Tracking Exercise Data Trends with NHANES
The skills the author demoed here can be learned through taking Data Science with Machine Learning bootcamp with NYC Data Science Academy.
Contributed by Thomas Kassel. He is currently enrolled in the NYC Data Science Academy remote bootcamp program taking place from January-May 2017. This post is based on his second class project, Shiny, focusing on the creation of an interactive web application using the Shiny package in R.
The National Health and Nutrition Examination Survey (NHANES) is one of the foremost assessments of health data statistics for children and adults in the United States. Sponsored by the Centers for Disease Control, NHANES combines interviews with physical examinations and laboratory tests for approximately 5,000 Americans each year. Results are compiled, anonymized and made publicly available at the program’s website on a rolling basis. These datasets are a major source of information for further studies ranging from simple national averages for physiological measurements (height or weight) to epidemiological trends for public health policy reform.
American Health Stats App
The 2013-2014 NHANES publication is made accessible through a Shiny application, providing an interactive environment for users to explore data trends. Because of the wide range of information gathered by the survey, the app focuses on a subset of the findings, drawing particular attention to the prevalence of exercise and associated health outcomes across demographic groups.
To use the live application, navigate here.
NHANES divides results into many separate tables for more manageable file download sizes. The app makes use of the following datasets (including but not limited to):
- Demographics – gender, age, ethnicity, education, annual household income, household size
- Exercise – minutes of various forms of physical activity per day
- Examination/Laboratory – physiological measurements
The common field across tables is the sequence ID, a unique identifier for each survey participant which can be used for joining operations across any/all tables. Let’s take a look at the initial raw data import of many NHANES tables and an example dplyr left_join operation to quickly aggregate information by survey participant.
The app is meant to be used as a starting point to explore high-level trends in the latest NHANES survey publication. It is recommended, however, that the program's survey methods and analytic guidelines be followed in order to draw any robust statistical conclusions from the underlying sample data.
What demographic cohorts are covered by NHANES?
The “Overview & Demographics” tab allows users to explore this question. NHANES appears to be a nationally representative sample of the U.S. population across gender, education level and income. However, certain sub-populations are more heavily represented, such as youths under 18 still undergoing much of their physiological development and at therefore at higher focus for many public health concerns.
Which populations regularly engage in physical activity?
The “Explore Exercise Trends” tab provides exploratory visualization of who is getting exercise, in what form and amount, and the possible health outcomes associated. Density curves help to illustrate exercise trends and confirm preconceptions about some of the grouping (i.e. demographic) variables involved in those trends.
Below we observe that, with each drop in level of education – from college/advanced graduate down to high school dropout – the density curve’s average moves from left to right. This implies a direct inverse relationship between minutes per day conducting “vigorous or moderate” physical activity at the workplace, i.e. manual labor, and education.
What health outcomes correlate with physical activity?
Minutes per day of physical activity – whether in the workplace, recreationally; vigorous, moderate or otherwise – are used as the predictor variable in a scatterplot to explore correlations with a variety of health outcomes such as weight, BMI, cholesterol, or blood pressure.
Above we observe a clear negative correlation between resting pulse (bpm) and average minutes per day of vigorous recreational activity. Through the use of the interactive graphing package Plotly, we gain further information from mouseover tooltips in this visual: among participants getting an average 180 minutes of exercise per day, none had a resting pulse of over 98 bpm.
Other relationships explored, such as cholesterol levels versus exercise habits, highlight the limitations of taking a bivariate approach in predicting health outcomes. For example, we might assume a negative correlation between amount of exercise and LDL cholesterol levels, but in the case below we observe this correlation to actually be positive. This suggests the study could benefit from other known associations to cholesterol, such as diet (perhaps individuals with moderate-exercise jobs tend to have higher-cholesterol diets) and genetic disposition.
As alluded to in the previous section, inclusion of more variables from the original NHANES dataset and the use of formal statistical methods could help to tell a more comprehensive story about the program data. Chief among these would be an exploration of dietary/nutritional and lifestyle habits (e.g. smoking, drug use), since NHANES has collected that data for each participant. As NHANES is an ongoing and continuously developing program, there is further potential to track the movement of these health trends over time in the general U.S. population.