big data in insurance
Jonathan Buchanan's Winter 2018 Independent Study
Today, my plan was to research the connection between home health and opioid use as much as I could, as well as develop forecasting models for the year-over-year growth of home care spending. I wanted to do research to narrow my focus to how medicare beneficiaries (typically 65 or older) are affected by the opioid epidemic because that's the population my data deals with. A historical context of home health care in the United States was also something I wanted to research because that might help me better understand the trends I saw in funding. Overall, my goal for today was to zoom in on the new trend I discovered in my data and research why that might be happening.
To start off my research, I searched the internet for the history of home health care in the United States. I found this paper, which was very helpful because it detailed the history from the creation of Medicare (1965) up to the present and even predicted future trends. Since my spending data covers 1981 to 2014, this was the time period I would be most interested in learning about. What I found is that home health care was least regulated after 1980 due to the Omnibus Reconciliation Act, which removed regulations on the number of visits, requirements for prior hospitalization, etc. In the graph, this de-regulation is reflected by jagged, sharp spikes. In 1997, with the passage of the Balanced Budget Act, an Interim Payment System was enacted to lower spending. However, this system cut funds much lower than expected, resulting in the only section of the chart where growth is negative. However, in 2000, the Prospective Payment System was put in place to replace the IPS, which led to higher funds but controlled spending, so the graph is smoother over time and does not contain any large spikes like the beginning. I also spent some time developing a forecasting model for the growth rate of home healthcare spending. My sponsor suggested I use a model called ARIMA, which stands for Auto-Regressive Integrated Moving Average. This model is generally good at predicting future values for a stationary time-series, which is what my data appears to be, at least post-2000. The ARIMA model accepts three different parameters: the first parameter p indicates the lag (delay) of the auto-regression, the second parameter d indicates how much to difference (integrate) the model, and q indicates the number of terms used for the moving average. After examining the ACF and PACF charts as well as experimenting, I found that the best model has parameters of (1, 1, 5). The model is not effective for predicting several years in the future. It will always tend towards the average. However, the model is useful for generating predictions for a year or two. You can see my model in the gallery above. Today was a little different because I didn't spend a lot of time looking at new data. Instead, I did research in a more contextual or historical sense, which was a nice break from looking at data all the time. It also helped me gain an understanding of what I was working on and why the trends I'm seeing might be happening. The forecasting model was also interesting because it was something I was completely new to and it was nice to learn about something like that. Tomorrow, I plan to make an objective comparison of the error of each model by testing with several sets of parameters and recording the error that each has.
1 Comment
Matt Buchanan
1/18/2018 08:57:10 am
This is really interesting, and I'm glad to see you research the historical perspective. Data can tell you a lot about a situation, but history can usually put things into a perspective that somehow provides more understanding.
Reply
Leave a Reply. |
ArchivesCategories |