Spatial measurement error models

This vignette introduces users to the spatial measurement error (ME) models implemented in the geostan package (Donegan, Chun, and Griffith 2021). Variations on this methodology have been examined previously by multiple authors (Bernadinelli et al. 1997; Kang, Liu, and Cressie 2009; Logan et al. 2020; Xia and Carlin 1998).

These models are designed for spatial models that use survey estimates as covariates. They are designed for use with American Community Survey (ACS) data and other large, government-backed surveys. A premise of this methodology is that the survey includes a systematic spatial sampling design (i.e., the sampling procedure was stratified by the areal unit of interest, whether they be block groups, counties, or states).

In a (spatial) regression analysis, measurement/sampling error in covariates will lead to model parameter estimates that are overly-confident and prone to bias (a terrible combination). This can lead to under- or over-estimation of population risks and needs, mainly because the noisy survey-based covariate passes directly into predictive values. This can impact real communities and service providers (see Bazuin and Fraser 2013).

Spatial measurement error models allow us to incorporate knowledge of data quality and human geography into our analyses, so that the final model results are a better reflection of our state of knowledge than they otherwise would be. These models account for sampling error, but cannot address other sources of bias and error that may be present in survey estimates.

Modeling errors of observation

The measurement error (ME) models implemented in geostan are hierarchical Bayesian models (HBMs) that incorporate two sources of information:

A sampling distribution for the survey estimates.
Generic background knowledge on social variables.

The model treats the true value $\boldsymbol x$ as an unknown parameter or a latent variable. The goal is to obtain a probability distribution for $\boldsymbol x$. That information can then be incorporated into any of geostan’s regression or disease mapping models.

To represent the general social and technical knowledge on which this model is premised, we write $\mathcal{M}$. This background knowledge includes, for example, that the data came from the valid small-area spatial sampling design.

The sampling distribution (a likelihood statement) states the following: if the true median income value is $x_i$, then the probability of obtaining survey estimate $z_i$ is \[ p(z_i | x_i , \mathcal{M}) = Gauss(z_i | x_i, s_i) \] where $s_i$ is the standard error of the estimate. This is saying, more directly, that the true value $x_i$ is probably within two standard errors of the estimate.

Our background knowledge on $x$ is the second component of the model. This is simply the knowledge that social variables tend to be spatially patterned. Specifically, extreme values are not implausible, although they tend to be clustered together.

This information is encoded into a prior probability distribution for the unknown values $\boldsymbol x$, using the spatial conditional autoregressive (CAR) model: \[ p(\boldsymbol x | \mathcal{M}) = Gauss(\boldsymbol x | \mu \boldsymbol 1, \boldsymbol \Sigma) \] where the bold symbols are used to indicate an object is a column vector or matrix.

The equation above states that the unknown values $\boldsymbol x$ are spatially correlated, with a mean value of $\mu$ and covariance matrix \[\boldsymbol \Sigma = (I - \rho C)^{-1} M,\] where $\rho$ is a spatial autocorrelation parameter, $I$ is the $n \times n$ identity matrix, and $C$ is a spatial connectivity matrix. $M$ is an $n \times n$ diagonal matrix; its diagonal contains the row-sums of $C$ multiplied by a scale parameter $\tau$ (Donegan 2022, Table 1 WCAR specification).

If $\boldsymbol x$ is a rate variable—e.g., the poverty rate—it can be important to use a logit-transformation (Logan et al. 2020) as follows: \[ p(logit(\boldsymbol x) | \mathcal{M}) = Gauss(logit(\boldsymbol x) | \mu \boldsymbol 1, \boldsymbol \Sigma) \]

The CAR parameters $\mu$, $\rho$, and $\tau$ all require prior probability distributions; geostan uses the following by default:

\[\begin{equation} \begin{split} \mu &\sim Gauss(0, 100) \\ \tau &\sim Student_t(10, 0, 40) \\ \rho &\sim Uniform(\text{lower_bound}, \text{upper_bound}) \end{split} \end{equation}\]

The default prior for $\rho$ is uniform across its entire support (determined by the extreme eigenvalues of $C$)

Discussion

For a sense of how all of this information gets combined in the model, consider two hypothetical survey estimates.

The first has a small standard error—it is a reliable data point—and it is also similar to its surrounding counties. In that case, the probability distribution for $x_i$ is going to look similar to the raw estimate and its standard error. Now consider a second county, where the estimate has a large standard error—it is unreliable—and the estimate is also dissimilar from its neighbors. In this case, we ought to consider it quite probable that the estimate $x_i$ is far from the truth. This means that the probability distribution for $x_i$ that comes from our model may be quite different from the raw estimate. That shift is often referred to as “shrinkage,” although the shift may be in either direction (up or down).

If a non-spatial prior model for $\boldsymbol x$ were used instead of the CAR model, then there would be a tendency for all survey estimates $z_i$ that are far from the mean to be shrunk towards the ‘global mean’, instead of the local area mean. This can lead to blatant social biases in ME models when they are applied to spatial data (Donegan, Chun, and Griffith 2021). For that and other reasons, the spatial CAR model is often essential for spatial ME models, and it remains important to graphically evaluate results (see below).

When we add a spatial ME model to a spatial regression model, the model results will automatically average over the uncertain inferences we make about $\boldsymbol x$. The inferential biases that follow from ignoring measurement/observational uncertainty will be addressed, meaning that the probability distributions for the parameters will reflect the information we have on data quality. It is important to understand that there may still be un-modeled forms of bias and error in the ACS estimates—due to response bias and other sources—beyond the amount that is attributed to sampling error.

Getting started

From the R console, load the geostan package.

library(geostan)
data(georgia)

The line data(georgia) loads the georgia data set from the geostan package into your working environment. You can learn more about the data by entering ?georgia to the R console.

Preparing the data

Users can add a spatial ME model to any geostan model. A list of data for the ME models can be prepared using the prep_me_data function. The function require at least two inputs from the user:

se A data frame with survey standard errors.
car_parts A list of data created by the prep_car_data function.

It also has optional inputs:

bounds A length-two vector giving the minimum and maximum values that the covariate may take on.
prior A list of prior distributions for the CAR model parameters.
logit A vector of logical values (TRUE or FALSE) specifying if the variate should be logit transformed. Only use this for rates (between 0 and 1).

The default prior distributions are designed to be fairly vague relative to ACS variables (including percentages from 0-100), assuming that magnitudes such as income have been log-transformed already.

The logit-transformation is often required for skewed rates, such as the poverty rate. Note that the user can not do this transformation on their own, they must use the logit argument in prep_me_data.

The bounds argument may be needed for rates, which are bounded by zero and one. Even if the logit argument is used, the sampling distribution still needs to be bounded. This functionality has limitations. If the bounds argument is used, it will apply to all of the modeled covariates (which might not make sense at all). It also causes the model to slow down. If none of the estimates ($\pm$ two standard errors or so) approach the boundaries, then the bounds argument can be safely ignored. This argument should be used if needed, but avoided when possible.

To demonstrate, take the percent of the population with health insurance coverage georgia$insurance. The estimates are stored as percentages; we are going to divide them and their standard errors by 100 to make them rates:

georgia$insurance <- georgia$insurance / 100
georgia$insurance.se <- georgia$insurance.se / 100

We want to check if any of the estimates are near their maximum and minimum values (zero and one):

max(georgia$insurance + georgia$insurance.se*2)

## [1] 0.9615319

min(georgia$insurance - georgia$insurance.se*2)

## [1] 0.6686201

None of the estimates are quite pushing against their maximum value (1), and none are anywhere near the minimum (0), so we can ignore the boundary constraint for now.

Now we need to gather the standard errors georgia$insurance.se into a data frame. The column name must match the name of the variable it refers to (“insurance”):

SE <- data.frame(insurance = georgia$insurance.se)

If we had additional survey-based covariates that we were intending to model, we would simply add their standard errors to this data frame.

To prepare the CAR data, we create a binary spatial connectivity matrix, and then pass it to prep_car_data to prepare our list of data for the CAR model:

C <- shape2mat(georgia, "B")
cars <- prep_car_data(C)

## Range of permissible rho values:  -1.661134 1

Now we use prep_me_data. We are going to use the logit-transform, since we are working with rates:

ME_list <- prep_me_data(se = SE,
                        car_parts = cars,
                        logit = TRUE
                        )

Now we are ready to begin modeling the data.

Spatial ME model

The following code chunk sets up a spatial model for male county mortality rates (ages 55-64) using the insurance variable as a covariate (without the spatial ME model):

fit <- stan_car(log(rate.male) ~ insurance, 
                data = georgia, 
                family = auto_gaussian(),
                car_parts = cars)

In practice, a Poisson model should be used to model rates like this. This little example is only for the purpose of demonstrating the ME models.

To add the spatial ME model to the above specification, we simply provide our ME_list to the ME argument:

fit <- stan_car(log(rate.male) ~ insurance, 
                ME = ME_list,
                family = auto_gaussian(),
                data = georgia, 
                car_parts = cars,
                iter = 650, # for demo speed 
                refresh = 0, # minimizes printing
                )

##   location scale
## 1     -4.2     5
##   location scale
## 1        0    15
##   df location scale
## 1 10        0     3
##   lower upper
## 1  -1.7     1

The CAR models in geostan can be highly efficient in terms of MCMC sampling, but the required number of iterations will depend on many characteristics of the model and data. Often the default iter = 2000 is more than sufficient (with cores = 4). To speep up sampling with multi-core processing, use cores = 4 (to sample 4 chains in parallel).

Visual diagnostics

It is important to check the models graphically. The me_diag provides some useful diagnostics for understanding how the raw estimates compare with the modeled values.

First, it provides a point-interval scatter plot that compares the raw survey estimates (on the horizontal axis) to the modeled values (on the vertical axis). The probability distributions for the modeled values are represented by their 95% credible intervals. The intervals provide a sense of the quality of the ACS data—if the credible intervals on the modeled value are large, this tells us that the data are not particularly reliable. Data quality places a limit on the kinds of inferences that can possibly be made, especially with multivariate data.

The me_diag function also provides more direct visual depictions of the difference between the raw survey estimates $z_i$ an the modeled values $x_i$: \[\delta_i = z_i - x_i.\] We want to look for any strong spatial pattern in these $\delta_i$ values, because that would be an indication of a bias. However, the magnitude of the $\delta_i$ value is important to consider—there may be a pattern, but if the amount of shrinkage is very small, that pattern may not matter. (The model returns $M$ samples from the posterior distribution of each $x_i$; or, indexing by MCMC sample $x_i^m$ ($m$ is an index, not an exponent). The reported $\delta_i$ values is the MCMC mean $\delta_i = \frac{1}{M} \sum_m z_i - x_i^m$.)

Two figures are provided to evaluate patterns in $\delta_i$: first, a Moran scatter plot and, second, a map.

me_diag(fit, 'insurance', georgia)

In this case, the results do not look too concerning insofar as there are no conspicuous patterns. However, a number of the credible intervals on the modeled values are large, which indicates that this data is not of high quality. The fact that some of the $\delta_i$ values are substantial also points to low data quality.

It can be important to understand that the ME models are not independent from the rest of the model. When the modeled covariate $\boldsymbol x$ enters a regression relationship, that regression relationship can impact the probability distribution for $\boldsymbol x$ itself. (This is a feature of all Bayesian ME models). To examine the ME model in isolation, you can set the prior_only argument to TRUE (see ?stan_glm or any of the spatial models).

Working with MCMC samples from ME models

geostan consists of pre-compiled Stan models, and users can always access the Markov chain Monte Carlo (MCMC) samples returned by Stan. When extracted as a matrix of samples (as below), each row of the matrix represents a draw from the joint probability distribution for all model parameters, and each column consists of samples from the marginal distribution of each parameter.

The ME models return samples for $x_i$ as well as the model parameters $\mu$ (“mu_x_true”), $\rho$ (“car_rho_x_true”), and $\tau$ (“sigma_x_true”). We can access these using as.matrix (also as.array and as.data.frame).

mu.x <- as.matrix(fit, pars = "mu_x_true")
dim(mu.x)

## [1] 1300    1

head(mu.x)

##           parameters
## iterations mu_x_true[1]
##       [1,]     1.922295
##       [2,]     1.754260
##       [3,]     1.826907
##       [4,]     1.813635
##       [5,]     1.722004
##       [6,]     1.721516

mean(mu.x)

## [1] 1.787654

We can visualize these using plot or print a summary:

print(fit$stanfit, pars = c("mu_x_true", "car_rho_x_true", "sigma_x_true"))

## Inference for Stan model: foundation.
## 4 chains, each with iter=650; warmup=325; thin=1; 
## post-warmup draws per chain=325, total post-warmup draws=1300.
## 
##                   mean se_mean   sd 2.5%  25%  50%  75% 97.5% n_eff Rhat
## mu_x_true[1]      1.79       0 0.07 1.64 1.74 1.79 1.83  1.93   794    1
## car_rho_x_true[1] 0.91       0 0.06 0.76 0.87 0.92 0.96  0.99   898    1
## sigma_x_true[1]   0.48       0 0.04 0.42 0.46 0.48 0.51  0.56  1139    1
## 
## Samples were drawn using NUTS(diag_e) at Mon Jul 11 13:28:34 2022.
## For each parameter, n_eff is a crude measure of effective sample size,
## and Rhat is the potential scale reduction factor on split chains (at 
## convergence, Rhat=1).

To extract samples from the joint probability distribution for $\boldsymbol x$, use the generic parameter name “x_true”:

x <- as.matrix(fit, pars = "x_true")
dim(x)

## [1] 1300  159

If we wanted to calculate the mean of each of these marginal distributions (one for every $x_i$), we could use apply with MARGIN = 2 to summarize by column:

x.mu <- apply(x, 2, mean)
head(x.mu)

## x_insurance[1] x_insurance[2] x_insurance[3] x_insurance[4] x_insurance[5] 
##      0.8369614      0.8001021      0.8516808      0.8520926      0.9179429 
## x_insurance[6] 
##      0.8700983

The vector x.mu contains estimates (posterior means) for $x_i$. We might want to use these to plot the residuals or fitted values against the predictor:

rs <- resid(fit)$mean
plot(x.mu, rs)
abline(h = 0)