Questions Being Asked:
How much do different indicator measures of rurality in Texas counties each contribute to a basic index of rurality both statistically and spatially? How much does the weighted basic index differ from a simple calculated mean for rurality?
Tools and Approaches Used
The major statistical method I used to answer these questions was principal component analysis (PCA) via the prcomp function in R and subsequent rurality indicator variable weighting for counties in Texas. The raw indicator data used in this analysis are as follows: Texas county median household income (2010 Census Data), Texas county population density (2010 Census data), and rasterized discrete data of land use for the state of Texas (2011 National Land Cover Database).
Description of Analysis Steps
1) Data Wrangling: Both income and population density data were natively in county-level polygon form, so no conversion was required in ArcGIS. Land use required extensive conversion in order to produce the indicator variable that would both be at the appropriate spatial definition and measure the correct aspect of land use. The raw land use discrete data was first reclassified from many categories of land, such as “deciduous forest,” “barren land,” and “scrub/shrub,” into binary “developed” and “undeveloped” classification groups for the entire state of Texas. Then, the resulting binary raster was converted to unsimplified polygons and the “tabulate intersection” ArcGIS tool was utilized in order to intersect Texas county boundaries with the produced unsimplified polygons. This tool allowed for calculation of percent of area of each Texas county that is considered developed land by the USGS.
2) Data Scaling: After the 3 variables were in analyzable form, each were scaled on a 0 to 100-point scale for visual comparison, PCA analysis, and computation of index weights.
3) Creation of Maps of Scaled Variables for Comparison: ArcGIS was utilized to create maps of each of the scaled variables to visually compare the 3 factors to one another.
4) PCA and Weight Computation: The PCA was completed in R using the prcomp function and variable weights were extracted from the procedure by determining the proportion of variance each variable contributed to. This specific step helps answer how much each indicator variable contributes to rurality in Texas.
5) Simple Means Map, Weighted Index Map, and Statistical Comparison: The scaled variables were used to calculate both an unweighted mean rurality score and weighted mean rurality score for all Texas counties. Maps of each of these measures were produced for visual comparison. A student’s t-test was then completed to compare unweighted and weighted rurality scores for counties.
Results
Maps of Scaled Variables of Texas Counties for Comparison
Rough but similar spatial patterns can be seen in all three maps. Counties with large cities and generally more urbanized counties have higher median income, population density, and percent of land area developed (Green=high, yellow=medium, red=low). There are more extreme values in the population density and percent of land area developed maps than median income.
PCA and Weighting Computation
The biplot of the PCA procedure helps visualize how the samples of the 3 variables relate to one another and how much each variable contributes to the first principal component. Important factors to notice in this plot are the direction of the arrows and clustering of the points. Typically, the more the variable arrow points toward the principal component 1 (PC1) axis, the more it contributes to the proportion of variance. The population density and percent land area developed variables happen to overlap one another in this plot, as they have highly similar contributions to the variance in PC1.
This table shows that population density contributes to 44.4% of the variance in PC1, income contributes to 11.11%, and percent land area developed contributes 44.5%. This PCA procedure indicates that when weighting rurality scores by county in Texas using these 3 variables, percent land area developed should have the highest weight, population density should have the second highest weight, and income should have the lowest. The weighting of variables in the weighted county rurality mean score follows these proportions.
Simple Means Map, Weighted Index Map, and Statistical Comparison
Visually comparing the two maps of rurality scores, there is is some overlap, but quite stark differences. County scores overall have become more rural (lower scores) based on simple visual comparison. Statistical analysis will allow a more nuanced view of the actual differences between the two.
The computed t-test indicates that the weighted index, on average, produces county rurality scores that are significantly more rural (lower scores on the 0-100 scale) than the unweighted index (p<0.001). The overall unweighted rurality score mean for all Texas counties is ~17, while the overall weighted rurality score mean is ~7.5. This indicates that the basic weighted index may better capture the multidimensionality of rurality for Texas counties than unweighted mean scores.
Critique
Based on my understanding, PCA is quite useful for creation of weighted indexes. Unfortunately, the variables I utilized for this analysis have somewhat varying patterns of clustering and possible outliers, which can be seen in the above biplot. These possible outliers could have an effect on the weighting of variables in the index. Due to the small sample size of counties and in order to produce a complete picture of all counties in Texas, I did not want to remove any counties from the analysis. PCA as it relates to index weighting also requires that there is either an a priori understanding or statistical proof of a correlation between variables. Based on the directions of the arrows in the biplot, there is likely correlation between the variables, but more considerations may be needed to confirm this. Also, the scaling of the median income variable is slightly off because the maximum value is over 100. This will need to be corrected in a later analysis.
My future analysis will include more health-related variables in the PCA procedure to determine how health variables contribute to rurality in comparison to the basic rural indicator variables analyzed in this exercise.