BP Oil Spill

In April, 2010 the Deepwater Horizon drilling rig exploded off the cost of Louisiana, starting one of the largest environmental disasters in US history that ultimately leaked 4.9 million barrels of oil into the Gulf of Mexico.

In 2011, I finished grad school with degree in applied economics, and an interest and experience with environmental problems.

It's not that surprising that I soon started working for an environmental economics consulting company contracted to the government to (among other things) quantify the value of the lost recreation on Gulf beaches due to the spill.

Trips lost due to the spill

The lost recreation value due to the spill basically came down to:

I worked mostly on quantifying the number of forgone trips. In theory:

Both numbers on the right hand side involve varying degrees of guessing.

Take the easier one: number of trips that actually took place. How would you go about calculating that?

"Ideally", maybe you have turnstiles at the entrance to every single beach in the gulf, then you can just count up the totals and see. In practice, obviously that's impossible

Instead we did some standard surveying and sampling techniques, dividing the gulf into segments and flying a plane up and down the coast multiple times a week snapping aerial photos. The photos were sent back to company headquarters, where we had a team that would count the people.

But snapshots only tell you the number of people at any particular moment in time. To go from that to full trips you need to get some sense of how long people were there.

For example, if everyone at the beach got there early and stayed all day, then the n of people in your picture == n of trips. Everyone was just there all day.

If everyone stayed for exactly (and randomly) half the day, then you'd have to multiply your snapshot count by 2. You counted the people there during the time you snapped a picture, but each person at the beach that day only had a .5 probability of being in the picture.

We got these durations (among other information) from a staff of 100+ people on the ground conducting interviews.

The result is cascading series of multiplying:

n people in a photo * 1/average length of time people stayed * 1/(how often you took a picture)


And that's just to get the actual number of trips taken post spill, which is the conceptually easier number. More abstract is the hypothetical: how many would people have taken had there been no spill?

Estimating the counterfactual

A few possible ways to approximate the number of trips people would have taken if there had been oil spill:

  1. Figure out how many people took trips in 2009 (right before the spill).
  2. Look at how many people took trips to places not affected by the spill, extrapolate based on that.
  3. Figure out how many people took trips after things returned to normal.

All of these have issues. No one had the data for (1) — doing all the counting and interviewing for actual trips in the aftermath of the spill was very expensive, and no one had been doing anything like it pre spill.

(2) required good, comprehensive data on how the spatial impacts over time, which was another effort. Also it was possible more people than usual might be visiting non-affected sites, which would make the true number harder to calculate.

That left (3), which had its own issues (How do you know when things have returned to normal? What if things never did? What if people who hadn't been taking trips went all at once after things got back to normal, overstating the "baseline" n of trips lost?)

To deal with that, the government (specifically NOAA, who was in charge) decided we should continue counting people on the beach until the number of trips leveled off and stayed that way for a year.

To deal with seasonality inherent in recreation trips to the gulf, that required two years of sampling after things had returned to normal (a year longer than strictly necessary) in order to figure out what "normal" was.

Controlling for weather

Another issue with using later, "back to normal" years as a baseline is potential uncontrolled for differences in other factors that might effect trips.

For example, imagine the weather in 2010-2011 (when oil was affecting recreation trips) was unusually cold. Pretend it was freezing, too cold for anyone to go to the beach

What would have been the impact of the spill on recreation in that case? Nothing! If the weather was so bad and cold that people wouldn't have taken any trips anyway, then the spill didn't actually have any effect on recreation.

Obviously in real life, the weather wasn't that bad, but we had to account for issues (another one was differences in gas prices between years) like it.

Officially, we did so using some fairly basic statistical techniques (adjusting broadly for good and bad weather days across years), but — as a check — we also built a more complicated parametric model that gave us more fine grained control, I helped design and did all of the coding for that, which was fun.

Me and BP

When I got there in the fall of 2011, the data collection effort was well underway, but we hadn't started turning that raw data into counts of overall trips, much less trips lost.

My role became programming up all processing, multiplying and adding. I did it in Stata, which I had used a bit previously, but not that much (I haven't used it in since ~2013).

Since it was for litigation with a lot stake financially, we had a subcontractor doing the same weighting-up coding, but in SAS.

I have many (fond?) memories of talking through obscure data edge cases with him, trying to figure out why our trip totals (totaling in the 100s of millions) differed by 17 trips or whatever.

Aside from the unfortunate circumstances, I enjoyed working on the project. We had a team of big name academic experts, and it was fun hashing things out with some of the same people that had written my college textbooks and coding up their ideas.

Working on it was a big part in me realize I liked programming (and data and analysis), even more so than working specifically on environmental problems per se.