Real Progress - 5 digit ZCTAs

It has been 4 months since real progress was made - but tonight represented a real break through. I got the code up and running on my new laptop and transitioned over to using Eclipse, a great IDE that has integrated debugging of Ruby. But talking about how I am developing is not the purpose of this post. The purpose is to breath life into the project once more.

I've gotten many e-mails over the past few months asking for help in creating maps. One person in particular had a very specific ask that I thought would both expand the functionality of the code & test a new data source. For this set of maps, I needed to start mapping 5-digit zip codes (Zip Code Tabulation Areas to the Census), my first foray into zip codes in this work.

As proof of life, enjoy the screenshot below which shows all of the zip codes for the state of Maine, colored with random colors.



Now that the code is running again and I've got some specific projects to get going again, I look forward to getting back to a much more regular schedule. Please keep up the contact.

Federal Election Commission Campaign Contribution Data

I am stretching my wings beyond Census data - and what could be more interesting then campaign contribution data? I have an interest in politics and thought it would be fascinating to see if there were ways to make the data about the amount spent on elections available. Using the same basic framework I created to read in Census data, I decided that the Federal Election Commission (FEC) would be a good source of interesting data about elections.

The Census provides boundary files for the last eight Congresses. The FEC provides a wealth of data and I have just started to explore the full extent of what is possible. I started by using the Candidate Financial Summary Without PAC Breakdown data. This data, of course, is in a rather complex format, so writing a Ruby module to read it out was the first order of business.

The Census boundary files for the 110th and 109th are in a relatively similar format (the 108th and before begin to deviate substantially) so I used these as my polygon files. I used the Candidate Financial Summary Without PAC Breakdown (CFS) files from the FEC. These files are the most current but do have some potential accounting issues. As the FEC states:

The cost of this timelines, though, is that some of the information available here is less precise than for "cansum". For example, in "cansum" you can see how much a campaign received from Corporate PACs or Labor PACs, while here there is only one value for the total received from "other political committees." This includes all PAC contributions, but it may also contain contributions from other candidates, and some other types of committees we don't typically think of as PACs. We can't do the full breakdowns until all the information about specific contributions has been entered into the database.

When using these summary files you need to be aware of some possible double counting of activity. Some candidates have more then one committee authorized to raise and spend funds on their behalf. The activity reflected in this file represents the sum of those committees. If they transfer funds back and forth among each other, this activity would be counted twice. Information about "transfers from authorized committees" and "transfers to authorized committees" is included in the file and if there are values in both of these fields it is necessary to subtract these from total receipts and total disbursements to obtain a more accurate value for actual activity.

In creating a data structure to store the CFS data, I created a somewhat flexible way to aggregate the data using some of the interesting features of Ruby. In particular the eval statement make it quite easy to pass in free text to allow the caller of the aggregation function to specify which field to aggregate quite easily.

I encountered two main challenges mapping the CFS data back to the Census Congressional District polygons. First, the FEC uses the two-letter acronym to identify the state and the Census uses FIPS codes to identify states. Using regular expressions in TextMate, I converted the list from the Census website to a couple of different Ruby hash tables so that I could convert back and forth from two-letter acronym to two-digit code.

Second, the FEC files are somewhat inconsistent (at least based on how I am reading it) about how they handle states with only one Congressional District. The Census bureau is pretty clear that it uses "00" to identify districts that are the sole district for a given state. The FEC seems to follow this convention for the most part, except in a couple of states. For example, in Wyoming there are candidates for the House of Representatives listed in Congressional Districts "00" and "01". Wyoming has only one seat in the 110th congress. Here is a snippet from the webl06.zip file which contains the CFS data for the 2005-2006 election cycle (the ellipses represent where I have cut from the line for the sake of readability):

  H4WY00055CUBIN, BARBARA L                      I2REP...WY01 W W48...
  H6WY01025TRAUNER, GARY S                        1DEM...WY01 W L47...
  H6WY00118WINNEY, JUSTIN WILLIAM JR              2REP...WY00     0...


According to the documentation for the file, the two digits following the state acronym represent the district. In this case, it would appear to suggest that there are candidates for House in district "00" and "01", when in fact there is only one district in Wyoming. To handle this, I simply rolled up all of the House candidates, regardless of the district in the FEC file for one district states. This may be the wrong thing to do, so I've got an e-mail into the FEC to find out the actual answer.

Alright, enough with that background, let's see some pictures. In the following screenshots, I am mapping the total amount received by House of Representative candidates in each Congressional District for a given election cycle. For each $1,000 received, the district gets one meter in height. The districts with higher amounts are more green, those with lower amounts are more red. Missouri is missing from the 109th Congress because the polygon metadata file is missing from the Census website.

109th Congress from above:


110th Congress from above:


109th looking North:


110th looking North:


109th looking West:


110th looking West:


109th looking East:


110th looking East:


109th looking South:


110th looking South:


That is it for now. I'm close to sharing the KMZs for these, so be on the look out for a post soon.

"Dynamic" Labels & StyleMaps

I've been unhappy with how Google Earth handles labeling of shapes - the Placemark KML object is pretty good at containing a single shape, but breaks down when handling multiple shapes. Placemark elements contain the name and description elements, but only Point elements pick up on these. As I have talked about before, you can't label a Polygon without using a Point. I started to use the MultiGeometry to group together Polygon elements that belong to one geography (i.e. when a County has a couple of noncontiguous shapes). The annoying part is that I needed to add a Point for labels to show up.

When dealing with complex geographies, like Block Groups, having the labels show up the entire time doesn't work very well. For example in the last map I shared, there are ~4,300 Block Groups in WA and having the labels all show up doesn't work very well because they overlap and make it quite confusing. You can play with breaking out the labels into a different Placemark and perhaps a folder structure at the County/County Subdivision level might help, but it still isn't perfect because you would have to hunt and peck for what you were looking. In looking through the KML spec, I was excited to find the StyleMap element. The StyleMap element provides a mechanism to have a Placemark respond to mouseover/click/highlight events. You can define a Style element for both normal and highlight classes. I thought this would be a great way to provide a label: when a user moves their mouse over a geography (e.g. Block Group, County Subdivision, County) the label could show up - the normal style would have it transparent and the normal would have it opaque.

Well, turns out it doesn't quite work that way. Unfortunately, the only thing that sparks the transformation is the user moving their mouse over the icon of a Point: the style doesn't change if they mouseover the label or any of the Polygon elements in the Placemark. What is odd, however, is that all of the elements of the Placemark do respond to the new style when you mouseover the icon.

What you'll see below are maps of Median Household Value (variable H85 MEDIAN VALUE (DOLLARS) FOR ALL OWNER-OCCUPIED HOUSING UNITS [1] from Summary File 3) by County Subdivision. I'm not entirely happy with my new labeling, but it works such that when you mouseover the icon at the center of the polygon, the name of the County Subdivision shows up and the border is highlighted in white. I like the effect, but I am not happy with how I have to have an icon show up. In the maps below, the more red an area/shape, the higher the median household value - the greener, the lower the median household value. In the 3D maps, each $1,000 of value adds one meter of height.

Movie:


Screenshot (notice how Leavenworth-Lake Wenatchee is highlighted):

Median Household Value for Washington state by Block Group

Using the new functionality I've discussed in the last two posts, I am pushing forward in creating new maps. Today I'm going to share a few maps of median household value (variable H76 - MEDIAN VALUE (DOLLARS) FOR SPECIFIED OWNER-OCCUPIED HOUSING UNITS [1] from Summary File 3). There is another variable, H85, which might be better for what I want, but I'm going to go ahead and share these maps before going back. These new maps show the median household value by block group for Washington state. There are ~4,300 block groups in Washington and a wide array of values in the data, so this data provides a good test of the new functionality (labels, excluded polygons, logarithmic color scales). Given the large number of block groups, I've found that it takes several different map formats to fully explore the data. I've created maps that have the 3D views I've shared before and maps that are flat with some transparency so that you can see the underlying geography.

One of the challenges I found in creating these maps was gathering the Census data at the block group for the entire state. NHGIS doesn't provide many variables beyond the basic population and economic ones and the Census FactFinder website doesn't make it easy to download all of the block groups for a given state at once - you have to download each county separately. Given these challenges, I looked into download the raw data from Summary File 3. There raw files are available by FTP but, of course, are in a very complex format. The Census Bureau provides an Access database template that contains empty versions of each of the ~80 tables needed to work with the data (including import specs which is quite helpful). Feeling intrepid I downloaded all of the data for Washington state (FTP site) and loaded the tables I needed to into the Access template. This worked pretty well, but is somewhat confusing, particularly because joining the geographic identifiers for each record is not quite as straight forward as the Census documentation would lead you to believe. I finally got it to work and this provided the data for the maps provided below - perhaps you can now understand why I haven't gone back and re-run the maps using the H85 variable yet.

In the maps below, the more red an area/shape, the higher the median household value - the greener, the lower the median household value. In the 3D maps, each $1,000 of value adds one meter of height. You'll note that some areas show up as white, this is because the data provided by the Census for these block groups is 0. This makes sense for places like Mt. Rainier, but not for a region of downtown Seattle that shows up as white. I'm going to look into this next.

Now, some maps (don't forget you can click on each picture to get a larger version)!

Entire state (flat from overhead):

Entire state (flat from overhead, borders around each block group, some transparency):

Entire state (3D from overhead):

Entire State (3D looking North):

Entire State (3D looking East):

Seattle (3D looking Southeast):

Seattle (flat from overhead, borders around each block group, some transparency):

Hawaii Population Data & New Features

As promised, my break gave me a new burst of energy. I spent my week off in Hawaii and figured it would be a good region to experiment with a few new map features. I tackled two items that were on my to-do list: labels & logarithmic color scales. I have generated a new map of population by County Subdivsion for Hawaii to demonstrate these.

First, on labels. I've found that Google Earth is quite powerful except when it comes to labeling polygons. You can provide "names" for polygons, but these only appear in the "Places" panel on the left of the map view. I suppose this may have to do with the difficulty in figuring out where to put these names on the map display (given how oddly shaped a polygon can be), but I would have thought there was some good default behavior for this (if I am missing a feature of KML, please let me know!). To place a label on the map display, you have to create a "Point" placemark. Out of a desire to keep moving, I pushed forward without labels. The Census Bureau's shape files do actually include a center point for each polygon, so I have now gone back and updated my code to generate "Point" placemarks for each of these center points. With these points, I can now have labels appear on the map. This is very helpful, particularly when dealing with geographies below County (County Subdivision, Block Group, etc.).

Second, on logarithmic color scales. One of the challenges in mapping any sort of data that has a very wide range and is not very evenly distributed across the range is that in can be hard to find a color scheme that provides clarity at either extreme. I have talked about this in a couple of previous posts, but finally gotten around to implementing a logarithmic scheme that more evenly distributes the data across the range. I'm not entirely happy with what I've implemented, so I plan to work on it further.

On to some screenshots. Below you'll find 3 perspectives of population, by County Subdivision, from the 2000 Census for Hawaii. Every meter in height represents 5 people; Greener represents lower population, Red higher population. The labels come in handy because I'm not that familiar with the islands. The new logarithmic color scheme comes in handy because Honolulu has a much higher population that all of the other County Subdivisions. I've used the same Green to Red color scheme I've used before, but with the logarithmic scaling it now does a much better job of helping one to distinguish between the Subdivisions on the lower end of the population range. Without this new scheme, Honolulu would be red and everything else green.

From directly overhead:


From an angle, looking North:


From an angle, looking South (so you can see the northern side of Oahu):


One other note of interest: I was perplexed for a few minutes because the Midway Islands and the other islands west of Kauai all showed up with tall, red polygons - meaning they have a high population (I excluded them from the screenshots for this reason). It turns out that the Honolulu County Subdivision includes all of these islands, hence they get the data for the entire Subdivision. I suppose this demonstrates one of the perils of geographic aggregation when working with an island chain.

More Migration Analysis

I'm finding the migration variables fascinating. These questions are a part of the long form and can be found in Summary File 3 on the Census website. For the maps in the screenshots below, I used the Population 5 years and over: Different house in 1995; In United States in 1995; Different county; Different state; ... variables. The variables allow the identification of where people are moving from, which is quite interesting. The variables are broken up into 4 regions: Northeast, Midwest, South, and West. What I have mapped is the number of people (over the age of 5) who have moved from a state in a particular region to another state (which may also be in that region). For example, in the Northeast map, if a person lived in Maine and moved to Arizona (the map below will show this appears to be quite a popular destination for New Englanders), they would be counted in the county they moved to in Arizona. If a person lived in Maine and moved to Vermont, they would be counted in the county they moved to in Vermont.

This data is broken up by county and the more red and taller a county, the more people that moved there. The heights are quite exaggerated: each person adds 10 meters of height to the county. These maps show how the linear color scale I've been employing to date only really work on datasets that have quite small ranges. I am working on a logarithmic scaling technique that should help on these sorts of datasets, where there may be a smaller number of values that may distort the distribution of values.

Also, I realize it might not have been intuitive: you can click on the pictures for a much larger version of the image. This is true for all of the pictures on the blog.

Northeast: New Englanders appear to be moving to Florida, Arizona, and California in droves. Chicago & Seattle get a fair number as well.


Midwest: Midwesterners are more focused on Arizona than California and quite drawn to Chicago.


South: Southeners appear to be moving all over including California, Arizona, Georgia , Texas, North Carolina, and DC.


West: Westerners shun moving to the Midwest, South, or East, favoring consolidation in Las Vegas (yes, the more northern red spike is Vegas) and Phoenix. You can see some movements to Hawaii & Alaska in the distance.

Getting better color schemes...

I'm making good progress in finding color schemes that actually make sense. As a sneak peak, I've posted a video of a tour over the median household income data from the 2000 Census by County. Green is low median income and red is high (the other way around, which may be more logical there is just way too much red on the screen). The main hurdle I am working on now is how to reduce the file size. The file shown in the movie is ~7 mb.



Here are some of the color schemes that I have put together by hand (enough of the auto-hex code generation that led me to a random walk across hues & saturations). The movie above uses the scheme third from left.

Median Household Income

I have made progress and am now generating maps at the County level with data. I'm not happy with the color schemes yet, so I will be pushing on this before I share the KMZ file. The orientation of the screenshot below is looking Northeast over New England at an altitude of 900 KM. The data is Median Household Income by County from the 2000 Census (Summary File 3). The brighter & lighter a color the higher the income. The heights of each County represent one meter for every dollar of median household income. If anyone knows of any good resources for color pallets, please drop me a line in the comments!

Counties

It is amazing how life can get in the way of making progress! I've been out of town for the past few days and did not have a chance to push the project forward. I finally got some time today and in an effort to generate output I can share broadly, I'm working to generate some interesting KMZs to share. I wanted to put together KMZs with median income by county subdivision, but could not get NHGIS to generate this data. Given the complication with downloading the data from the Download Center on the Census website at the County Subdivision level, I've decided to roll up to the County level. The Census makes it easy to download variables for the entire country at this level and I am working to chart a couple of variables. If there any you an particularly interested in seeing just leave a comment.

I've gotten the code ready for Counties and I wanted to share a screenshot of what the 48 contiguous states look like broken out at the County level. Real data soon!

Finally got 3D colors working!

Thanks to the very first (and still only) comment left so far on CensusKML I have finally gotten good (well, maybe just bright) colors working when extruding the Census data out on Google Earth. I admit that I am terrible at picking color schemes, so please forgive the white-to-red one I'm using here. The trick was relatively simple: the coordinates must be listed in counter-clockwise order in order for Google Earth to properly draw the colors. I don't understand why this is the case, but it is. The Census boundary files list the coordinates for each polygon in clockwise fashion. Ruby made it very simple to reverse the coordinates. The coordinates are initially stored in an array in the order they are read from the Census files - when it comes time to read them out and put them into the KML file, all you have to do is call reverse! on the coordinate array and the array is reversed in place.

The following is a screen shot of the population data from the 2000 Census, broken out at the County Subdivision level. The height of each subdivision is equal to 1 meter for every 10 people. The colors also represent population, but use the state's max county subdivision population as the denominator, meaning they are only relevant within the state, not comparable across states. The largest subdivision will be bright red and the small ones will be white. I'm experimenting with this to try and get data on two levels - national and state. Among all of the data, note Boston on the upper-left, Florida on the upper-far-right in the distance, and Chicago on the bottom-right. Here you go:


You can compare the previous picture to this one to see how big a difference it makes to list the coordinates in a counter-clockwise fashion: