The objective of this prototype cluster analysis it to perform unsupervised learning to identify distinct and homogeneous market areas in Dallas County with respect to their
characteristics of the underlying census tracts.
The package “ClustGeo” is the workhorse of this analysis. It performs hierarchical cluster analysis using the Ward parametrization of the Lance & Williams algorithm under an imposed spatial contiguity constraint.
Potential improvements could be:
Set the spatial tessellation of the census tracts up and prepare the contiguity space dissimilarity among the census tracts.
data(tractShp) tractShp <- tractShp[tractShp$NIGHTPOP!=0, ] # Remove 2 airport tracts with NA's ## ## Identify spatial neighboring census tracts and process them ## nb <- spdep::poly2nb(tractShp, queen=F) # extract first order neighbors links B <- spdep::nb2mat(nb, style="B") # convert neighbor list to binary matrix ## Visualize first order neighbors plot(tractShp, col="palegreen3", border=grey(0.9), axes=T) plot(nb, coords=coordinates(tractShp), pch=19, cex=0.1, col="blue", add=T) title("Spatial Neighbors Links among Tracts")
## Transform the contiguity matrix into spatial dissimilarity matrix geoDist <- 1-B; diag(geoDist) <- 0; geoDist <- as.dist(geoDist) ## ## Alternative spatial dissimilarity matrices ## # # Calculate shortest path distances between the tracts # BNa <- ifelse(B==0,NA,B) # recode 0's to NA's # allPath <- e1071::allShortestPaths(BNa) # calucate the shortest path among all nodes # topoDist <- as.dist(allPath$length) # number of steps from origin to destination node # e1071::extractPath(allPath, 1, 527) # example: path between tract 1 and tract 527 # # # spherical distances matrix among tracts in km # sphDist <- as.dist(sp::spDists(tractShp, longlat=T))
Set the variables of the census tracts up, impute 25 missing median home values and define the feature space dissimilarity matrix.
## ## Select variables for feature dissimilarity organized into three groups: ## [a] demographic features ## [b] socio-economic features ## [c] housing infrastructure features ## ## Calculate additional features tractShp$PRE1960 <- tractShp$PCTB1950+tractShp$PCTB1940+tractShp$PCTBPRE tractShp$POST2000 <- tractShp$PCTB2000+tractShp$PCTB2010 ## Feature list varKeep <- c("PCTWHITE","PCTBLACK","PCTHISPAN","MEDAGE", "PCTUNIVDEG","HHMEDINC","PCTUNEMP","PCTNOHINS", "POPDEN","PCTDAYPOP","PCTHUVAC","MEDVALHOME","PRE1960","POST2000") xVars <- tractShp@data xVars <- xVars[varKeep] summary(xVars) R> PCTWHITE PCTBLACK PCTHISPAN MEDAGE R> Min. : 5.586 Min. : 0.000 Min. : 0.00 Min. :20.30 R> 1st Qu.: 48.713 1st Qu.: 5.062 1st Qu.:17.13 1st Qu.:29.70 R> Median : 66.040 Median :14.509 Median :34.02 Median :33.10 R> Mean : 62.669 Mean :21.250 Mean :38.36 Mean :34.55 R> 3rd Qu.: 79.745 3rd Qu.:27.923 3rd Qu.:59.53 3rd Qu.:38.60 R> Max. :100.000 Max. :93.397 Max. :93.71 Max. :65.70 R> R> PCTUNIVDEG HHMEDINC PCTUNEMP PCTNOHINS R> Min. : 1.00 Min. : 14754 Min. : 0.000 Min. : 0.00 R> 1st Qu.:11.00 1st Qu.: 39048 1st Qu.: 3.460 1st Qu.:12.10 R> Median :22.00 Median : 51139 Median : 5.393 Median :22.00 R> Mean :30.64 Mean : 62459 Mean : 5.973 Mean :21.61 R> 3rd Qu.:47.00 3rd Qu.: 72096 3rd Qu.: 7.697 3rd Qu.:30.48 R> Max. :91.00 Max. :250001 Max. :25.191 Max. :56.51 R> R> POPDEN PCTDAYPOP PCTHUVAC MEDVALHOME R> Min. : 82.61 Min. :13.96 Min. : 0.000 Min. : 12600 R> 1st Qu.: 2363.97 1st Qu.:26.96 1st Qu.: 4.512 1st Qu.: 93225 R> Median : 3708.51 Median :33.11 Median : 7.355 Median : 133450 R> Mean : 5231.91 Mean :38.77 Mean : 8.230 Mean : 208912 R> 3rd Qu.: 5382.85 3rd Qu.:44.88 3rd Qu.:11.221 3rd Qu.: 238850 R> Max. :80322.40 Max. :99.81 Max. :43.878 Max. :2000001 R> NA's :25 R> PRE1960 POST2000 R> Min. : 0.000 Min. : 0.000 R> 1st Qu.: 2.506 1st Qu.: 3.147 R> Median :11.378 Median : 8.905 R> Mean :22.478 Mean :15.460 R> 3rd Qu.:37.149 3rd Qu.:22.819 R> Max. :92.321 Max. :89.306 R> ## Replace missing values in MEDVALHOME by neighbors average xVars <- DMwR::knnImputation(xVars, k=5, scale=T, meth="weighAve") summary(xVars$MEDVALHOME) R> Min. 1st Qu. Median Mean 3rd Qu. Max. R> 12600 93950 134000 208647 242093 2000001 ## ID each row of xVars row.names(xVars) <- 1:nrow(xVars) ## Calculate feature distance matrix featDist <- dist(scale(xVars))
Identify the \(\alpha\) parameter which balances the convex mixture of the feature space dissimilarities (black curve) with the contiguity space dissimilarities (red curve):
\[(1-\alpha)*FeatureDissimilarity + \alpha*ContiguityDissimilarity\].
Preference should be given to the feature dissimilarity (black) and for neither dissimilarity metric the intra-cluster homogeneity should not drop rapidly for the sequence of \(\alpha\)-values. Low \(\alpha\) values may lead to spatially split clusters.
For this prototype analysis the number of clusters is set to \(K=12\), because \(12\) is the maximum number of factor levels that the function “mapColorQual” can display. A work-around to increase the number of clusters is to split \(K>12\) clusters across several maps of maximal \(12\) factor levels. Please see step 8 below.
The selected mixture parameter \(\alpha\) is set at \(\alpha=0.2\) because for smaller \(\alpha\)-values the intra-cluster homogeneity for spatial contiguity metric drops rapidly.
Show the cluster generation history in a dendrogram and highlight the 12 market areas of census tracts.
Evaluate the number of census tracts per market area.
Map the identified homogeneous market area. An interpretation can be performed by their locations within Dallas County and by the levels and dispersion of their underlying features.
Note that some internally homogeneous clusters of market areas are distributed disjunctly across Dallas County. This implies that areas of similar characteristics can be found in several locations of Dallas County.
Evaluate the unique feature compositions for each market area. A market area should significantly differ from the remaining market areas in its variable profile by at least one feature or a combination of features.
To maintain a sufficient visual resolution, the variable box-plots by market areas are split into two panels.
To map just a set of selected clusters or, alternatively, for more than 12 generated clusters a sequence of maps with consecutive subsets of clusters the remaining clusters are set to NA. Tract with NA’s are displayed in light grey.
Clusters 4 and 10 are predominant employment centers with a proportionally higher number of day-time population relative to the nighttime population.
## Select subset nLevels[c(4,10)] and set remaining clusters to NA nLevels <- levels(neighClus) neighClusSet <- as.factor(ifelse(neighClus %in% nLevels[c(4,10)], neighClus, NA)) mapColorQual(neighClusSet, tractShp, map.title ="Clusters of Major Employment Centers", legend.title="Cluster\nId.", legend.cex=0.9) plot(lakesShp, col="skyblue", border="skyblue",add=T) plot(hwyShp, col="cornsilk2", lwd=4, add=T)