For the past five years at least, I have been suffering from migraines. Though I don’t suffer from “classic” migraines–I don’t have the painful headaches that force one to lie down in a darkened room, for instance–my migraines are debilitating enough, as they make me sensitive to light (and sometimes to noise), they can last up to seven days, and in the worst of cases, I have severe nausea and have to spend a few days in bed while I recover.
My migraines have gotten a lot better since coming to Minnesota, thank God, but one of the things I have been abstaining from these past few years has been drinking wine, because wine is often thought to cause migraines. The upside is that I have discovered the wonderful world of whiskeys as a result of not drinking wine; the downside is… Well, having to abstain from wine is a serious downside.
After a few years of not drinking wine and still suffering from migraines, I decided at the beginning of this year to determine once and for all whether my migraines were actually caused by drinking wine. So I did something simple enough: I ran an RCT on myself. I took a spreadsheet and, for a period of 50 days, I decided to randomize each day into treatment (i.e., a day where I would drink wine) and control (i.e., a day where I would abstain from wine).
(And before the critical thinkers among you chime in with the same comment: Yes, I know that this does not allow doing a double blind study, thanks.)
Further, within each treatment day, I randomized how much wine I should drink (i.e., 0.25 or 0.5 cups). In my spreadsheet, other than the date and the number of the observation (t), I thus created two columns: wine_treat (a dummy for whether a given day was randomly assigned as a “wine” day), and q_treat (a dummy for how much wine I was to drink on a given wine day). These variables can be used to estimate the intent to treat (ITT).
I know myself too well, however, and I knew that I might have a hard time sticking to random assignment. So I made sure to create two more variables, whose names are self-explanatory: wine_actual and q_actual, which measure whether I actually stuck to a day’s random assignment and how much wine I really had that day. These variables can be used to estimate the local average treatment effect, instrumenting _actual variables with _treat variables.
Finally, I decided to add in some controls: red, a dummy variable for whether I had had red wine (red wine being thought as “more likely” to give migraines by the medical profession than white wine), other_alc for whether I had other kinds of alcohol that day (i.e., whiskey, beer), and stress for whether this had been a particularly stressful day. And of course, for each date, I have a variable called migraine, which measures whether I have experienced migraine onset the next day (when I have a migraine, I almost always wake up with it and to my knowledge, I have never had a migraine start later during the day, after the normal time where I would have wine, i.e., dinner time).
You can find the data here in Excel format and have fun with it, but in short:
- Regressing migraine onset the next day on random assignment of whether I should drink wine on a given day yields a statistically insignificant ITT estimate, with or without controls (i.e., other alcohol, stress, and a linear trend). Even if that ITT were significant, it would be negative, i.e., assignment to treatment would be associated with a decrease in the likelihood of migraine.
- Likewise for the ITT estimate of how much wine I should drink.
- Regressing migraine onset the next day on whether I actually drank wine on a given day instrumented by random assignment of whether I should drink wine yields a statistically insignificant LATE estimate, with or without controls. Note, however, that in both cases, the first-stage F-statistic is less then 10, i.e., there is a weak instrument problem. (That is just how bad I am at following instructions!)
- Likewise for the LATE estimate of how much wine I should drink.
- This could well be underpowered. For n=49 (I lose one observation because migraine onset is a lead variable), I experienced migraine onset only once. Great for me; bad for statistical power.
- Nevertheless, there came a time where I just “knew” that the effect of wine on onset of my migraines was not causal. It’s difficult to explain when that time came, or why, but I guess if you do something often enough without something else happening, you just make the induction that the former does not cause the latter.
Finally, in an effort to check whether stress makes me drink, I created a dummy variable for whether I defied control, i.e., a variable equal one on those days where I was assigned to not have wine but decided to have wine nevertheless, and equal to zero otherwise. It turns out that being stressed out made me less likely to have wine when I was not supposed to!
This all goes to show that it is certainly possible to use the methods we currently use to evaluate impact to make your own life better. As the descendant of French immigrants to North America, I could not imagine living the rest of my life without wine.