You are feeling not so well. You go to the doctor. She is a good doctor. She runs some tests, tells you nothing is wrong with you and you leave, ready to get back to work. Why are you so much more ready to work now then you were before you saw your doctor?
Insight into this question is very nicely captured in a new paper by Andrew Dillon, Jed Friedman and Pieter Serneels. Before getting to some of what makes this paper neat, a little background is in order. Dillon and co. are studying the effects of malaria treatment on workers on a Nigerian sugar cane plantation. Workers on this plantation can choose between two tasks: cutting cane and scrabbling (which is collecting the cut cane and packing it for transport). Cutting cane is paid by the amount of sugarcane you cut, and hence gives a really nice measure of productivity. Productivity in scrabbling is harder to observe, so this gets a fixed daily wage -- and a daily wage that is significantly less than the average one would get cutting cane. These workers are randomly (over time) exposed to a testing and treatment program, where those who are found to have clinical levels of malaria get Artemisinin based Combination Therapy (ACT). With this set up, Dillon and co. can unpack what happens when you visit the doctor. There are two big headline results from this paper:
First, workers who test positive respond to treatment both in terms of improved health (which we knew from the medical literature) and in terms of increased labor supply (which we didn't). This leads to an increase in earnings of about 10 percent.
Second, workers who test negative (i.e. do not have clinical malaria) also have higher earnings after being tested, with a lot of this effect being driven by a switch from the lower wage scrabbling to the higher wage (but more physically demanding) cane cutting work. Interestingly, this result seems to be concentrated among workers who were surprised by the negative result.
This second result is important because it shows us the productive impacts of simply providing health information. A really nice paper by Boozer and Philipson in 2000 lays out some of the theory on this (including priors) and then looks at the impacts of getting information on HIV status on sexual behavior. And all of the published literature in this vein since then (to my knowledge) has similarly focused on sexual behavior. With Dillon and co. we get some insight into how this information can shift work responses.
So that is what’s neat on the level of the results. The way they get to them is also helpful for thinking about how we might do more of this kind of analysis in the future (especially with medically proven interventions).
Dillon and co's identification comes from the fact that the testing and treatment was rolled out randomly across and within eight different work groups (nb: all workers are ultimately offered the test and (possibly) treatment, the only random component is the timing). This lets them look at a number of things. First, to look at the combined impact of testing on both the positive and negative individuals, they compare those who get testing at some point in time with those who get it later (the fact that the treatment takes awhile to kick in helps them here) -- this is the intent to treat estimate. Second, to look at the effect on those who have malaria, they look at those who test positive (and are treated) and compare them with people who test positive later (but don't get treated at the same time due to their random test date). Here, they are helped out by the fact that malaria outbreaks last 14-17 days. This approach gives them an estimate of the effect of treatment on the treated. Finally, to get the result on those who test negative, they compare those who test negative now with those who test negative later (they call this group the TmUT - the treatment on the medically untreated).
An example might make this clearer. To look at the one week effects of treatment (they show effects over a range of weeks) they will compare 2nd week outcomes for workers who were tested in week 1 (the treatment group) with those who will be tested in week 3 (the control group) and they will also compare 3rd week outcomes for those tested in week 2 with those to be tested in week 4, etc.
A couple of thoughts. First, when they look at the effect on just the positives, they are comparing workers who are known to be sick (the treatment group) with those who will test positive in the future (the control group). As the authors point out, this requires the assumption that those who test positive in the control are also going to be experiencing the illness at the same time as those in the treatment group. The first point to allay any concern with this is the 14-17 day length of malaria outbreaks. However, if there are some folks who test positive later (i.e. in the control group) who aren't actually sick when the treatment group gets tested, this suggests their estimates are actually a lower bound.
Second, when they look at the effect on the negatives, a similar assumption is required. That is, those who test negative in later weeks (the control) are also negative when then the treatment group is tested. On the face of this, this is more concerning since the control group could be sick, but recover by the time they get tested. But there are a number of reasons to be not so worried about this. First, the result that a lot of the effect for the negative population is being driven by task choice rather than just labor supply would require a pretty weird pattern of behavior to be consistent with a violation of this assumption. Second, Dillon and co actually asked people if they felt sick in the past four weeks. When they re-run the estimates on just this sample of the malaria-negative folks, the results don't change. Finally, they also ask about fatigue (a common symptom of malaria) and they measure the number of malaria parasites in individuals' blood (side note: a lot of these folks have some). Those who are more fatigued, or have higher (but not high enough for a positive diagnosis) levels of the parasite, show positive wage and earnings effects, while those who aren't, don't. This suggests folks who are more likely to believe that they have malaria, and hence be surprised by the result, are the ones for who this effect really hits home.
All in all, this is a thought provoking paper both in terms of its results -- which show us not only how malaria treatment can pay off, but also how health information matters for how we work -- but also for its methods.
Insight into this question is very nicely captured in a new paper by Andrew Dillon, Jed Friedman and Pieter Serneels. Before getting to some of what makes this paper neat, a little background is in order. Dillon and co. are studying the effects of malaria treatment on workers on a Nigerian sugar cane plantation. Workers on this plantation can choose between two tasks: cutting cane and scrabbling (which is collecting the cut cane and packing it for transport). Cutting cane is paid by the amount of sugarcane you cut, and hence gives a really nice measure of productivity. Productivity in scrabbling is harder to observe, so this gets a fixed daily wage -- and a daily wage that is significantly less than the average one would get cutting cane. These workers are randomly (over time) exposed to a testing and treatment program, where those who are found to have clinical levels of malaria get Artemisinin based Combination Therapy (ACT). With this set up, Dillon and co. can unpack what happens when you visit the doctor. There are two big headline results from this paper:
First, workers who test positive respond to treatment both in terms of improved health (which we knew from the medical literature) and in terms of increased labor supply (which we didn't). This leads to an increase in earnings of about 10 percent.
Second, workers who test negative (i.e. do not have clinical malaria) also have higher earnings after being tested, with a lot of this effect being driven by a switch from the lower wage scrabbling to the higher wage (but more physically demanding) cane cutting work. Interestingly, this result seems to be concentrated among workers who were surprised by the negative result.
This second result is important because it shows us the productive impacts of simply providing health information. A really nice paper by Boozer and Philipson in 2000 lays out some of the theory on this (including priors) and then looks at the impacts of getting information on HIV status on sexual behavior. And all of the published literature in this vein since then (to my knowledge) has similarly focused on sexual behavior. With Dillon and co. we get some insight into how this information can shift work responses.
So that is what’s neat on the level of the results. The way they get to them is also helpful for thinking about how we might do more of this kind of analysis in the future (especially with medically proven interventions).
Dillon and co's identification comes from the fact that the testing and treatment was rolled out randomly across and within eight different work groups (nb: all workers are ultimately offered the test and (possibly) treatment, the only random component is the timing). This lets them look at a number of things. First, to look at the combined impact of testing on both the positive and negative individuals, they compare those who get testing at some point in time with those who get it later (the fact that the treatment takes awhile to kick in helps them here) -- this is the intent to treat estimate. Second, to look at the effect on those who have malaria, they look at those who test positive (and are treated) and compare them with people who test positive later (but don't get treated at the same time due to their random test date). Here, they are helped out by the fact that malaria outbreaks last 14-17 days. This approach gives them an estimate of the effect of treatment on the treated. Finally, to get the result on those who test negative, they compare those who test negative now with those who test negative later (they call this group the TmUT - the treatment on the medically untreated).
An example might make this clearer. To look at the one week effects of treatment (they show effects over a range of weeks) they will compare 2nd week outcomes for workers who were tested in week 1 (the treatment group) with those who will be tested in week 3 (the control group) and they will also compare 3rd week outcomes for those tested in week 2 with those to be tested in week 4, etc.
A couple of thoughts. First, when they look at the effect on just the positives, they are comparing workers who are known to be sick (the treatment group) with those who will test positive in the future (the control group). As the authors point out, this requires the assumption that those who test positive in the control are also going to be experiencing the illness at the same time as those in the treatment group. The first point to allay any concern with this is the 14-17 day length of malaria outbreaks. However, if there are some folks who test positive later (i.e. in the control group) who aren't actually sick when the treatment group gets tested, this suggests their estimates are actually a lower bound.
Second, when they look at the effect on the negatives, a similar assumption is required. That is, those who test negative in later weeks (the control) are also negative when then the treatment group is tested. On the face of this, this is more concerning since the control group could be sick, but recover by the time they get tested. But there are a number of reasons to be not so worried about this. First, the result that a lot of the effect for the negative population is being driven by task choice rather than just labor supply would require a pretty weird pattern of behavior to be consistent with a violation of this assumption. Second, Dillon and co actually asked people if they felt sick in the past four weeks. When they re-run the estimates on just this sample of the malaria-negative folks, the results don't change. Finally, they also ask about fatigue (a common symptom of malaria) and they measure the number of malaria parasites in individuals' blood (side note: a lot of these folks have some). Those who are more fatigued, or have higher (but not high enough for a positive diagnosis) levels of the parasite, show positive wage and earnings effects, while those who aren't, don't. This suggests folks who are more likely to believe that they have malaria, and hence be surprised by the result, are the ones for who this effect really hits home.
All in all, this is a thought provoking paper both in terms of its results -- which show us not only how malaria treatment can pay off, but also how health information matters for how we work -- but also for its methods.
Join the Conversation