Our correlation plots showed that obesity is highly correlated with other diseases. Hence, we decided to try to predict obesity rates using other variables in our datasets. Successful prediction models possibly allow policies to be derived to reduce obesity.

We trained our prediction models with data aggregated by state and county-wise data. While the county data is more extensive, we could include other interesting variables in our prediction models, such as GDP, exercise, and fast food availability. 

We decided to use two different models: Our linear regression model was comparatively simple and easy to understand and interpret. Our random forest regression model is more flexible and hence performs better. However, that's why it is more complex to understand and interpret.

We trained our models on 80% of the data and tested them on the remaining 20%. We split the data randomly. We evaluate the data using the mean squared error (MSE) and R-squared value (R2).  

Obesity Prediction per State

We did not want the models to get too complex as we only had a few data points available. Therefore, we included variables describing exercising, GDP, smoking, and fast food availability. The results show that the simple linear regression outperforms the random forest model. A reason for that is very likely that there are small amounts of available data points. The linear regression model seems able to explain an amount of the variability in the obesity rates using only the four provided variables. The plot of the regression coefficients shows that smoking is the dominant variable.

Obesity Prediction per County

We did not have access to GDP, exercise, or fast food variables for the prediction per county. Therefore, we added drinking, health insurance, self-rated health status, mammography, and dental visits to our variable set. We chose mammography and dental visits as higher values indicate a more heightened health-related awareness and education. 

With a lot more data using county-wise data points, our random forest outperforms the linear regression. We still provide the regression coefficients, as they are easily interpretable, and our model can explain considerable variation within the data. Again, smoking is the most important factor. Health insurance is the second most crucial variable. Having higher rates of insured people could mean more people can afford enough food in the first place.