On Jul 13, 2018, at 7:30 AM, Roger Bivand <Roger.Bivand at nhh.no> wrote:
On Fri, 13 Jul 2018, Benjamin Lieberman wrote:
All-
I would like to note that as the data is proprietary, and for obvious privacy concerns, the lat/long pairs were randomly generated, and were not taken directly from the data.
Thanks for the clarification. Note that if the data are a sample, that is not a complete listing for one or more study areas, you don't know who the first order neighbour (the most proximate other voter) is, because that indidivual may not be in the sample. Your fallback then is to treat the data as aggregates, unless you rule out local sampling variability.
Roger
--
Benjamin Lieberman
Muhlenberg College 2019
Mobile: 301.299.8928
On Jul 13, 2018, at 6:58 AM, Benjamin Lieberman <bl250604 at muhlenberg.edu> wrote:
Roger anf Facu,
Thank you very much for the help. In terms of the data, I only provided the ID and Lat/Long pairs because they were the only covariates which were necessary. The data set we are using was purchased and contains voter registration information, voter history, and census tract information, after some geocoding took place. The locations are the residents houses, in this instance.
I have rerun the knn with longlat = T, but I am still hung up on the idea of the first order neighbors. I have reread the vignette and section 5 discusses High-Order Neighbors, but there isn?t any mention of first or second order neighbors, as you mentioned above (?first order neighbors are not defined?). One of the pieces of literature I found said that polygons are problematic to work with, as are tesslations for precisely the reason you mentioned, non-planarity. For this reason, I am hung up on the idea of how to find all first order neighbors for a point, especially as the number of first order neighbors varies from point to point, and such knearneigh would not be appropriate here.
If this is something that does not seem feasible, maybe another tactic is necessary.
Again, thank you all for the help.
Warmest
--
Benjamin Lieberman
Muhlenberg College 2019
Mobile: 301.299.8928
On Jul 13, 2018, at 6:11 AM, Roger Bivand <Roger.Bivand at nhh.no <mailto:Roger.Bivand at nhh.no> <mailto:Roger.Bivand at nhh.no <mailto:Roger.Bivand at nhh.no>>> wrote:
On Fri, 13 Jul 2018, Facundo Mu?oz wrote:
Dear Benjamin,
I'm not sure how you define "first order neighbors" for a point. The
first thing that comes to my mind is to use their corresponding voronoi
polygons and define neighborhood from there. Following your code:
Thanks, the main source of confusion is that "first order neighbors" are not defined. A k=1 neighbour could be (as below), as could k=6, or voronoi neighbours, or sphere of influence etc. So reading vignette("nb") would be a starting point.
Also note that voronoi and other graph-based neighbours should only use planar coordinates - including dismo::voronoi, which uses deldir::deldir() - just like spdep::tri2nb(). Triangulation can lead to spurious neighbours on the convex hull.
v <- dismo::voronoi(coords)
par(mfrow = c(1, 2), xaxt = "n", yaxt = "n", mgp = c(0, 0, 0))
plot(coords, type = "n", xlab = NA, ylab = NA)
plot(v, add = TRUE)
text(x = coords[, 1], y = coords[, 2], labels = voter.subset$Voter.ID)
plot(coords, type = "n", xlab = NA, ylab = NA)
plot(poly2nb(v), coords, add = TRUE, col = "gray")
?acu.-
On 07/12/2018 09:00 PM, Benjamin Lieberman wrote:
Hi all,
Currently, I am working with U.S. voter data. Below, I included a brief example of the structure of the data with some reproducible code. My data set consists of roughly 233,000 (233k) entries, each specifying a voter and their particular latitude/longitude pair.
Using individual voter data is highly dangerous, and must in every case be subject to the strictest privacy rules. Voter data does not in essence have position - the only valid voting data that has position is of the voting station/precinct, and those data are aggregated to preserve anonymity.
Why does position and voter data not have position? Which location should you use - residence, workplace, what? What are these locations proxying? Nothing valid can be drawn from "just voter data" - you can get conclusions from carefully constructed stratified exit polls, but there the key gender/age/ethnicity/social class/etc. confounders are handled by design. Why should voting decisions be influenced by proximity (they are not)? The missing element here is looking carefully at relevant covariates at more aggregated levels (in the US typically zoning controlling social class positional segregation, etc.).
I have been using the spdep package with the hope of creating a CAR model. To begin the analysis, we need to find all first order neighbors of every point in the data.
While spdep has fantastic commands for finding k nearest neighbors (knearneigh), and a useful command for finding lag of order 3 or more (nblag), I have yet to find a method which is suitable for our purposes (lag = 1, or lag =2). Additionally, I looked into altering the nblag command to accommodate maxlag = 1 or maxlag = 2, but the command relies on an nb format, which is problematic as we are looking for the underlying neighborhood structure.
There has been numerous work done with polygons, or data which already is in ?nb? format, but after reading the literature, it seems that polygons are not appropriate, nor are distance based neighbor techniques, due to density fluctuations over the area of interest.
Below is some reproducible code I wrote. I would like to note that I am currently working in R 1.1.453 on a MacBook.
You mean RStudio, there is no such version of R.
# Create a data frame of 10 voters, picked at random
voter.1 = c(1, -75.52187, 40.62320)
voter.2 = c(2,-75.56373, 40.55216)
voter.3 = c(3,-75.39587, 40.55416)
voter.4 = c(4,-75.42248, 40.64326)
voter.5 = c(5,-75.56654, 40.54948)
voter.6 = c(6,-75.56257, 40.67375)
voter.7 = c(7, -75.51888, 40.59715)
voter.8 = c(8, -75.59879, 40.60014)
voter.9 = c(9, -75.59879, 40.60014)
voter.10 = c(10, -75.50877, 40.53129)
These are in geographical coordinates.
# Bind the vectors together
voter.subset = rbind(voter.1, voter.2, voter.3, voter.4, voter.5, voter.6, voter.7, voter.8, voter.9, voter.10)
# Rename the columns
colnames(voter.subset) = c("Voter.ID", "Longitude", "Latitude")
# Change the class from a matrix to a data frame
voter.subset = as.data.frame(voter.subset)
# Load in the required packages
library(spdep)
library(sp)
# Set the coordinates
coordinates(voter.subset) = c("Longitude", "Latitude")
coords = coordinates(voter.subset)
# Jitter to ensure no duplicate points
coords = jitter(coords, factor = 1)
jitter does not respect geographical coordinated (decimal degree metric).
# Find the first nearest neighbor of each point
one.nn = knearneigh(coords, k=1)
See the help page (hint: longlat=TRUE to use Great Circle distances, much slower than planar).
# Convert the first nearest neighbor to format "nb"
one.nn_nb = knn2nb(one.nn, sym = F)
Thank you in advance for any help you may offer, and for taking the time to read this. I have consulted Applied Spatial Data Analysis with R (Bivand, Pebesma, Gomez-Rubio), as well as other Sig-Geo threads, the spdep documentation, and the nb vignette (Bivand, April 3, 2018) from earlier this year.
Warmest,
Ben
--
Benjamin Lieberman
Muhlenberg College 2019
Mobile: 301.299.8928
[[alternative HTML version deleted]]