About the Data

The United States Centers for Disease Control maintains a database of behavioral health risk factors survey data. This data is freely available for download. Numerous visualizations are also available.

Most geographic visualizations either capture current values on a chloropleth or heat map, or they capture trends using a line chart. Sometimes, though, it is useful to see both the spatial and temporal trends in one graphic, which is why this visualization tool was made.

For this tool, ten specific indicators were chosen based on the recorded summary of CDC BRFSS survey responses. Eight of these indicators come directly from specific categories directly computed by the CDC and made available as part of the data.

In addition, two new categories were used for this visualization. For body mass index, the CDC uses a 5-category response that records overweight and obesity prevalence as separate categories. For this dataset, both categories were combined into a single overweight / obese category.

The raw data tracks the portion of the population who have ever had their cholesterol checked, and the portion of the population who have ever had their cholesterol checked *and* been told that it was high. For this visualization, the prevalence of high cholesterol was estimated simply by taking the fraction who had been told their cholesterol was high and dividing by the fraction of people ever tested. Using this approach as a quantitaive estimate implies that these two factors are independent, which may not be true. As a qualitative indicator, though, it seems reasonable.

The specific questions and answers used were:

About the Trends

Trends in health indicators and behavioral risk factors are clearly important to try and understand, but how can we visualize them easily? For this tool, I chose a categorization approach. If you look at the data, you wll see that over 7 years, the trends can be more complex than simply an increase or a decrease. In Charlotte, NC for instance, diabetes and overweight prevalences appear to have crept up steadily, while the prevalence of smokers dropped in the last 3-4 years after holding steady for the first 3-4 years. You will also notice that the data has some noise, and most year-to-year changes are well within the margin of error, so, with only 7 years of data, complex time-series analysis is hard to justify.

Ideally, what would be nice is to have something that goes a step beyond just a positive or negative trend, something that could help identify the locales where more interesting changes may be taking place. To accomplish that, the positive and negative trends have been divided into three sub-classes, "speeding up", "steady", and "slowing/reversing". This was done by comparing linear and quadratic time-series model fits. For the linear cases, an analysis of variance was performed based on the reported means and estimated variances of the data series. Where ANOVA indicated a change with a Type I error probability of 5% or less, the trend was classified as positive (progress) or negative (decline) based on the difference in "best fit" values (using least squares regressio with inverse variance weighting), and based on common assumptions about what constituted a "healthy" change. Specifically, increases in diabetes, hypertension, overweight/obesity, and high blood cholesterol prevalence were "unhealthy", as were increases in alcohol consumption and smoking prevalence. When these trends are described with the term "decline", it means that they indicate a decline in the health of the population, not that the prevalence itself is declining. I have always tried to make this distinction clear in the visualizations. Increases in physical activity, blood cholesterol screening, health insurance coverage, and self-reported general health were considered "healthy".

Once a trend was established as significant, the adjusted squared Pearson correlation coefficients for the linear and quadratic time-dependent models were compared. If the quadratic fit gave a higher adjusted R-squared, then the absolute value of the estimated slope (vs. time) at the first and last year were compared. A higher absolute value of the slope was deemed "speeding up" (or "accelerating"), while a lower value was classified as "slowing/reversing". Note that speeding up vs. slowing down can be determined independently from whether or not a reversal is indicated in the trend, but to keep things simple the slowing / reversing effects were lumped together as described. If the linear (in time) fit provided a higher adjusted R-squared, then the trend was classified as improving or declining "steadily".

If the results on the "summary" page are examined, you will see that most trends tended to fall into the "steady" category. This is useful as it fulfills the purpose of pointing out what may be the more interesting trends. For instance, for smoking prevalence, out of almost 200 metro areas with a discernible trend, only 17 were classified as "accelerating improvement / progress speeding up" while only two were classified as having a "decline speeding up" / "accelerating decline". A look at the map page will show that many of the 17 areas where progress has accelerated lately are in Kentucky, Tennessee, and the Carolinas -- which gives an interesting insight into what may be a recently emerging health trend in the heart of tobacco country. (To see this for yourself, click on almost any metro area, then find the graph for smoking -- click the "behaviors" radio button if you don't see it -- then click on "compare trends", or, if you are quite unlucky and chose one of the two out of about 250 metro areas where no smoking data is available, just click on a different metro area and the graph will populate with data.)

One of the two metro areas where the smoking trend is an "accelerating decline" is my former home of Akron, OH. I will use it to briefly mention some of the limitations of this visualization approach. To find Akron, if you are not an expert in geography, and you don't feel like reading all the tooltips, you could take a side trip to a map server, or with the smoking trends still displayed, you could look click on one of the two darkest brown dots on the screen. Hint: Akron is the harder dot to find, being obstructed on most screens by other nearby metros. Alas, making the metro "dots" large enough to click on easily makes a few of them more difficult to find if you are looking for one in particular. Thus, a search box is a good choice for a future feature. If you do manage to look at the data for Akron, you will notice there are four data points for most measures. Four is an interesting number as it does allow for an adjusted R-squared comparison between linear and quadratic models, but with only four points an L-, U-, or V-shaped pattern can happen by chance quite frequently, which means that a larger than normal number of fits are misclassified as non-linear. You'll notice quite a few trends in Akron classified as non-linear. You may also notice that the trends in smoking and alcohol consumption are quite similar, yet the smoking trend is classified as a "decline (in health) speeding up" because the 2015 prevalence is slightly higher than the 2011 prevalence, while for alcohol use, the trend is merely "progress slowing/reversing" because the 2015 prevalence is slightly lower than the 2011 prevalence. Thus, the non-linear trends, although doing a good job of making the hunt for interesting data easier, need to be interpreted with more caution!

About the Statistics

Please note that the statistical methods used here are reasonable choices for "exploratory grade" analysis. They are not "publication grade" and not intended for publication and peer review (sometimes data can simply be informative and -- academics, avert your eyes! -- fun).

The reason for the "exploratory" grade involve two simplifications made with the data. First, the complex sample design used to get population estimates by CDC was not taken into account. When, for instance, it is reported that about 15% of respondents in Austin were current smokers, this does not mean that 15% of survey respondents answered yes. About 15% answered yes, but then the CDC adjusted the proportion to take into account the characteristics of the sample, in order to estimate what the responses would have been for a perfectly representative sample. If interested, pleae feel free to read all about it on the CDC website.

What the above means for statistics is that the sampling errors are not what one would normally calculate for a given number of responses. In many cases, the CDC has helpfully provided 95% confidence intervals, but from that we can only approximate the expected variance in the measurements. Which leads us to simplification #2: for this analysis, the response data is assumed to be normally distributed. These data were collected from surveys, often with several hundred to several thousand responses, so the calculated variance in responses based on the binomial distribution should be well-approximated by a normal distribution. The complex sampling design, however, may introduce unexpected features in the distribution. If you want to do publication-quality statistics, you should read the CDC's published papers (or ask them about it).

Where this matters most is in the custom categories (overweight and high cholesterol prevalence). For these categories, the error bars you see are 95% confidence intervals based on assumptions of normality as well as independence of the component measurements and a disregard of the complex sampling design. To see how much this matters, I compared the CDC's own multi-category derived error estimates with ones I computed using the above simplifications. The results agreed within about 10%, but did not coincide perfectly. Thus, the error bars on those two categories should be treated as approximate. Otherwise, the error bars correspond to the 95% confidence limits reported by the CDC, and as long as you simply take them at face value and don't try to further manipulate them for statistical purposes, they should be reliable.

Thanks!

If you're still reading after all that, then you probably like data as much as I do. So you may be interested in knowing (fi you don't already) that the CDC is a great place to get your hands on clean, free data sets that can be quite massive. In addition to thanking them, I would also like to thank Michelle Chandra, whose bl.ocks.org US map came in handy, as well as Trang Tran, whose d3-colorbar also also added a nice touch. The GeoJSON source data (from the US census bureau) came via the helpful site by Eric Celeste, with some helpul hints from a helpful article by Florian Ledermann.

The data munging was done wtih Python3 in a Jupyter Notebook using Pandas, SciPy.stats and the linear regression models from scikit-learn, and the visualization was done with a big chunk of d3.js (v5). Hint: pure d3 is great on your own time because it will pay you back with better graphics skills, but if someone else is paying by the hour, use Plotly.js if you can! Finally, I couldn't have been good enough with these tools to consider doing all this as a quick project without Danny Clifford and his fellow instructors at the UC Berkeley Data Analytics program. If you like Excel but would love to take it to the next level quickly, it's worth the effort!