Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=Transcript of the talk, from the video's captions. Auto-generated: speaker names in particular are unreliable.
=

# Open Dumps and the Global Trade in Garbage

Authors:   
Discussant: None  
Video: https://www.youtube.com/watch?v=PHYYOyrSItw&t=18773s


## Talk (05:12:53 – 05:21:25)

[05:12:53] >> Thank you. The next speaker is Matt Gordon.  
[05:13:01] >> Hi everyone. Thanks for having me here. Uh I'm Matthew Gordon from the Paris School of Economics. Uh Anna Pat, my co-author is in the back and Monica Shandell is uh starting a new job in Manila soon. Um, so I want to start with  
[05:13:15] two stylized facts here. First of all, the default the average piece of garbage in a low and middle inome country is going to be disposed of in an open dump.  
[05:13:23] And that open dump is going to generate a whole bunch of really bad externalities. Uh, let me just tell you about one of these in particular. So organic waste when it decomposes actually generates a whole bunch of heat. And in some cases that heat can be so intense that it actually  
[05:13:37] spontaneously combusts. the resulting toxic fumes are not that pleasant as you can imagine. Um we don't have really good data though however to to measure these externalities. Um second fact um  
[05:13:52] despite these poor management practices we actually ship a whole bunch of waste from highincome countries to low and middle inome countries. Now what happens to that waste when it arrives? Uh again we don't really have uh very good data  
[05:14:06] on that. Um despite you know the the logic being what Larry Summers once described as impeccable. Um [laughter] uh these questions have become increasingly policy relevant since 2018  
[05:14:20] uh when China disrupted international recycling markets by putting a complete import ban on all recycled materials. Um prior to that they accounted for about 50% of the world's traded waste imports.  
[05:14:33] Um it went to zero within a few years after the ban. and you see corresponding increases in places like Indonesia, Malaysia, and Turkey around the same time. So, uh, what we're going to do in this paper is try to begin to collect  
[05:14:46] the data to study the welfare consequences of this trade and of these practices. Um, and we're going to do it in a in a in a kind of creative way, right? Because we're going to start out with uh data that we know is completely unrepresentative, right? Because we're  
[05:15:01] going to crowdsource it. And we're going to use this seed data to train our machine learning model which uh could be arbitrarily biased or inaccurate. So we're going to call it our trash model.  
[05:15:12] Um and because of these reasons, we're not going to trust the model's predictions. But what we're going to do is we're going to use these models predictions to inform an active learning strategy that at the end of the day is going to give us the first ever globally  
[05:15:25] representative time series data on open dumps. We also go to a few smaller geographies where we're going to try to collect the the universe of open dumps to get a little bit more precise information about those externalities.  
[05:15:38] Um we're going to use this data to learn about where dumps are. Um how many people live close to them, how often they're on fire. Um and we're also going to use it to study the welfare consequences of this import shock. Um,  
[05:15:52] no spoilers, but we find that kind of imports of waste material generate really low economic value relative to the scale of the the externalities that they generate.  
[05:16:03] So to start out, we created this website, the Atlas of Plastic Waste. Uh, we partnered with an NGO uh that asked their membership to submit examples of open dumps around the world that they knew about. Um we use these example  
[05:16:18] dumps along with some other data that we were able to find to to train the first iteration of our machine learning model.  
[05:16:24] Um and this model can recognize the the spectral signature of dumps in satellite data. So on the left here you see an open dump in Albania um in high resolution imagery and on the right you see the brightly colored cluster of  
[05:16:38] pixels corresponding to our our model's predictions about you know whether or not each point is a dump. Um, and you know, so we do a pretty decent job of capturing the the size and the extent of these dumps, but also changes over time.  
[05:16:52] For example, in 2021, they capped this dump and built some football stadiums on top of it, and our model is no longer identifying there being a dump there.  
[05:17:02] Um, still, we don't want to lean too heavily on our model's predictions. Um, for all the reasons that I've discussed, they could be, you know, differentially biased or inaccurate across locations or over time. So instead what we're going to do is we're going to use these  
[05:17:16] predictions to think about where should we go and check whether or not there is in fact a dump there. And the basic process is like this. We can give that high resolution imagery to an expert that can look at it and classify whether or not there is in fact a dump in the  
[05:17:30] image. Um this is time consuming and tedious and expensive. Uh and so we want to use this uh verification procedure kind of sparingly and strategically. Um, and what we learn is that if you want to  
[05:17:45] make our resulting estimates about the scale and prevalence of these dumps as precise as possible, what we want to do is oversample the areas where our model is most uncertain, right? And this allows us to rule out a whole lot of farmland and forests and places that  
[05:18:00] look nothing like dumps um, and oversample the more difficult to classify images. Um at the end of the day this is going to give us like a globally representative data set that we can use you know to study dumps even if our machine learning model's predictions  
[05:18:15] were totally uh miscalibrated or even complete noise. The the benefit that the model gets us is that if our model is good then it'll increase the precision of our resulting estimates.  
[05:18:26] So what do we learn? Uh around the world dumps are concentrated at the 95th and 99th percentile of the population density distribution. Um but we actually see some interesting heterogeneity by income here. Right? So uh in the yellow  
[05:18:40] dots you see that we don't find many open dumps in the Manhattans of the world. Um more in the New Jerseys of the world at kind of the 80th percentile. Um  
[05:18:52] but in in low and middle inome countries we find a whole bunch of dumps in the in the Jakartas and the Mumbai and the really dense cities of low and middle inome countries.  
[05:19:05] um using [clears throat] the China waistband as a natural experiment. Uh we look at whether import shocks of waste lead to more open dumping. Uh and we do find that countries that received more imports after the the China waistband  
[05:19:19] saw an increase in the extent in the prevalence of these dumps. Um and we also find that these dumps seem to be on fire a whole lot of the time. Uh maybe 15% of dumps have at least one fire in any given year.  
[05:19:34] Um, as you can imagine, these fires generate a lot of pretty noxious pollution. Um, and we actually find an increase in pollution in Java, uh, in the areas closest to dumps in the years after the 2018 China waistband, which  
[05:19:48] you can see in the red and and the yellow lines showing elevated pollution after 2018.  
[05:19:54] Um, and we're going to model this change in pollution from those resulting imports flexibly as a function of distance to the dump and wind direction.  
[05:20:03] This gives us these nice looking plume diagrams on the right hand side. So, you see this shows about a 5% increase in particulate matter in the places closest to the dump. Um, and what we're going to do is we're going to use this to do a very kind of simple back of the envelope  
[05:20:18] calculation. We overlay these plumes on all of the dumps in our Java sample.  
[05:20:24] Calculate how many people are affected by what level of particulate matter increase. Uh apply a dose response function from the public health literature. Um and we calculate that about 4,400 additional deaths per year in Java  
[05:20:38] resulted from the influx of imports caused by the China wastpan. Um now people are paying a positive price for these imports. So we can do a very kind of stylized calculation about what does the consumer surplus look like from this  
[05:20:52] supply shock of imports. Uh and when we compare that consumer surplus to the amount of deaths created, we find that the kind of economic value of these imports was about $12,600  
[05:21:05] per premature death with a pretty wide error band. But no matter what, it's kind of orders of magnitude lower than standard estimates of the the value of a statistical life.  
[05:21:16] Um, so then eight minutes was about one minute for every year that we've worked on this project. Um, there's a whole lot more in the paper, but I'm uh looking forward to discussing with you more