UUorld

Dear Friends,

We’ve spent the last year involved in several interesting projects, gathering information on new uses of data visualization and collaborating with industry leaders to design UUorld 2.0 … and now we are pleased to announce that we’re starting its development!

We’re going to build a new, better version of our software: we’ll upgrade our data portal and significantly increase the power and flexibility of our rendering engine, in part through the much requested automation of outputs. Read: we’re building an application programming interface (API).

For our current users, among the first new features will be long-awaited multi-variable analysis functions, and also more refined variable scaling, data filtering, and support for map legends. A better data portal will help our users tap into myriad open data sources driven by transparency mandates and standardization of linked data resources. We look forward to rolling out these features and hearing from you how else we can improve the software for you. For that matter, if you want to get anything into the development agenda, now would be a great time to drop us a note! 2011 is going to be a really good year for us, and for anyone who wants a better, easier way to understand data.

If you’re interested in learning more about our plans, please drop us a line at .

Here’s looking forward to a great year! — the UUorld Team

For images and videos covering the number of reported cases since 4/22/09, please see the previous post.

Of the 123 new cases, 74 are in Illinois, which is the largest single day increase for any state.

Note: Colors are by logarithmic scale. See further explaination and comparisons below.

Over the weekend 119 new cases of Swine Flu were reported in the US by the CDC, bringing the count from 160 on Friday to 279 on Monday morning. Note: scroll down for videos of the whole trend, from the first reported cases onward, or search “UUorld” on Vimeo.

To start with, here is an image showing the most recent distribution of cases by state:

Next is the same data, with colors assigned by a logarithmic scale, so that the breadth of states reporting any cases is emphasized. This scaling may be appropriate if we are concerned to take special note of where the Flu is just beginning to appear. Notice the difference in the color legend. We have this version of the map in both a 3D version and in a flat, colors-only version…

Here is a video showing the rise of reported cases from 4/22/09 when no cases had been reported in the US, through 11:00am on 5/4/09…

And a video that shows the same data, with colors by logarithmic scale…

Last, at bottom is a video where colors are by rate of change, such that states with the greatest increase in reported cases appear in dark red and those states with no increase appear in pale yellow. States in blue occur when the number of reported cases decreases, which may happen if an initially reported case cannot be confirmed.

Notice the alternation of high growth in the leading states (CA, TX, NY) with higher rates of increase in a broader group of states (IN, DE, MD, SC, LA, OK, CO). The first cases are reported and growth happens in the leading states, then on 4/30, many other states report increasing numbers, then the leading states are growing fastest again, and last we see a wave of increases coming again from a broader range of states, notably including many new cases in Louisiana and Delaware.

Over the past several days, Mac and Linux users were unable to access the Data Portal through UUorld. This was due to a server misconfiguration that was blocking communication between the UUorld application and the Data Portal. We apologize for the error, and greatly appreciate the help from the UUorld community in finding and isolating this issue.

Hi.  We’ll update this post as new numbers become available — this is the second post.

Our data is from the CDC : http://www.cdc.gov/swineflu/

Please let us know if you see new numbers before we do.  Thanks!

Note: Colors by log scale.

And, here is a video, showing the change in reported cases from 4/30….

And below we can take a look at rates of change by color, with Red as highest growth rate, and Blue as the lowest. In many cases, the number of cases is very small, and since this number is used as the denominator in calculating rate of change, some states with relatively high rates of change have very few actual cases. It may be useful to compare the map below to the top map to assess which states have relatively high numbers of cases and relatively high growth rates. Also worth noting is that none of these numbers have been placed in the context of other public health numbers. For example, if these maps were scaled relative to regular flu cases, it would be very difficult to see any swine flu cases or rates of growth. Nonetheless, if the CDC regards these Swine Flu cases as an important trend to watch, maybe these maps are useful without needing broader context. Comments?

Colors above show rate of change from 4/30, with dark red indicating a growth rate of greater than .1 and pale yellow showing no growth.

Hi.  We’ll update this post as new numbers become available.

Our data is from the CDC : http://www.cdc.gov/swineflu/

Please let us know if you see new numbers before we do.  Thanks!

Map backgrounds are provided in both black and white in case you want to post them anywhere else. If you want to adjust colors, scale, view angle or other details, please use our software, which can be downloaded in a free version from our main page.

In an Economic News Release on Friday, the Bureau of Labor Statistics published their latest unemployment figures, covering the first two months of 2009¹. Not that it comes as a surprise to anyone, but it was more bad news: more unemployment in virtually every sector of US industry.

Unemployment Rate by State in 2007 (click to enlarge)	Unemployment Rate by State in 2008 (click to enlarge)

The overall negative trend is not consistent across sectors, however. While a few areas have managed to stop hemorrhaging workers, others are proportionally hit so hard that another 3 percent of the entire workforce lost their jobs between December and February of this year. Specifically, the goods-producing sectors (generally construction and manufacturing) are still suffering massive job losses each month; professional and business services were hardest hit by sheer numbers, with 180,000 jobs lost in February; and by contrast, some of the other services-providing sectors have slowed their losses considerably. A minority of services have even seen slight recovery, with government, education and health-sector jobs actually increasing marginally in January and February.

While professional and business services did suffer major losses this past month, the 180,000 jobs lost represent one percent of the national work force in that sector; meanwhile, the 168,000 jobs lost in manufacturing, which compound repeated one-percent-and-greater losses of the workforce the past several months, represent almost 1.5 percent of the work force of January.

To expose some of these differing trends, I took the figures for the last three recorded months — December of ‘08 and January and February of ‘09 — and rescaled them as percentage changes within the workforce for a given sector. I then plugged the data into IBM’s online visualization toolkit, “Many Eyes²,” to create the graph below. You can click the image to interact with it and look at my data (note: I couldn’t get the labels to appear without being clipped, so hover over a bar in the graph to get a complete view of the category name).

The disproportion between industries is striking, indeed. And since the graph is showing percentage change in employment figures, the compounded effect of a dwindling labor force month by month makes each job lost in goods-producing industries count further. It should be noted that in addition, not only are construction and manufacturing establishments shedding employees, but they are also reducing the number of hours on average that their employees are at work, which in turn reduces their wages. According to the BLS’s numbers, the average hours of work reported in manufacturing jobs for December, January and February gradually declined from 39.9 to 39.8 to 39.6 hours¹, despite steady averages across all sectors.

–

Taking a closer look at manufacturing…

As UUorld sends its employees between Silicon Valley and Washington DC — areas primarily involved in technology and government, respectively — the immediate effects of failing manufacturing are not always obvious in my everyday experience. Clearly, however, there are parts of the country that rely heavily on manufacturing. I thought it would be useful to include some analysis of regional patterns as the unemployment crisis has affected them.

I started by mapping recorded instances of new unemployment, in the form of mass layoff events, measured whenever a company has at least 50 initial claims for unemployment insurance (UI) filed against it during a 5-week period³. The most recent data being the records for January of 2009, I decided to contrast that data with the same measures for January of 2008. The regional trends are immediately identifiable:

Mass Layoff Events by State, January 20084 (click to enlarge)	Mass Layoff Events by State, January 2009 (click to enlarge)

I’ve scaled the data by dividing each state’s number of recorded mass layoffs by every 100,000 members of its labor force. So, for instance in the upper-right image above, Michigan had 4.6 mass layoffs reported for every 100,000 workers in the state this past January. With a total labor force of about 4.9 million, that means that more than 220 mass layoffs occurred in January alone — more than seven per day.

To get an idea of the consistency of these layoffs, I’ve also generated maps for the last three recorded months. The data in the rightmost map below is identical to the January 2009 map above-right, but the colors are scaled relative to the other maps in the set of three. Because Kentucky suffered an enormous number of mass layoffs in December of 2008 (middle map) which surpasses all other measures, there is some diminution of the high values in January 2009. Given that these recordings are separate from month to month, it is all the more shocking to realize that the effects are cumulative. In other words, while Kentucky suffered 6.4 mass layoffs per 100,000 workers in December, the trend didn’t reverse itself in January, which saw the state suffer another 3.9 mass layoffs per 100,000.

Mass Layoff Events, November ‘08 (click to enlarge)	Mass Layoff Events, December ‘08 (click to enlarge)	Mass Layoff Events, January ‘09 (click to enlarge)

I also made some maps based on the actual total number of claimants³ for UI in a given period of time. The regional patterns are similar, and the rate of new unemployment seems consistent across both time periods as well (the January-to-January comparison and the month-by-month comparison). In this case I scaled the data by dividing by every 1000 members of the labor force; so for example Kentucky saw 10 of every 1000 workers file for unemployment insurance in December of 2008 (the middle small map below).

Initial Claimants for Unemployment Insurance, January 2008 (click to enlarge)	Initial Claimants for Unemployment Insurance, January 2009 (click to enlarge)

Claimants for UI, November 08 (click to enlarge)	Claimants for UI, December 08 (click to enlarge)	Claimants for UI, January 09 (click to enlarge)

Across all of the maps, it is obvious that some states are much harder hit than others. In particular, the Great Lakes states and many of those in the Southeast are the worst off. And that is logical, given the recent data published by BLS: many, if not most, of the Great Lakes and Southeastern states rely heavily on manufacturing.

To go one step further, I took the most recent data from the Census Bureau’s American Community Survey5 to get some maps showing exactly which states were most reliant. Of course all of the hard-hit states show up in dark blue: Michigan, Ohio, Indiana and Wisconsin all see close to a fifth of their entire workforce employed in the sector. I took several snapshots from uncommon angles as well, shown in the small maps below. Looking south across the US, one can see how little

% of Work Force Employed in Manufacturing, by State (click to enlarge)

of the Rocky Mountain states’ labor force is employed in manufacturing — which corresponds with their relative resistance to job losses in the past few months.

Claimants for UI, November 08 (click to enlarge)	Claimants for UI, December 08 (click to enlarge)	Claimants for UI, January 09 (click to enlarge)

What is perhaps surprising in all of this is that, unlike with the housing market or wall street traders, the idea of over-inflated prices self-correcting doesn’t really apply. The school of thought that we should welcome recession as a means of readjusting values to their appropriate levels is somewhat lost on an industry which has actually been in decline for several years running, and which was already suffering — albeit less severely — before the current recession hit.

But even if the recession didn’t cause the problems in manufacturing, it seems that tough economic times have exacerbated the problem. For states in the Rust Belt and states in general that rely heavily on goods-production, GDP has decreased for several years running, and the situation only seems to get worse with time. (Example: ).

Between 2003 and 2007, the three worst-performing states in GDP growth were Michigan, Indiana and Ohio, all vastly under the national levels (and even averaging recessive growth in the case of Michigan). Nonetheless, through the entire period about a quarter of the GDP generated by those states, and some of those in the Southeast (Kentucky, Louisiana, the Carolinas), has remained tied to the sector.

% of GDP Derived from Manufacturing, by State, 1997-2007 (click to play video)

Ultimately, is the solution an “industrial revolution,” as called for by some economists today? Or will the recession force the Great Lakes states and others to redistribute their workforce across other sectors? In any event, all appearances seem to indicate that further tough times lie ahead, and preliminary statistics show a further employment slide as we progress through this year. The BLS’s predictions for the years to come indicate similar trends: in their 10-year projections for industry sectors, a large majority of those listed as “most rapidly declining” are directly related to or are subsets of manufacturing⁶.

It will be interesting to see whether programs like the Obama Stimulus Package can at least slow down the job loss rates in the coming months. We will be watching the release of new statistics closely …

—

Footnotes:

1. Bureau of Labor Statistics. Economic News Release: employment situation summary for February, 2009.
2. research group. Many Eyes online data visualization tool.
3. Bureau of Labor Statistics. Mass layoff statistics program.
4. States that are missing from the maps did not provide data that meets the disclosure standards of either the BLS or local state agency. BLS confidentiality standards.
5. US Census Bureau. American Community Survey. Data available at: UUorld Data Portal.
6. Bureau of Labor Statistics. Employment projections.

You can now follow on Twitter here:

We will be posting development progress and participating in a general conversation about the direction of the product. Jump in!

The “American Recovery and Reinvestment Plan¹,” as christened by President Obama, is meant to channel funds into areas critical to both the short- and long-term health of the US economy. Amidst online discussions of the plan, I came across a series of interesting articles on National Public Radio’s blog, in a section dedicated to economics called “Planet Money.”

	These articles, authored by Alan Cordova2, analyze five elements of the stimulus package, based on estimates released by the White House3 last week. Mr. Cordova did a bit of data crunching with other US-government datasets, which I’ve emulated (household counts and employment figures from the Census Bureau)4, and produced some flat maps to
Alan Cordova’s map on Job Creation - click to enlarge

show which states would be most affected by each part of the stimulus package. He used IBM’s online visualization generator⁵, “Many Eyes,” to create the maps; incidentally, Many Eyes is part of what IBM calls its “Collaborative User Experience” project, and is a nifty interface to create some simple, incisive visualizations. But I digress.

Mr. Cordova’s post piqued my curiosity, and gave me a hankering to actually interact with the numbers myself; so, I reproduced his data and imported it into UUorld. To get spending numbers for individual states, I drew from the memo released by White House economic advisor Brian Deese last week. You can find that memo here.

The memo breaks benefits into five areas, in accordance with new programs; I thought I’d use the same. At first I played around with the sensitivity and ground-level values in UUorld for a bit so that each variable would display with a complete range of colors (I found it hard to differentiate between states in some of the brown 2D maps). In so doing, I came across an array of patterns I wasn’t aware of at first take:

1. The number of jobs projected to be created or saved over the next two years, as a percentage of total employment in a state. (click images to enlarge)

2. The percentage of working-age adults eligible to receive a “Making Work Pay” tax cut of up to $1000. (click images to enlarge)

These first two variables measure benefits that will, in theory, directly combat poverty. The first is of course job creation; the second translates to fiscal support for working adults who are below the poverty line — it’s a basic extension of the Earned Income Tax Credit⁶ (established under Ford, one of the rare domestic fiscal-policy creations that has garnered support from both political parties).

Interestingly, the two benefits seem to complement each other almost perfectly; in states where fewer jobs will be created by percentage, almost without exception, more families are supported by the tax credit. Also interestingly, a few southern states in particular — Lousiana, Alabama and South Carolina (and to an extent Kentucky) — will reap significant benefits from both parts of the program. These states are among the poorest in the Union (all in the bottom 10 by income)⁷, so at first glance, the added support seems well placed.

3. The percentage of families that will become eligible for government aid called the American Opportunity Tax Credit, which is geared specifically towards making college affordable for poorer families. (click images to enlarge)

The AOTC is meant to eliminate around two-thirds of college costs for families in need by furnishing a $4000 tax credit in exchange for 100 hours of community service performed by the candidate⁸.

It immediately jumped out at me that Washington DC, at 9.2% of families eligible, has about twice the number of possible candidates by percentage than most of the rest of the states in the Union. Conventional “wisdom” would hazard me to guess that this is because even though the District is not the poorest state — actually by per capita income it ranks as high as 16th⁷ — its population is severely divided economically. Note to self: look into statistics on that for a future post.

4. The percentage of working-age adults eligible to receive an additional $100 per month in unemployment insurance benefits. (click images to enlarge)

5. The percentage of schools that will undergo so-called “modernization.” (click images to enlarge)

Between these two measures, again it looks like benefits are well distributed around the country. While the additional unemployment credit will have greater effect in the northern states, it seems that southern states will benefit more from the school “modernization” project⁹ encapsulated within the stimulus plan.

–

When I started working with this data, my intention was to determine whether certain states were favored by the stimulus package (by measure of these variables, anyway). As it turned out, interestingly, the benefits seem to be more or less evenly distributed across the country … and while that may be a positive sign for the plan itself, it doesn’t do much to answer my question.

To dig a bit deeper, I took standard deviations and calculated z-scores¹⁰ for all five variables; then I created a new index variable by taking the average z-score for each state. Here are the results:

The states in darker blue (and elevated in the second image) are the ones most positively affected, in the aggregate, by the five benefits of the stimulus package as originally delineated by the Deese memo.

It bears noting that Alaska ranks dead last, receiving the fewest benefits on average as measured by our given variables. However, it would be hard to argue that Obama is using the stimulus package to play favorites, at least on a state-wide level, given that his home state of Hawaii is quite far down the list as well (about two-thirds of the states fare better than Hawaii using this index).

Back during the immediate aftermath of the election, I looked at a range of statistics comparing voting trends — and my colleagues and I wrote some blog posts about those trends. Most notable, to my mind, is that almost without exception the country voted more democratically — that is, more in favor of Obama — than it had four years prior; this held true for even most states that are typically staunchly Republican. While the state may still have gone to John McCain in 2008, the overall

percentage of Democratic votes often dramatically increased from the Bush-Kerry election of 2004. The one exception to this was a strong tendency towards more Republican voting within a band spanning one of the poorest parts of the country: namely, the mouth of the Mississippi River in Mississippi and Louisiana, stretching up through Arkansas, and across Tennessee into Kentucky and West Virginia.

As I was familiar with this trend, I was surprised to note that the stimulus-plan index I created seems to show those same states are for the most part going to benefit proportionally more than many of their peers. Somewhat ironic, perhaps … or could the stimulus plan intentionally favor those constituents with whom Obama is currently least popular? To assert the numerical validity of that assumption, I pitted state-level election results (percentage of the vote won by the Democrats) against the 5-variable index by testing for statistical correlation¹¹. The result was a correlation coefficient of 0.166, meaning that if anything there is a slight tendency for those states that voted for Obama to benefit more from the plan.

While 0.166 is not a strong correlation indicator either way, it does represent the strength of all five variables, averaged. As I had already found out, the benefits of each part of the plan are spread across the country in different proportions. Thus, I also tested for correlation between voting results and the 5 variables individually:

ALL. Five-Variable Index and % of 2008 Obama Votes: Correlation Coefficient = 0.1658 [small positive correlation]

1. Job Creation and % of 2008 Obama Votes: Correlation Coefficient = -0.6336 [large negative correlation]

2. Adults with MWP and % of 2008 Obama Votes: Correlation Coefficient = 0.1162 [small positive correlation]

3. Families with AOTC and % of 2008 Obama Votes: Correlation Coefficient = 0.1240 [small positive correlation]

4. $100 Benefits and % of 2008 Obama Votes: Correlation Coefficient = 0.3741 [positive correlation]

5. Schools Modernization and % of 2008 Obama Votes: Correlation Coefficient = 0.3721 [positive correlation]

This seems to indicate that the job creation forecasted by the White House tends strongly to states that voted against Obama. What the exact measures are to be taken to create those jobs is somewhat unclear, but that Republican-voting states are the predicted beneficiaries is striking.

Of course, correlation is not at all the same thing as causation. I remain wary that the demographic and economic differences that render certain parts of the country susceptible to one element of the stimulus package may be unrelated to whatever factors underpin ideological tendencies in a given direction.

As always, UUorld has helped me explore my questions, answer them, and then find many more.

—

Footnotes:

1. Wikipedia America Recovery and Reinvestment Plan.
2. Alan Cordova is a writer for “Planet Money” at National Public Radio. .
3. Boston Herald. “White House estimates new jobs in stimulus plan” by the Associated Press.
4. American Community Survey. The US Census Bureau’s 2006 tally of households by state; and American Community Survey. The US Census Bureau’s 2006 count of total employed adults by state.
5. research group. Many Eyes online data visualization tool.
6. Internal Revenue Service. Earned Income Tax Credit definition, questions and answers.
7. Wikipedia States of the US by Income.
8. . American Opportunity Tax Credit - Definition and Overview.
9. National Clearinghouse for Educational Facilities. Federal funding stimulus for school facilities: description and comparison of bills.
10. Wikipedia Standard score. Standard deviations were calculated like this: The differences between each value and the mean value are squared, summed, and then divided by the total number of measurements.
11. .

I recently read a post on Cognitive Daily entitled “Reading graphs — How we do it, and what it tells us about making better ones.” The post exposed the research performed by cognitive scientist Raj Ratwani¹, with the goal of tracking eye movements of respondents analyzing a basic population density map.

The inferences were fascinating, and reinforce many of the concepts that have inspired our work here at UUorld. In a nutshell, when posed questions that required analysis (rather than simple observation), respondents were stymied by disparate elements of the presentation: a four-level color scheme and cryptic legend essentially forced them to absorb the data in several steps, processing each step one after the other. The study found that respondents first had to “integrate the graph visually — that is, determine which cluster goes with which data. Then, [they had to] cognitively integrate — figure out the relationship between the clusters.”

Maryland counties population density in 2000 (US Census Bureau) - click to enlarge

Ratwani and his crew used maps very similar to the one I’ve put together here (above). While they used a fictional state, they distributed the counties across it such that obvious “bands” of higher and lower population density were immediately apparent. In my map here, I’ve chosen Maryland because its counties are distributed in a similar fashion. Immediately, we can see that a 4-color spectrum is useless when it comes to differentiating most of the counties in the state.

Of course, making information clearer through maps is basically UUorld’s mission statement, and so I jumped on the opportunity to use our visualization engine to improve upon the Maryland map, and counter the array of “tough spots” exposed by Ratwani’s research.

For starters, much of the difficulty respondents had with the simple density maps came from shifting focus between legend and map, because the map itself failed to convey information on its own about a given county’s measurement. It might be tempting to conclude from this that simply a better legend was needed; but on the contrary, to my mind, the major shortcoming is that the globs of color convey almost no information beyond identifying the most basic regional trends.

Maryland counties population density (US Census Bureau) - click to enlarge	Maryland counties population density (US Census Bureau) - click to enlarge

My first “fix” with UUorld, therefore, was to choose a better color scheme — one which uses a gradual scale, where colors are adjusted by value (above-left). Then, I added labels directly to the counties so that questions akin to those posed in Ratwani’s tests — e.g. “What is the population of X county” — could be answered instantly (above-right).

What continues to amaze me is just how remarkable such small changes can be. Immediately, using the newer Maryland maps, one can detect all kinds of subtleties amongst those counties which were nothing but red in the first image. And, for instance, whereas in the original map Howard county is grouped with Anne Arundel and Baltimore counties, all of them appearing solid yellow, our newer maps show clearly that Howard is significantly less dense than its eastern neighbors (about 200 people per square mile less so).²

Maryland counties population density (US Census Bureau) - click to enlarge	Maryland counties population density (US Census Bureau) - click to enlarge

… And of course all exposed patterns are even clearer in 3D.

What interests me so much about the study Mr. Ratwani put together is that his point of departure was to measure how much thinking was required to answer a question using different maps. That strikes a chord with me because our company was created in recognition of the shortcomings of many charts and graphs, and the incredible informative potential of thematic mapping. It is exciting to envision a scientific analysis of how exactly 3-dimensional mapping improves cognitive efficiency.

For my part, I find that handling statistics with UUorld pushes me towards new discovery. In the case of Maryland population density, I was intrigued that the county-level maps seemed to indicate Washington DC suburbs were even more densely populated than the county that actually contains Baltimore. To put my musings to rest, I loaded up the zip-code level data from our Data Portal and very quickly had the answer: Baltimore is the most densely populated part of Maryland; the county-level maps just aren’t fine-grained enough to pick up the nuance.

Now I’m curious to go back and check on other major metropolitan areas around the country…

Maryland zip codes population density (US Census Bureau) - click to enlarge	Maryland zip codes population density (US Census Bureau) - click to enlarge
Maryland zip codes population density (US Census Bureau) - click to enlarge	Maryland zip codes population density (US Census Bureau) - click to enlarge

—

Footnotes:

1. Mr. Ratwani’s research, published in the Journal of Experimental Psychology, can be found here.
2. US Census Bureau 2000 population density figures: Howard County 983, Baltimore County 1259, Anne Arundel County 1177. (persons per square mile)