=200, phi=phi, method=1, 
                      weighted=FALSE,  plot=FALSE, par0=NULL, control.mcmc=NULL, messages = TRUE, plot_profile = TRUE)

Discussion

Using SDALGCP package for analysis of spatially aggregated data provides two main advantages. One, it allows the user to make spatially continous inference irrespective of the level of aggregation of the data. Second, it is more computationally efficient than the lgcp model for aggregated data that was implemented in package.