Some inverse distance weighting hacks – using R and spatstat

For a recent project I was mapping survey responses to attitudes towards the police, and I wanted to make a map of those responses. The typical default to accomplish this is inverse distance weighting. For those familiar with hot spot maps of crime, this is similar in that is produces a smooth isarithmic map, but instead of being a density it predicts values. For my project I wanted to explore two different things; 1) estimating the variance of the IDW estimate, and 2) explore different weighting schemes besides the default inverse distance. The R code for my functions and data for analysis can be downloaded here.

What is inverse distance weighting?

Since this isn’t typical fodder for social scientists, I will present a simple example to illustrate.

Imagine you are a farmer and want to know where to plant corn vs. soy beans, and are using the nitrogen content of the soil to determine that. You take various samples from a field and measure the nitrogen content, but you want predictions for the areas you did not sample. So say we have four measures at various points in the field.

Nit     X   Y
1.2     0   0
2.1     0   5
2.6    10   2
1.5     6   5

From this lets say we want to estimate average nitrogen content at the center, 5 and 5. Inverse distance weighting is just as the name says, the weight to estimate the average nitrogen content at the center is based on the distance between the sample point and the center. Most often people use the distance squared as the weight. So from this we have as the weights.

Nit     X   Y Weight
1.2     0   0   1/50
2.1     0   5   1/25
2.6    10   2   1/34
1.5     6   5   1/ 1

You can see the last row is the closest point, so gets the largest weight. The weighted average of nitrogen for the 5,5 point ends up being ~1.55.

For inverse distance weighted maps, one then makes a series of weighted estimates at a regular grid over the study space. So not just an estimate at 5,5, but also 5,4|5,3|5,2 etc. And then you have a regular grid of values you can plot.

Example – Street Clean Scores in LA

An ok example to demonstrate this is an LA database rating streets based on their cleanliness. Some might quibble about it only makes sense to estimate street cleanliness values on streets, but I think it is ok for exploratory data analysis. Just visualizing the streets is very hard given their small width and irregularity.

So to follow along, first I load all the libraries I will be using, then set my working directory, and finally import my updated inverse distance weighted hacked functions I will be using.


MyDir <- "C:\\Users\\axw161530\\Dropbox\\Documents\\BLOG\\IDW_Variance_Bisquare\\ExampleAnalysis"

#My updated idw functions

Next we need to create an point pattern object spatstat can work with, so we import our street scores that contain an X and Y coordinate for the midpoint of the street segment, as well as the boundary of the city of Los Angeles. Then we can create a marked point pattern. For reference, the street scores can range from 0 (clean) to a max of 3 (dirty).

CleanStreets <- read.csv("StreetScores.csv",header=TRUE)
BorderLA <- readOGR("CityBoundary.shp", layer="CityBoundary")

#create Spatstat object and window
LA_Win <- as.owin(BorderLA)
LA_StreetPP <- ppp(CleanStreets$XMidPoint,CleanStreets$YMidPoint, window=LA_Win, marks=CleanStreets$StreetScor)

Now we can estimate a smooth inverse distance weighted map by calling my new function, idw2. This returns both the original weighted mean (equivalent to the original spatstat idw argument), but also returns the variance. Here I plot them side by side (see the end of the blog post on how I calculate the variance). The weighted mean is on the left, and the variance estimate is on the right. For the functions the rat image is the weighted mean, and the var image is the weighted variance.

#Typical inverse distance weighted estimate
idw_res <- idw2(LA_StreetPP) #only takes a minute
plot(idw_res$rat) #this is the weighted mean
plot(idw_res$var) #this is the weighted variance

So contrary to expectations, this does not provide a very smooth map. It is quite rough. This is partially because social science data is not going to be as regular as natural science measurements. In spatial stats jargon street to street measures will have a large nugget – a clean street can be right next to a dirty one.

Here the default is using inverse distance squared – what if we just use inverse distance though?

#Inverse distance (linear)
idw_Lin <- idw2(LA_StreetPP, power=1)

This is smoothed out a little more. There is essentially one dirty spot in the central eastern part of the city (I don’t know anything about LA neighborhoods). Compared to the first set of maps, the dirty streets in the northern mass of the city are basically entirely smoothed out, whereas before you could at least see little spikes.

So I was wondering if there could maybe be better weights we could choose to smooth out the data a little better. One I have used in a few recent projects is the bisquare kernel, which I was introduced by the geographically weighted regression folks. The bisquare kernel weight equals [1 - (d/b)^2]^2, when d < b and zero otherwise. Here d is the distance, and b is a user chosen distance threshold. We can make a plot to illustrate the difference in weight functions, here using a bisquare kernel distance of 2000 meters.

#example weight functions over 3000 meters
dist <- 1:3000
idw1 <- 1/dist
idw2 <- 1/(dist^2)
b <- 2000
bisq <- ifelse(dist < b, ( 1 - (dist/b)^2 )^2, 0)

Here you can see both of the inverse distance weighted lines trail to zero almost immediately, whereas the bisquare kernel trails off much more slowly. So lets check out our maps using a bisquare kernel with the distance threshold set to 2000 meters. The biSqW function is equivalent to the original spatstat idw function, but uses the bisquare kernel and returns the variance estimate as well. You just need to pass it a distance threshold for the b_dist parameter.

#BiSquare weighting, 2000 meter distance
LA_bS_w <- biSqW(LA_StreetPP, b_dist=2000)

Here we get a map that looks more like a typical hot spot kernel density map. We can see some of the broader trends in the northern part of the city, and even see a really dirty hot spot I did not previously notice in the northeastern peninsula.

The 2,000 meter distance threshold was just ad-hoc though. How large or small should it be? A quick check of the spatial correlogram is one way to make it slightly more objective. Here I use the correlog function in the ncf package to estimate this. I subsample the data first (I presume it has a call to dist somewhere).

#correleogram, random sample, it is too big
subSamp <- CleanStreets[sample(nrow(CleanStreets), 3000), ]
fit <- correlog(x=subSamp$XMidPoint,y=subSamp$YMidPoint,z=subSamp$StreetScor, increment=100, resamp=0, quiet=TRUE)

Here we can see points very nearby each other have a correlation of 0.2, and then this trails off into zero before 20 kilometers (the distances here are in meters). FYI the rising back up in correlation for very large distances often occurs for data that have broader spatial trends.

So lets try out a bisquare kernel with a distance threshold of 10 kilometers.

#BiSquare weighting, 10000 meter distance
LA_bS_w <- biSqW(LA_StreetPP, b_dist=10000)

That is now a bit oversmoothed. But it allows a nicer range of potential values, as oppossed to simply sticking with the inverse distance weighting.

A few notes on the variance of IDW

So I hacked the idw function in the spatstat package to return the variance of the estimate as well as the actual weighted mean. This involved going into the C function, so I use the inline package to create my own version. Ditto for creating the maps using the bisquare weights instead of inverse distance weighting. To quick see those functions here is the R code.

Given some harassment on Crossvalidated by Mark Stone, I also updated the algorithm to be a more numerically safe one, both for the weighted mean and the weighted variance. Note though that that Wikipedia article has a special definition for the variance. The correct Bessel correction for weighted data though (in this case) is the sum of the weights (V1) minus the sum of square of the weights (V2) divided by V1. Here I just divide by V1, but that could easily be changed (not sure if in the sum of squares I need to worry about underflow). I.e. change the line MAT(var, ix, iy, Ny) = m2 / sumw; to MAT(var, ix, iy, Ny) = m2 / (sumw - sumw/sumw2); in the various C calls.

Someone should also probably write in a check to prevent distances of zero. Maybe by capping the weights to never be above a certain value, although that is not trivial what the default top value should be. (If you have data on the unit square weights above 1 would occur quite regularly, but for a large city like this projected in meters capping the weight at 1 would be fine.)

In general these variance maps did not behave like I expected them to, either with this or other data. When using Bessel’s correction they tended to look even weirder. So I would need to explore some more before I go and recommend them. Probably should not waste more time on this though, and just fit an actual kriging model though to produce the standard error of the estimates.

Added code snippets page

I’ve written quite a few blog posts over the years, and it is getting to be hard for me to keep all of them orderly. So I recently created a page with a simple table of contents with code snippets broken down into categories. There are also a few extra SPSS macros in there that I have not written blog posts about.

Every now and then I see a broken link and update it, but I know there are more I am missing. So just send an email if I have a link to some of my code and it is currently broken.


The spatial clustering of hits-vs-misses in shootings using Ripley’s K

My article, Replicating Group-Based Trajectory Models of Crime at Micro-Places in Albany, NY was recently published online first at the Journal of Quantitative Criminology (here is a pre-print version, and you can get my JQC offprint for the next few weeks).

Part of the reason I blog is to show how to replicate some of my work. I have previously shown how to generate the group based trajectory models (1,2), and here I will illustrate how to replicate some of the point pattern analysis I conducted in that paper.

A regular task for an analyst is to evaluate spatial clustering. A technique I use in that article is to use Ripley’s K statistic to evaluate clustering between different types of events, or what spatial statistics jargon calls a marked point pattern. I figured I would illustrate how to do this on some example point patterns of shootings, so analysts could replicate for other situations. The way crime data is collected and geocoded makes generating the correct reference distributions for the tests different than if the points could occur anywhere in the study region.

An example I am going to apply are shooting incidents that have marks of whether they hit the intended victim or missed. One theory in criminology is that murder is simply the extension of general violence — with the difference in aggravated assault versus murder often being happenstance. One instance this appears to be the case from my observations are shootings. It seems pure luck whether an individual gets hit or the bullets miss everyone. Here I will see if there appear to be any spatial patterning in shots fired with a victim hit. That is, we know that shootings themselves are clustered, but within those clusters of shootings are the locations of hits-and-misses further clustered or are they random?

A hypothetical process in which clustering of shooting hits can occur is if some shootings are meant to simply scare individuals vs. being targeted at people on the street. It could also occur if there are different tactics, like drive-bys vs. being on foot. This could occur if say one area had many shootings related to drug deals gone wrong, vs. another area that had gang retaliation drive by shootings. If they are non-random, the police may consider different interventions in different places – probably focusing on locations where people are more likely to be injured from the shooting. If they are random though, there is nothing special about analyzing shootings with a person hit versus shootings that missed everyone – you might as well analyze and develop strategy for all shootings.

So to start we will be using the R statistical package and the spatstat library. To follow along I made a set of fake shooting data in a csv file. So first load up R, I then change the working directory to where my csv file is located, then I read in the csv into an R data frame.


#change R directory to where your file is
MyDir <- "C:\\Users\\andrew.wheeler\\Dropbox\\Documents\\BLOG\\R_shootingVic\\R_Code"

#read in shooting data
ShootData <- read.csv("Fake_ShootingData.csv", header = TRUE)

The file subsequently has four fields, ID, X & Y coordinates, and a Vic column with 0’s and 1’s. Next to conduct the spatial analysis we are going to convert this data into an object that the spatstat library can work with. To do that we need to create a study window. Typically you would have the outline of the city, but for this analysis the window won’t matter, so here I make a window that is just slightly larger than the bounding box of the data.

#create ppp file, window does not matter for permuation test
StudyWin <- owin(xrange=c(min(ShootData$X)-0.01,max(ShootData$X)+0.01),yrange=c(min(ShootData$Y)-0.01,max(ShootData$Y)+0.01))
Shoot_ppp <- ppp(ShootData$X, ShootData$Y, window=StudyWin, marks=as.factor(ShootData$Vic), unitname = c("meter","meters"))

Traditionally when Ripley’s K was originally developed it was for points that could occur anywhere in the study region, such as the location of different tree species. Crimes are a bit different though, in that there are some areas in any city that a crime basically cannot occur (such as the middle of a lake). Also, crimes are often simply geocoded according to addresses and intersections, so this further reduces the locations where the points can be located in a crime dataset. To calculate the sample statistic for Ripley’s K one does not need to account for this, but to tell whether those patterns are random one needs to simulate the point pattern under a realistic null hypothesis. There are different ways to do the simulation, but here the simulation keeps the shooting locations fixed, but randomly assigns a shooting to be either a victim or a not with the same marginal frequencies. That is, it basically just shuffles which events are counted as a victim being hit or one in which there were no people hit. The Dixon article calls this the random relabelling approach.

Most of the spatstat functions can take a separate list of point patterns to use to simulate error bounds for different functions, so this function takes the initial point pattern, generates the permutations, and stuffs them in a list. I set the seed so the analysis can be reproduced exactly.

#generate the simulation envelopes to use in the Cross function
MarkedPerms <- function(ppp, nlist) {
  myppp_list <- c() #make empty list
  for (i in 1:nlist) {
    current_ppp <-  ppp(ppp$x, ppp$y,window=ppp$window,marks=sample(ppp$marks))  #making permutation      
    myppp_list[[i]] <- current_ppp                                               #appending perm to list

#now making a set of simulated permutations
set.seed(10) #setting seed for reproducibility
MySimEvel <- MarkedPerms(ppp=Shoot_ppp,nlist=999)

Now we have all the ingredients to conduct the analysis. Here I call the cross K function and submit my set of simulated point patterns named MySimEvel. With only 100 points in the dataset it works pretty quickly, and then we can graph the Ripley’s K function. The grey bands are the simulated K statistics, and the black line is the observed statistic. We can see the observed is always within the simulated bands, so we conclude that conditional on shooting locations, there is no clustering of shootings with a victim versus those with no one hit. Not surprising, since I just simulated random data.

#Cross Ripleys K
CrossK_Shoot <- alltypes(Shoot_ppp, fun="Kcross", envelope=TRUE, simulate=MySimEvel)
plot(CrossK_Shoot$fns[[2]], main="Cross-K Shooting Victims vs. No Victims", xlab="Meters")

I conducted this same analysis with actual shooting data in three separate cities that I have convenient access to the data. It is a hod podge of length, but City A and City B have around 100 shootings, and City C has around 500 shootings. In City A the observed line is very near the bottom, suggesting some evidence that shootings victims may be further apart than would be expected, but for most instances is within the 99% simulation band. (I can’t think of a theoretical reason why being spread apart would occur.) City B is pretty clearly within the simulation band, and City C’s observed pattern mirrors the mean of the simulation bands almost exactly. Since City C has the largest sample, I think this is pretty good evidence that shootings with a person hit are spatially random conditional on where shootings occur.

Long story short, when conducting Ripley’s K with crime data, the default way to generate the simulation envelopes for the statistics are not appropriate given how crime data is recorded. I show here one way to account for that in generating simulation envelopes.

One sided line buffers in R using rgeos

I’ve started to do more geographic data manipulation in R, and part of the reason I do blog posts is for self-reference, so I figured I would share some of the geographic functions I have been working on.

The other day on StackOverflow there was a question that asked how to do one sided buffers in R. The question was closed (and the linked duplicate is closed as well), so I post my response here.

The workflow I describe to make one sided buffers is in a nutshell

  • expand the original polyline into a very small area by using a normal buffer, calls this Buf0
  • do a normal two sided buffer on the original polyline, without square or rounded ends, call this Buf1
  • take the geographic difference between Buf1 and Buf0, which basically splits the original buffer into two parts, and then return them as two separate spatial objects

Here is a simple example.


TwoBuf <- function(line,width,minEx){
  Buf0 <- gBuffer(line,width=minEx,capStyle="SQUARE")
  Buf1 <- gBuffer(line,width=width,capStyle="FLAT")

Squig <- readWKT("LINESTRING(0 0, 0.2 0.1, 0.3 0.6, 0.4 0.1, 1 1)") #Orig
TortBuf <- TwoBuf(line=Squig,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue'))  #First object on left, second on right

If you imagine travelling along the polyline, which in this example goes from left to right, this is how I know the first red polygon is the left hand and the blue is the right hand side buffer. (To pick a specific one, you can subset like TortBuf[1] or TortBuf[2].)

If we reverse the line string, the order will subsequently be reversed.

SquigR <- readWKT("LINESTRING(1 1, 0.4 0.1, 0.3 0.6, 0.2 0.1, 0 0)") #Reversed
TortBuf <- TwoBuf(line=SquigR,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue'))  #Again first object on left, second on right

Examples of south to north and north to south work the same as well.

SquigN <- readWKT("LINESTRING(0 0, 0 1)") #South to North
TortBuf <- TwoBuf(line=SquigN,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue'))  #Again first object on left, second on right

SquigS <- readWKT("LINESTRING(0 1, 0 0)") #North to South
TortBuf <- TwoBuf(line=SquigS,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue'))  #Again first object on left, second on right

One example in which this procedure does not work is if the polyline creates other polygons.

Square <- readWKT("LINESTRING(0 0, 1 0, 1 1, 0 1, 0 0)") #Square
TortBuf <- TwoBuf(line=Square,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue','green'))

#Switch the direction
SquareR <- readWKT("LINESTRING(0 0, 0 1, 1 1, 1 0, 0 0)") #Square Reversed
TortBuf <- TwoBuf(line=SquareR,width=0.05,minEx=0.0001)
plot(TortBuf, col=c('red','blue','green'))                #Still the same order

This messes up the order as well. If you know that your polyline is actually a polygon you can do a positive and negative buffer to get the desired effect of interest. If I have a need to expand this to multipart polylines I will post an update, but I have some other buffer functions I may share in the mean time.

Some plots to go with group based trajectory models in R

On my prior post on estimating group based trajectory models in R using the crimCV package I received a comment asking about how to plot the trajectories. The crimCV model object has a base plot object, but here I show how to extract those model predictions as well as some other functions. Many of these plots are illustrated in my paper for crime trajectories at micro places in Albany (forthcoming in the Journal of Quantitative Criminology). First we are going to load the crimCV and the ggplot2 package, and then I have a set of four helper functions which I will describe in more detail in a minute. So run this R code first.


long_traj <- function(model,data){
  df <- data.frame(data)
  vars <- names(df)
  prob <- model['gwt'] #posterior probabilities
  df$GMax <- apply(prob$gwt,1,which.max) #which group # is the max
  df$PMax <- apply(prob$gwt,1,max)       #probability in max group
  df$Ord <- 1:dim(df)[1]                 #Order of the original data
  prob <- data.frame(prob$gwt)
  names(prob) <- paste0("G",1:dim(prob)[2]) #Group probabilities are G1, G2, etc.
  longD <- reshape(data.frame(df,prob), varying = vars, v.names = "y", 
                   timevar = "x", times = 1:length(vars), 
                   direction = "long") #Reshape to long format, time is x, y is original count data
  return(longD)                        #GMax is the classified group, PMax is the probability in that group

weighted_means <- function(model,long_data){
  G_names <- paste0("G",1:model$ng)
  G <- long_data[,G_names]
  W <- G*long_data$y                                    #Multiple weights by original count var
  Agg <- aggregate(W,by=list(x=long_data$x),FUN="sum")  #then sum those products
  mass <- colSums(model$gwt)                            #to get average divide by total mass of the weight
  for (i in 1:model$ng){
    Agg[,i+1] <- Agg[,i+1]/mass[i]
  long_weight <- reshape(Agg, varying=G_names, v.names="w_mean",
                         timevar = "Group", times = 1:model$ng, 
                         direction = "long")           #reshape to long
pred_means <- function(model){
    prob <- model$prob               #these are the model predicted means
    Xb <- model$X %*% model$beta     #see getAnywhere(plot.dmZIPt), near copy
    lambda <- exp(Xb)                #just returns data frame in long format
    p <- exp(-model$tau * t(Xb))
    p <- t(p)
    p <- p/(1 + p)
    mu <- (1 - p) * lambda
    t <- 1:nrow(mu)
    myDF <- data.frame(x=t,mu)
    long_pred <- reshape(myDF, varying=paste0("X",1:model$ng), v.names="pred_mean",
                         timevar = "Group", times = 1:model$ng, direction = "long")

occ <- function(long_data){
 subdata <- subset(long_data,x==1)
 agg <- aggregate(subdata$PMax,by=list(group=subdata$GMax),FUN="mean")
 names(agg)[2] <- "AvePP" #average posterior probabilites
 agg$Freq <-$GMax))[,2]
 n <- agg$AvePP/(1 - agg$AvePP)
 p <- agg$Freq/sum(agg$Freq)
 d <- p/(1-p)
 agg$OCC <- n/d #odds of correct classification
 agg$ClassProp <- p #observed classification proportion
 #predicted classification proportion
 agg$PredProp <- colSums(as.matrix(subdata[,grep("^[G][0-9]", names(subdata), value=TRUE)]))/sum(agg$Freq) 

Now we can just use the data in the crimCV package to run through an example of a few different types of plots. First lets load in the TO1adj data, estimate the group based model, and make our base plot.

out1 <-crimCV(TO1adj,4,dpolyp=2,init=5)

Now most effort seems to be spent on using model selection criteria to pick the number of groups, what may be called relative model comparisons. Once you pick the number of groups though, you should still be concerned with how well the model replicates the data at hand, e.g. absolute model comparisons. The graphs that follow help assess this. First we will use our helper functions to make three new objects. The first function, long_traj, takes the original model object, out1, as well as the original matrix data used to estimate the model, TO1adj. The second function, weighted_means, takes the original model object and then the newly created long_data longD. The third function, pred_means, just takes the model output and generates a data frame in wide format for plotting (it is the same underlying code for plotting the model).

longD <- long_traj(model=out1,data=TO1adj)
x <- weighted_means(model=out1,long_data=longD)
pred <- pred_means(model=out1)

We can subsequently use the long data longD to plot the individual trajectories faceted by their assigned groups. I have an answer on cross validated that shows how effective this small multiple design idea can be to help disentangle complicated plots.

#plot of individual trajectories in small multiples by group
p <- ggplot(data=longD, aes(x=x,y=y,group=Ord)) + geom_line(alpha = 0.1) + facet_wrap(~GMax)

Plotting the individual trajectories can show how well they fit the predicted model, as well as if there are any outliers. You could get more fancy with jittering (helpful since there is so much overlap in the low counts) but just plotting with a high transparency helps quite abit. This second graph plots the predicted means along with the weighted means. What the weighted_means function does is use the posterior probabilities of groups, and then calculates the observed group averages per time point using the posterior probabilities as the weights.

#plot of predicted values + weighted means
p2 <- ggplot() + geom_line(data=pred, aes(x=x,y=pred_mean,col=as.factor(Group))) + 
                 geom_line(data=x, aes(x=x,y=w_mean,col=as.factor(Group))) + 
                geom_point(data=x, aes(x=x,y=w_mean,col=as.factor(Group)))

Here you can see that the estimated trajectories are not a very good fit to the data. Pretty much eash series has a peak before the predicted curve, and all of the series except for 2 don’t look like very good candidates for polynomial curves.

It ends up that often the weighted means are very nearly equivalent to the unweighted means (just aggregating means based on the classified group). In this example the predicted values are a colored line, the weighted means are a colored line with superimposed points, and the non-weighted means are just a black line.

#predictions, weighted means, and non-weighted means
nonw_means <- aggregate(longD$y,by=list(Group=longD$GMax,x=longD$x),FUN="mean")
names(nonw_means)[3] <- "y"

p3 <- p2 + geom_line(data=nonw_means, aes(x=x,y=y), col='black') + facet_wrap(~Group)

You can see the non-weighted means are almost exactly the same as the weighted ones. For group 3 you typically need to go to the hundredths to see a difference.

#check out how close
nonw_means[nonw_means$Group==3,'y'] -  x[x$Group==3,'w_mean']

You can subsequently superimpose the predicted group means over the individual trajectories as well.

#superimpose predicted over ind trajectories
pred$GMax <- pred$Group
p4 <- ggplot() + geom_line(data=pred, aes(x=x,y=pred_mean), col='red') + 
                 geom_line(data=longD, aes(x=x,y=y,group=Ord), alpha = 0.1) + facet_wrap(~GMax)

Two types of absolute fit measures I’ve seen advocated in the past are the average maximum posterior probability per group and the odds of correct classification. The occ function calculates these numbers given two vectors (one of the max probabilities and the other of the group classifications). We can get this info from our long data by just selecting a subset from one time period. Here the output at the console shows that we have quite large average posterior probabilities as well as high odds of correct classification. (Also updated to included the observed classified proportions and the predicted proportions based on the posterior probabilities. Again, these all show very good model fit.)

# group AvePP Freq   OCC      ClassProp PredProp
#1 1 0.9975787 22 6666.84046 0.05820106 0.05835851
#2 2 0.9810445 55 303.94300 0.14550265 0.14681030
#3 3 0.9443613 134 30.90624 0.35449735 0.35262337
#4 4 0.9651479 167 34.98890 0.44179894 0.44220783

A plot to accompany this though is a jittered dot plot showing the maximum posterior probability per group. You can here that groups 3 and 4 are more fuzzy, whereas 1 and 2 mostly have very high probabilities of group assignment.

#plot of maximum posterior probabilities
p5 <- ggplot(data=subD, aes(x=as.factor(GMax),y=PMax)) + geom_point(position = "jitter", alpha = 0.2)

Remember that these latent class models are fuzzy classifiers. That is each point has a probability of belonging to each group. A scatterplot matrix of the individual probabilities will show how well the groups are separated. Perfect separation into groups will result in points hugging along the border of the graph, and points in the middle suggest ambiguity in the class assignment. You can see here that each group closer in number has more probability swapping between them.

#scatterplot matrix
sm <- ggpairs(data=subD, columns=4:7)

And the last time series plot I have used previously is a stacked area chart.

#stacked area chart
nonw_sum <- aggregate(longD$y,by=list(Group=longD$GMax,x=longD$x),FUN="sum")
names(nonw_sum)[3] <- "y"
p6 <- ggplot(data=nonw_sum, aes(x=x,y=y,fill=as.factor(Group))) + geom_area(position='stack')

I will have to put another post in the queue to talk about the spatial point pattern tests I used in that trajectory paper for the future as well.

Online geocoding in R using the NYS GIS server

Previously I wrote a post on using the NYS ESRI geocoding server in python. I recently wrote a function in R to do the same. The base url server has changed since I wrote the Python post, but it is easy to update that (the JSON returned doesn’t change.) This should also be simple to update for other ESRI servers, just change the base variable in the first function. This uses the httr package to get the url and the jsonlite package to parse the response.

#Functions for geocoding using online NYS GIS Esri API,

#getting a single address, WKID 4326 is WGS 1984, so returns lat/lon
get_NYSAdd <- function(address,WKID='4326'){
  base <- ""
  soup <- GET(url=base,query=list(SingleLine=address,maxLocations='1',outSR=WKID,f='pjson'))
  dat <- fromJSON(content(soup,as='text'),simplifyVector=TRUE)$candidates
#looping over a vector of addresses, parsing, and returning a data frame
geo_NYSAdd <- function(addresses,...){
  #make empy matrix
  l <- length(addresses)
  MyDat <- data.frame(matrix(nrow=l,ncol=3))
  names(MyDat) <- c("Address","Lon","Lat")
  for (i in 1:l){
    x <- get_NYSAdd(address=addresses[i],...)
    if (length(x) > 0){
      MyDat[i,1] <- x[,1]
      MyDat[i,2] <- x[,2][1]
      MyDat[i,3] <- x[,2][2]
  MyDat$OrigAdd <- addresses

The first function takes a single address, gets and parses the returning JSON. The second function loops over a list of addresses and returns a data frame with the original addresses, the matched address, and the lat/lon coordinates. I use a loop instead of an apply type function because with the web server you really shouldn’t submit large jobs that it would take along time anyway. The NYS server is free and has no 2,500 limit, but I wouldn’t submit jobs much bigger than that though.

AddList <- c("100 Washington Ave, Albany, NY","100 Washington Ave Ext, Albany, NY",
             "421 New Karner Rd., Albany, NY","Washington Ave. and Lark St., Albany, NY","poop")
GeoAddresses <- geo_NYSAdd(addresses=AddList)

We can compare these to what the google geocoding api returns (using the ggmap package):

googleAddresses <- geocode(AddList,source="google")
GeoAddresses$G_lon <- googleAddresses$lon
GeoAddresses$G_lat <- googleAddresses$lat

And we can see that the nonsense "poop" address was actually geocoded! See some similar related funny results from the google maps geocoding via StackMaps.

We can also see some confusion between Washington Ave. Ext as well. The NYS online server should theoretically have more up to date data than Google, but as above shows it is not always better. To do geocoding well takes some serious time to examine the initial addresses and the resulting coordinates in my experience.

To calculate the great circle distance between the coordinates we can use the spDists function in the sp library.

spDists(x = as.matrix(GeoAddresses[1:4,c("Lon","Lat")]),
        y = as.matrix(GeoAddresses[1:4,c("G_lon","G_lat")]),
        longlat=TRUE,diagonal=TRUE) #distance in kilometers

But really, we should just project the data and calculate the Euclidean distance (see the proj4 library). Note that using the law of cosines is typically not recommended for very small distances, so the last distance is suspect. (For reference I point to some resources and analysis showing how to calculate great circle distances in SPSS on Nabble recently.)

Custom square root scale (with negative values) in ggplot2 (R)

My prior rootogram post Jon Peck made the astute comment that rootograms typically are plotted on a square root scale. (Which should of been obvious to me given the name!) The reason for a square root scale for rootograms is visualization purposes, the square root scale gives more weight to values nearby 0 and shrinks values farther away from 0.

SPSS can not have negative values on a square root scale, but you can make a custom scale using ggplot2 and the scales package in R for this purpose. Here I just mainly replicated this short post by Paul Hiemstra.

So in R, first we load the scales and the ggplot2 package, and then create our custom scale function. Obviously the square root of a negative value is not defined for real numbers, so what we do is make a custom square root function. The function simply takes the square root of the absolute value, and then multiplies by the sign of the original value. This function I name S_sqrt (for signed square root). We also make its inverse function, which is named IS_sqrt. Finally I make a third function, S_sqrt_trans, which is the one used by the scales package.


S_sqrt <- function(x){sign(x)*sqrt(abs(x))}
IS_sqrt <- function(x){x^2*sign(x)}
S_sqrt_trans <- function() trans_new("S_sqrt",S_sqrt,IS_sqrt)

Here is a quick example data set in R to work with.

#rootogram example, see
MyText <- textConnection("
Dist Val1 Val2
1 0.03 0.04
2 0.12 0.15
3 0.45 0.50
4 0.30 0.24 
5 0.09 0.04 
6 0.05 0.02
7 0.01 0.01
MyData <- read.table(MyText,header=TRUE)
MyData$Hang <- MyData$Val1 - MyData$Val2

And now we can make our plots in ggplot2. First the linear scale, and second update our plot to the custom square root scale.

p <- ggplot(data=MyData, aes(x = as.factor(Dist), ymin=Hang, ymax=Val1)) + 
     geom_hline(aes(yintercept=0)) + geom_linerange(size=5) + theme_bw()

p2 <- p + scale_y_continuous(trans="S_sqrt",breaks=seq(-0.1,0.5,0.05), name="Density")

Venn diagrams in R (with some discussion!)

The other day I had a set of three separate categories of binary data that I wanted to visualize with a Venn diagram (or a Euler) diagram of their intersections. I used the venneuler R package and it worked out pretty well.

MyVenn <- venneuler(c(A=74344,B=33197,C=26464,D=148531,"A&B"=11797, 
MyVenn$labels <- c("A\n22","B\n7","C\n5","D\n58")

Some digging around on this topic though I came across some pretty interesting discussion, in particular a graph makeover of a set of autism diagnoses, see:

for background. Below is a recreated image of the original Venn diagram under discussion (from Kosara’s American Scientist article.)

Applying this example to the venneuler library did not work out so well.

MyVenn2 <- venneuler(c(A=111,B=65,C=94,"A&B"=62,"A&C"=77,"B&C"=52,"A&B&C"=51))
MyVenn2$labels <- c("PL-ADOS","clinician","ADI-R")

Basically there is a limit on the size the intersections can be with the circles, and here the intersection of all three sets is very large, so there is no feasible solution for this example.

This is alittle bit different situation than typical for Venn diagrams though. Typically these charts all one is interested in is the overlaps between each set. But the autism graph that is secondary. What they were really interested in was the sensitivity of the different diagnostic measures (i.e. percentage identifying true positives), and to see if any particular combination had the greatest sensitivity. Although Kosara in his blog post says that all of the redesigns are better than the original I don’t entirely agree, I think Kosara’s version of the Venn diagram with the text labels does a pretty good job, although I think Kosara’s table is sufficient as well. (Kosara’s recreated graph has better labelling than the original Venn diagram, mainly by increasing the relative font size.)

For the autism graph there are basically two over-arching goals:

  • identifying the percent within arbitrary multiple intersections
  • keeping in mind the baseline N for each of the arbitrary sets

It is not immediately visually obvious, but IMO it is not that hard to arbitrarily collapse different categories in the original Venn diagram and make some rough judgements about the sensitivity. To me the first thing I look at is the center piece, see it is quite a high percentage, and then look to see if I can make any other arbitrary categories to improve upon the sensitivity of all three tests together. All others are either very small baselines or do not improve the percentage, so I conclude that all three combined likely have the most sensitivity. You may also see that the clinicians are quite high for each intersection, so it is questionable whether the two other diagnostics offer any significant improvement over just the clinicians judgement, but many of the clinician sets have quite small N’s, so I don’t put as much stock in that.

Another way to put it is if we think of the original Venn diagram as a graphical table I think it does a pretty good job. The circles and the intersections are a lie factor in the graph, in that their areas do not represent the baseline rates, but it is an intuitive way to lay out the textual categories, and only takes a little work to digest the material. Kosara’s sorted table does a nice job of this as well, but it is easier to ad-hoc combine categories in the Venn diagram than in table rows that are not adjacent. Visually the information does not pop out at you, like a functional relationship in a scatterplot, but the Venn diagram has the ingredients that allow you to drill down and estimate the information you are looking for. Being able to combine arbitrary categories is the key here, and I don’t think any of the other graphical representations allow one to do that very easily.

I thought a useful redesign would be to keep the Venn theme, but have the repeated structures show the base rate via Isotype like recurring graphs. Some of this is motivated by using such diagrams in interpreting statistics (see this post by David Spieghalter for one example, the work of Gerd Gigenzer is relevant as well). I was not able make a nice set of contained glyphs though. Here is a start of what I am talking about, I just exported the R graph into Inkscape and superimposed a bunch of rectangles.

This does not visualize the percentage, but one way to do that would be to color or otherwise distinguish the blocks in a certain way. Also I gave up before I finished the intersecting middle piece, and I would need to make the boxes a bit smaller to be able to squeeze it in. I think this idea could be made to work, but this particular example making the Venn even approximately proportional is impossible, and so sticking with the non-proportional Venn diagram and just saying it is not proportional is maybe less likely to be misleading.

I think the idea of using Isotype like repeated structures though could be a generally good idea. Even when the circles can be made to have the areas of the intersection exact, it is still hard to visually gauge the size of circles (rectangles are easier). So the multiple repeated pixels may be more useful anyway, and putting them inside of the circles still allows the arbitrary collapsing of different intersections while still being able to approximately gauge base rates.

Transforming KDE estimates from Logistic to Probability Scale in R

The other day I had estimates from several logistic regression models, and I wanted to superimpose the univariate KDE’s of the predictions. The outcome was fairly rare, so the predictions were bunched up at the lower end of the probability scale, and the default kernel density estimates on the probability scale smeared too much of the probability outside of the range.

It is a general problem with KDE estimates, and there are two general ways to solve it:

  • truncate the KDE and then reweight the points near the edge (example)
  • estimate the KDE on some other scale that does not have a restricted domain, and then transform the density back to the domain of interest (example)

The first is basically the same as edge correction in spatial statistics, just in one dimension instead of the two. Here I will show how to do the second in R, mapping items on the logistic scale to the probability scale. The second linked CV post shows how to do this when using the log transformation, and here I will show the same with mapping logistic estimates (e.g. from the output of a logistic regression model). This requires the data to not have any values at 0 or 1 on the probability scale, because these will map to negative and positive infinity on the logistic scale.

In R, first define the logit function as log(p/(1-p) and the logistic function as 1/(1+exp(-x)) for use later:

logistic <- function(x){1/(1+exp(-x))}
logit <- function(x){log(x/(1-x))}

We can generate some fake data that might look like output from a logistic regression model and calculate the density object.

x <- rnorm(100,0,0.5)
l <- density(x)  #calculate density on logit scale

This blog post goes through the necessary math, but in a nut shell you can’t simply just transform the density estimate using the same function, you need to apply an additional transformation (referred to as the Jacobian). So here is an example transforming the density estimate from the logistic scale, l above, to the probability scale.

px <- logistic(l$x)  #transform density to probability scale
py <- l$y/(px*(1-px))

To make sure that the area does sum to one, we can superimpose the density calculated on the data transformed to the probability scale. In this example of fake data the two are pretty much identical. (Black line is my transformed density, and the red is the density estimate based on the probability data.)

dp <- density(logistic(x)) #density on the probability values to begin with

Here is a helper function, denLogistic, to do this in the future, which simply takes the data (on the logistic scale) and returns a density object modified to the probability scale.

logistic <- function(x){1/(1+exp(-x))}
logit <- function(x){log(x/(1-x))}
denLogistic <- function(x){
  d <- density(x)
  d$x <- logistic(d$x)
  d$y <- d$y/(d$x*(1-d$x))
  d$call <- 'Logistic Density Transformed to Probability Scale'
  d$bw <- paste0(signif(d$bw,4)," (on Logistic scale)")

In cases where more of the probability density is smeared beyond 0-1 on the probability scale, the logistic density estimate will look different. Here is an example with a wider variance and more predictions near zero, so the two estimates differ by a larger amount.

lP <- rnorm(100,-0.9,1)
test <- denLogistic(lP)

Again, this works well for data on the probability scale that can not be exactly zero or one. If you have data like that, the edge correction type KDE estimators are better suited.

Translating between the dispersion term in a negative binomial regression and random variables in SPSS

NOTE!! – when I initially posted this I was incorrect, I thought SPSS listed the dispersion term in the form of Var(x) = mean + mean*dispersion. But I was wrong, and it is Var(x) = 1 + mean*dispersion (the same as Stata’s, what Cameron and Trivedi call the NB2 model, as cited in the Long and Freese Stata book for categorical variables.) The simulation in the original post worked out because my example I used the mean as 1, here I update it to have a mean of 2 to show the calculations are correct. (Also note that this parametrization is equivalent to Var(x) = mean*(1 + mean*dispersion), see Stata’s help for nbreg.)

When estimating a negative binomial regression equation in SPSS, it returns the dispersion parameter in the form of:

Var(x) = 1 + mean*dispersion

When generating random variables from the negative binomial distribution, SPSS does not take the parameters like this, but the more usual N trials with P successes. Stealing a bit from the R documentation for dnbinom, I was able to translate between the two with just a tedious set of algebra. So with our original distribution being:

Mean = mu
Variance = 1 + mu*a

R has an alternative representation closer to SPSS’s based on:

Mean = mu
Variance = mu + mu^2/x

Some tedious algebra will reveal that in this notation x = mu^2/(1 - mu + a*mu) (note to future self, using Solve in Wolfram Alpha could have saved some time, paper and ink). Also, R’s help for dbinom states that in the original N and P notation that p = x/(x + mu). So here with mu and a (again a is the dispersion term as reported by GENLIN in SPSS) we can solve for p.

x = mu^2/(1 - mu + a*mu)
p = x/(x + mu)

And since p is solved, R lists the mean of the distribution in the N and P notation as:

n*(1-p)/p = mu

So with p solved we can figure out N as equal to:

mu*p/(1-p) = n

So to reiterate, if you have a mean of 2 and dispersion parameter of 4, the resultant N and P notation would be:

mu = 2
a = 4
x = mu^2/(1 - mu + a*mu) = 2^2/(1 - 2 + 4*2) = 4/7
p = x/(x + mu) = (4/7)/(4/7 + 2) = 2/9
n = mu*p/(1-p) = 2*(4/7)/(3/7) = 8/3

Here we can see that in the N and P notation the similar negative binomial model results in a fractional number of successes, which might be a surprising result for some that it is even a possibility. (There is likely an easier way to do this translation, but forgive me I am not a mathematician!)

Now we would be finished, but unfortunately SPSS’s negative binomial random functions only take integer values and do not take values of N less than 1 (R’s dnbinom does). So we have to do another translation of the N and P notation to the gamma distribution to be able to draw random numbers in SPSS. Another representation of the negative binomial model is a mixture of Poisson distributions, with the distribution of the mixtures being from a gamma distribution. Wikipedia lists a translation from the N and P notation to a gamma with shape = N and scale = P/(1-P).

So I wrapped these computations up in an SPSS macros that takes the mean and the dispersion parameter, calculates N and P under the hood, and then draws a random variable from the associated negative binomial distribution.

DEFINE !NegBinRV (mu = !TOKENS(1)
       /disp = !TOKENS(1) 
       /out = !TOKENS(1) )
COMPUTE #x = !mu**2/(1 - !mu + !disp*!mu).
COMPUTE #p = #x / (#x + !mu).
COMPUTE #n = !mu*#p/(1 - #p).
COMPUTE #G = RV.GAMMA(#n,#p/(1 - #p)).
FORMATS !Out (F5.0).

I am not sure if it is possible to use this gamma representation and native SPSS functions to calculate the corresponding CDF and PDF of the negative binomial distribution. But we can use R to do that. Here is an example of keeping the mean at 1 and varying the dispersion parameter between 0 and 5.

x <- expand.grid(0:10,1:5)
names(x) <- c("Int","Disp")
mu <- 1
x$PDF <- mapply(dnbinom, x=x$Int, size=mu^2/(1 - mu + x$Disp*mu), mu=mu)
#add in poisson 
t <- data.frame(cbind(0:10,rep(0,11),dpois(0:10,lambda=1)))
names(t) <- c("Int","Disp","PDF")
x <- rbind(t,x)
p <- ggplot(data = x, aes(x = Int, y = PDF, group = as.factor(Disp))) + geom_line()
#for the CDF
x$CDF <- ave(x$PDF, x$Disp, FUN = cumsum) 

Here you can see how the larger dispersion term can easily approximate the zero inflation typical in criminal justice data (see an applied example from my work). R will not take a dispersion parameter of zero in this notation (as the size would be divided by zero and not defined), so I just tacked on the Poisson distribution with a mean of zero.

Here is an example of generating random data from a negative binomial distribution with a mean of 2 and a dispersion parameter of 4. I then grab the PDF from R, and superimpose them both on a chart in SPSS (or perhaps I should call it a PMF, since it only has support on integer values). You can see the simulation with 10,000 observations is a near perfect fit (so a good sign I did not make any mistakes!)

*Simulation In SPSS.
LOOP Id = 1 TO 10000.

!NegBinRV mu = 2 disp = 4 out = NB.

*Making seperate R dataset of PDF.
mu <- 2
disp <- 4
x <- 0:11
pdf <- dnbinom(x=x,size=mu^2/(1 - mu + disp*mu),mu=mu)
#add in larger than 10
pdf[max(x)+1] <- 1 - sum(pdf[-(max(x)+1)])
MyDf <- data.frame(cbind(x,pdf))
FORMATS x (F2.0).
VALUE LABELS x 11 '11 or More'.

*Now superimposing bar plot and PDF from separate datasets.
FORMATS NB_Cat (F2.0).
VALUE LABELS NB_Cat 11 '11 or More'.

  SOURCE: Data=userSource(id("Data"))
  DATA: NB_Cat=col(source(Data), name("NB_Cat"), unit.category())
  DATA: COUNT=col(source(Data), name("COUNT"))
  SOURCE: PDF=userSource(id("PDF"))
  DATA: x=col(source(PDF), name("x"), unit.category())
  DATA: den=col(source(PDF), name("pdf"))
  TRANS: den_per = eval(den*100)
  GUIDE: axis(dim(1))
  GUIDE: axis(dim(2))
  SCALE: linear(dim(2), include(0))
  ELEMENT: interval(position(summary.percent(NB_Cat*COUNT)), shape.interior(shape.square))
  ELEMENT: point(position(x*den_per), color.interior(, size(size."8"))