1. Question Asked

Here I asked if there was evidence supporting temporal autocorrelation of age-adjusted Salmonella county rates within Oregon from 2008-2017 and if so what type of correlation structure is most appropriate. I also investigated spatial patterns of reported Salmonella rates as they related to various demographic variables like: Percent of county which is aged 0-4, percent of county which is aged 80+, percent of county which is female, median age, percentage of county residents who have completed high school, median county income, percent of county who is born outside the US, percent of county who speaks a language other than English at home, percentage of county estimated to be in poverty, and percent of children in a county estimated to be in poverty.

To answer these questions, I used the same data outlined in the first exercise blog post with newer demographic variables being taken from the American Community Survey and published on AmericanFactFinder which provides yearly estimates of demographic information at the county level. Unfortunately, yearly data prior to the year 2009 is unavailable which shortened the window of analysis by a year.

2. Names of analytical tools/approaches used

The methods used to answer these questions was first to create an exhaustive general linear model where county Salmonella rates were a function of the above listed demographic variables. A Ljung-Box Test was used to assess if there was evidence of non-zero autocorrelation of residuals of the model at various time lags. Following this, an ideal linear model was selected using a stepwise AIC selection process and then different variance structures were compared by AIC, BIC, and log Likelihood metrics as well as ANOVA testing. Following the selection of an appropriate base model and variance structure, I allowed for interaction between all variables and time and performed more ANOVA testing to select the best model which allowed for variable time interaction. A version of this model would later be used in geographically weighted regression analysis. I performed geographically weighted regression allowing the coefficients to vary across space in Oregon.

3. Description of the analytical process

I created a lattice plot of reported age-adjusted Salmonella rates over the 10-year period in every county to visually assess whether Salmonella rates were changing over time. After seeing that rates of reported Salmonella were changing over time in many Oregon counties I created a “full” linear model in which the rate of reported Salmonella cases in a county was a function of the demographic variables described above. Because this is longitudinal data measured over time I wanted to see if rates of Salmonella were correlated over time, meaning that a county’s rate one year could be predicted from the rate found in another year in the same county. I first performed a Ljung-Box test to assess the need to evaluate for temporal autocorrelation as well as tested normality assumptions, and log-transformed my outcome (Salmonella rates) based on those normality assumption tests. A simple backward step-wise AIC approach was used on my full model to identify and remove extraneous variables. This worked by removing variables in the full model in a stepwise fashion, comparing the AIC values between the two models, and continuing this process until the AIC values between the two models being compared are not significantly different. I then used this model to select an ideal variance structure to compare Salmonella rates autocorrelated at different time lags. The types of variance compared were: Independent variance, compound symmetric, AR1, MA1, ARMA (1,1), unstructured, and allowing for different variance structures across time. After AIC, BIC, log Likelihood, and ANOVA testing an ideal variance structure was determined and the model using this variance structure was evaluated for basic interaction with time. All variables present in the model were allowed to have time interaction including time itself (i.e. allowing for a quadratic time trend). Once again AIC, BIC, log Likelihood, and ANOVA testing were used to select the most ideal model.

Following this I moved on to GWR, where I was able to use the model identified above to create a new data frame containing Beta coefficient estimates of the significant variables in my final model for every county in Oregon. This data frame of coefficients was merged with a spatial data frame containing county level information for all of Oregon. Plots of the coefficient values for every different county were created.

4. Brief Description of Results

Every panel represents one of Oregon’s 36 counties and panels containing an error did not have any cases of reported Salmonella. Some counties are seen decreasing with time, others show slightly increasing trends, and others show a fairly level rate over time. Clearly there is evidence of some time trend for some counties.

Results of normality testing found that age- adjusted rates of reported Salmonella in Oregon were not normally distributed, for the ease of visualization and as an attempt to address the failure to meet the assumption of normality in linear modeling I log-transformed the rates of reported Salmonella. Results of the Ljung-Box test with my full model provided evidence of non-zero time autocorrelation in the data and a visual inspection of the lattice plot supports this with most counties showing a change in rates over time.

The stepwise AIC model selection approach yielded the following model:

logsal ~ %female + %childpov + medianage +year

Covariance structure comparison:

Covariance Model	logLikelihood	AIC	BIC
Independent	-431	874	896
Compound Symmetry	-423	860	887
AR(1)	-427	869	895
MA(1)	-427	869	895
ARMA(1,1)	-423	862	892
Unstructured	-387	859	1017
Compound Symmetry Different Variance Across Time	-412	854	910

 

Mostly AIC, BIC, and Log Likelihood values were clustered together for the different models. I decided to base my choice primarily on the lowest AIC because that’s how I did variable selection to this point. This resulted in me choosing a compound symmetric model which allowed for different variances across time.

Next, I built models which allowed for simple interaction with time meaning that any three-way interaction with time was not evaluated for. Subsequent ANOVA testing comparing the different interaction models to each other, to a model where no interaction was present, and a model where time was absent were used in my selection of a final model.

Final Model:  logsal ~ %female + %childpov + medianage + year +(medianage*year) This model follows a 5×5 compound symmetric correlation variance model which allows for variance to change over time.
Code: interact_m <- gls(logsal ~ female + childpov + medianage + year +(medianage*year), na.action=na.omit, correlation= corCompSymm(form= ~1|county),weights=varIdent(form=~1|year),data=alldata)
Within County Standard Error  (95% CI): 0.92
Estimate Name	Estimate (log-scale)	Std. Error	p-value
Intercept	-759.42	237.79	0.002
% Female	18.06	4.46	<0.001
% Child Poverty	0.03	3.27	0.001
Median Age	17.16	3.11	0.002
Year	0.38	3.18	0.002
Median Age*Year	-0.01	-3.12	0.002

 

Estimates are on the log scale making them difficult to interpret without exponentiation, however it can be seen that a percent change in the number of females in a county or a year change in median age are associated with much larger changes in rates of reported Salmonella incidence compared to changes in percent of child poverty and the year. Overall, incidence rates of reported Salmonella were shown to increase with time, county percentage females, county percentage of child poverty, and county median age with a significant protective interaction effect between time and median age.

For my GWR analysis I used a function derived from the rspatial.org website and looks like:

regfun <- function(x) {
dat <- alldata[alldata$county == x, ]
m <- glm(logsal ~ female + childpov + medianage + year + (medianage*year), data=dat)
coefficients(m)
}

As can be seen, I retained the same significant variables found in the OLS regression for my time series analysis. GWR in this case allows for the coefficient estimates to vary by county.

This allowed me to create a data frame of all coefficient estimates for every county in Oregon. Subsequent dot charts showed the direction and magnitude of all covariates varied across the counties. Plots and dot charts of Oregon for the different coefficient estimates were made.

For % Female

For % Child Poverty

For Median Age

For Year

For Median Age * Year

For the most part, county % female, median age, year, and the interaction term clustered close to 0 for most counties. Some counties were showed highly positive/negative coefficient estimates though no consistently high/low counties could be identified. The maps for the coefficients of median age and year are very similar though I do not have a clear idea why this is the case. The map of the coefficients of child poverty showed the most varied distribution across space. Autocorrelation analysis using Moran’s I of the residuals from this GWR model did not find any evidence of significant autocorrelation. I could not find evidence of a significant non-random spatial pattern for the residuals of my model.

5. Critique of Methods Used

While the temporal autocorrelation analysis was useful in that it provided evidence of temporal autocorrelation present in the data and prior univariate spatial autocorrelation provided limited evidence of variables being spatially autocorrelated at different spatial lags, I was unable to perform cross correlation analysis between my outcome and predictor variables. One important note: I do plan on performing this analysis I just need to figure out how the “ncf” package works in R. This is one of the more glaring shortcomings of my analysis so far is that I do not have evidence that my outcome is correlated with my predictor variables at various distances. Another critique is that the choice of an ideal temporal correlation structure was fairly subjective with my choice of model selection criteria being AIC. Basing my decision on other criteria would likely change my covariance structure. A similar argument could be said for my choice of variable selection being based on a backwards stepwise AIC approach where other selection criteria/methods would likely have different variables in the model.

Finally, the results of my GWR analysis do not show actual drivers of reported Salmonella rates. Rather it shows the demographic characteristics associated with higher reported rates. While this information is useful it does not identify any direct drivers of disease incidence. Further analysis will be required to see if these populations have different exposures or severity of risky exposures.