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.
=

# Green Subsidies with Demand Distortions

Authors:   
Discussant: None  
Video: https://www.youtube.com/watch?v=PHYYOyrSItw&t=24525s


## Talk (06:48:45 – 07:39:10)

[06:48:51] uh very last session. I'll try to make it worth uh your while. This paper is joint with Josh Dean who is here with us as well. Um, we're studying green subsidies in the presence of uh demand distortions. Thanks also to the organizers for putting this together with the the previous paper. I think  
[06:49:06] they they fit very nicely together. Um, so this is a paper about development economics and environmental economics.  
[06:49:12] Pigu um tells us that we should correct externalities positive or negative externalities at the margin. So the optimal policy would be for example to tax uh some amount that was equal to the marginal externality cost or if it was a  
[06:49:26] benefit to offer a marginal subsidy on this right things like a gasoline tax aim to do this. This corrects different margins for example when you think about a gasoline tax if you drive a lot there's selection people who drive a lot would save a lot more by buying a more  
[06:49:41] fuel efficient vehicle and you might also be encouraged to drive less. So that's a treatment effect right? So what's nice about these marginal pricing instruments is they correct all of these different margins. And this has inspired many programs across the world uh that  
[06:49:54] clean that subsidize the clean uh technology usage. And so some examples of this um for marginal cost subsidies, right? That's things like charging your electric vehicle or uh using your electric cookware, right? That's  
[06:50:08] distinct from a fixed cost subsidy that I'll go into. But if you own for example an electric vehicle, you might have an electricity price that is subsidized to encourage you to use that electric vehicle uh as well. Volvo has started paying for this as well. And we're  
[06:50:22] seeing this in Africa as well. So uh Uganda, sorry, Kenya has an electric vehicle tariff. Uganda has a cooking tariff. Uh and Kenya is currently debating whether they should have an electricity tariff specifically for people using electricity to cook uh as  
[06:50:36] well. Right? And so these are marginal pricing schemes that's distinct from a fixed cost subsidy where we're subsidizing just the technology itself.  
[06:50:44] So let's say some reduction in the price of buying the technology. Again we're seeing this with electric vehicles with heat pumps and then we're seeing it in Africa as well. Again Kenya is currently considering having a tax waiver for  
[06:50:57] importing electric stoves. uh and anytime you have an off or many times offset schemes also operate as fixed cost subsidies where there's transfers for every uh let's say improved stove uh that was sold. So we're interested in  
[06:51:11] these types of subsidy schemes in the presence of market distortions which everyone in this room knows are really prevalent in low and middle inome countries. Uh what this is going to mean is that we're going to simplify this.  
[06:51:24] We're going to uh characterize a market distortion as some gamma times your efficient willingness to pay. So normally we think of your efficient willingness to pay as your total private welfare gain from the technology. And we're going to say that what we observe  
[06:51:38] is some willingness to pay that is lower than that level. And this could be caused for example by inattention to the benefits. It could be caused by present bias or it could be caused by credit constraints. That's going to be the one that we'll focus on in this paper. uh  
[06:51:52] and we're going to hold this fixed. We're not going to study whether you might want to solve or address those credit constraints. We're saying because of informality, imperfect property rights, institutional constraints, etc., these are ingrained and we're trying to understand how to correct environmental  
[06:52:06] externalities in the presence of these types of distortions. Um you could characterize distortions in different ways. We're just going to study this relatively straightforward linear uh multiplier here. And so we're again in theory of the second best, right?  
[06:52:21] Piggovian marginal pricing may not be the optimal policy here. And so the first finding of this paper is going to be a theoretical contribution that these demand distortions can actually increase the efficacy of fixed cost subsidies.  
[06:52:35] And there's going to be two channels and I'll keep going back to these two channels throughout the talk today. So first, if private benefits and social benefits are positively correlated. So if you drive a lot, you're spending a lot of money on gasoline, you're also  
[06:52:49] emitting a lot. So those are going to be positively correlated. Then distortions that reduce adoption will increase the positive externality that's generated from the marginal adopter. All right? So that's a selection mechanism. That's the  
[06:53:02] first one. The second is that demand distortions can often increase demand elasticity and that's going to lower your subsidy expenditure that's required for uh for [clears throat] incentivizing one additional adoption.  
[06:53:16] And so then uh the question is these are regarding fixed cost subsidies. So we know that marginal cost subsidies also have advantages right these selection and treatment effects. And so are these benefits of fixed cost subsidies sufficient to overcome uh those  
[06:53:31] shortcomings. And the answer in our context and we're going to argue in many contexts is going to be yes. that the uh the demand distortions that are present here are going to increase not just the efficacy of fixed cost subsidies  
[06:53:43] relative to marginal cost subsidies but actually in absolute terms it's going to make every subsidy dollar more efficient in terms of how many tons of CO2 you can reduce I'll go through uh the theory for  
[06:53:58] this in just a second but I'll just mention that we're going to study this in the context of cook stoves just to give some background billions of people cook with uh biomass mass every day.  
[06:54:07] This is usually wood or charcoal. And just to give you a sense of the magnitude, for most households, this is their most polluting uh device, an average Kenyan household in our sample, when they cook all of their meals with charcoal, most of their meals, they emit  
[06:54:21] about four or five tons of CO2 per year doing so. The average American household that drives an average gasoline vehicle, an average number of miles, also emits between four and five tons of CO2.  
[06:54:34] Right? So this is the equivalent of billions of gasoline vehicles driving around. And so we also think this particular sector is of importance in and of itself.  
[06:54:44] Okay. So we're going to uh enroll households who are currently using a Kenyan charcoal stove, a GCO, and we'll offer them the opportunity to buy an induction stove and that comes with three pots that you can use on that stove. Uh so we'll have uh 2,000  
[06:54:58] households in the sample. We'll randomize three things. So marginal cost subsidies, so the cost of actually using the electric stove. There's a you have to pay the electricity company for those uh units. We're going to subsidize those units. We're going to randomize fixed  
[06:55:12] cost subsidies for the actual price of the stove, either 10% or 75% off of the standard price of $82. And then we're going to randomize the this demand distortion that we're interested in, which is credit constraints. So we're going to randomize whether you're paying  
[06:55:26] that upfront cost either uh with a loan or upfront. And then we'll have high frequency measurement of both charcoal stove usage as well as induction stove usage uh as well as inperson follow-up surveys. Um and that uh the two-minute  
[06:55:40] frequency temperature monitor. You can actually see that on the picture here on the left. There's a little temperature sensor that one of our field officers installed on one of their respondent cook stoves.  
[06:55:51] Okay. So I'll start just by presenting reduced form impacts. So adoption of the induction stove reduces your total energy spending by about $9 a month. The reduction reduction of about $10 in charcoal spending, a little bit of an increase in electricity spending, and  
[06:56:05] then some changes in wood and and LPG spending as well. Uh that's about $184 discounted over the 1.9year average lifespan of the stove. uh when you convert all of the the charcoal,  
[06:56:18] electricity, um gas uh emissions that are generated from from all those different fuels, when you take the total, we estimate a reduction of about 2.5 tons of CO2 for every 12 months that  
[06:56:31] the improved stove is in use. Um that is more than the best estimates about when a US household switches from a gasoline vehicle to an electric vehicle, the annual reduction that that achieves.  
[06:56:43] Um the demand distortion is significant here in the control group where people are paying uh without a loan willingness to pay is $20. With a loan average willingness to pay is $35. So if we think the private benefit here is $184.  
[06:56:58] Then the willingness to pay here uh the observed willingness to pay is basically that number multiplied by the distortion that we observe. So the distortion is 0.11 for people without a loan. So they're only willing to pay basically  
[06:57:11] 11% of the total benefit with a loan that willingness to pay increases to 19% of the benefit.  
[06:57:18] Okay. Uh we find that the marginal cost subsidies have no impacts on either willingness to pay or usage. This is despite us making them very salient, handing out a lot of information, flyers about how much it would cost to cook, etc. I'll show those in in just a  
[06:57:33] second. So how do the demand distortions affect the efficacy of the the fixed cost subsidies that we offer? So I mentioned the two channels. The first is that the demand distortion changes who the  
[06:57:48] marginal adopter is. So we see that marginal adopters in the group that is more distorted that is paying in cash.  
[06:57:55] They generate larger private savings $200 instead of 154. And they also generate a larger total emission reduction, five tons of CO2 instead of 4.2 tons of CO2. Right? So that's the first channel is that in this more  
[06:58:08] distorted world, you're incentivizing higher users, people with higher private and social benefits to adopt. The second channel is that because the demand distortion increases demand elasticity, it requires less subsidy spending to  
[06:58:22] incentivize each additional adoption. And so there's more abatement per adoption and it's a lower cost per adoption. So when you combine these numbers, we can either express this as a welfare gain per dollar of uh of  
[06:58:37] government of subsidy spending. So that increases by 78%.  
[06:58:41] Uh and you this is including both the environmental benefit as well as the private fuel savings and the cost of the stove. Um, or you can think about what is the subsidy spending required to abate one ton of CO2 and that goes down  
[06:58:53] from $22 to $13 in the more uh distorted group by 40%.  
[06:58:59] So, one question that you might ask is what would happen if we fully eliminated the demand distortion, right? Even the group where there's less of a demand distortion that has increased their willingness to pay, they're still only willing to pay 19% of the total value.  
[06:59:14] And so we're going to estimate the model to try and extrapolate uh beyond that.  
[06:59:18] So we're going to use two-step generalized method of moments to estimate five parameters by matching to seven moments. And we're going to use the random variation from the experiment to estimate this. And so at the 10% level with the the first bullet here,  
[06:59:31] which is basically what Kenya is at at the baseline, um the cost is to the cost to reduce emissions is $13 per ton of CO2.  
[06:59:42] If we run the counterfactual where people are actually willing to pay what they would save, the cost per ton goes up to 137. So it increases by uh more than a factor of 10.  
[06:59:54] So what does this mean for environmental policy more broadly? So across the world uh governments and private sector, philanthropists, investors, companies, offset markets combined are financing more than a trillion dollars in carbon  
[07:00:09] mitigation every year. There's lots of concerns about adverse selection and moral hazard that are undermining the efficiency of this market. And so what we should do is fi estimate the marginal abatement supply curve and identify the  
[07:00:23] lowest places the cheapest places where globally we should be abating. And so our point with this paper is to say that it's possible and likely that many of those lowest cost opportunities are going to be in locations in lower middle- inome countries where demand  
[07:00:38] distortions are larger. So just as an example, we're focusing on capital markets where uh we know that there are significant credit constraints uh uh at play. So just as an example for when we think about electric vehicles, let's say  
[07:00:50] in the US um and this is gasoline vehicles, 80% of new cars are sold on credit with an average APR of 6%. Uh and if you calculate the subsidy cost to uh abate one ton of CO2 for that, it's  
[07:01:05] $1,300. In Kenya, even with the digital collateral, so this is a remote shut off, all of these things for the induction stoves that we sell, if you wanted to buy that with credit in the in a shop, you would have to pay an APR of  
[07:01:19] 224%. And so what we would argue is the kind of the the 10 the 10 times multiplier factor here is part of the explanation for why this cost is so much lower than this first cost. Of course, it's partly also a difference in technologies, but  
[07:01:33] that's why we use the random variation to kind of to kind of disentangle that.  
[07:01:37] Okay, any questions at this point? >> Sorry, quick question. Why is the lifespan only 1.9 years? Am I missing something?  
[07:01:47] >> Um, it's a good question. I think it's partly behavioral. I think partly um there's we don't know and actually we have food diaries so we know exactly you know many people who buy the induction  
[07:02:02] stoves will continue cooking some of their meals with charcoal. So you might cook your rice or something else with one stove and then something else with the induction stove and at one point you might decide not to or it might break or you might decide to cook your meals with a different stove. We're not sure. We're  
[07:02:16] still trying to figure this out. Uh the 1.9 is from historical data from stove sales going back uh at least three years or something. So we're looking at when people are still using those stoves.  
[07:02:28] Okay, great. Um so uh again uh I'm going to present a simplified version of the model that's in the paper. We're going to assume that everybody has inelastic energy services, right? So you cook a certain number of meals every day or you drive a certain number of miles every  
[07:02:43] day. We're going to assume that that's perfectly inelastic. And you can satisfy that either with a dirty or a greener version of that technology. You're going to get some private fuel savings. That's the difference in price per unit. And you're going to generate some externality. That's the difference in  
[07:02:57] the emissions. So your demand for the green technology is whatever your private savings are multiplied by the gamma that we're interested in here.  
[07:03:07] Market demand is just going to say for any uh given subsidized price P minus S1, what fraction of the market or what number of people uh have a willingness to pay uh that is above that price? And  
[07:03:19] then there was a question. Yeah, >> perfect.  
[07:03:31] It seems like people might have >> Yes. In the paper, we have a preference parameter and I'll mention that in just a second. I'm just presenting the simplified version just to be able to show a simple graph before I do that.  
[07:03:40] Yeah. Yeah. I thought you had your hand up.  
[07:03:48] >> Yeah. >> Yeah. So, it could be in attention people are just not realizing the cost  
[07:04:03] difference. It could be that people really prefer some of their foods to be cooked on charcoal that the taste is very different. And so, we're actually currently running an SMS experiment to try to see which of those it is. Uh we're we're still waiting on the data  
[07:04:15] for that. Great. Okay. So, um we can we can align people by their willingness to pay, right? So, if you drive a lot, you would save a lot of money from a greener technology. So you would have a high private you would  
[07:04:29] have a high uh willingness to pay a high private marginal benefit. And so anybody with uh uh private benefits above the market price would adopt in this context. Right? We're assuming a linear distortion here. So in a distorted  
[07:04:43] market that demand curve would rotate inward. Right? You're just multiplying the willingness to pay for the green technology by and we're assuming here a homogeneous uh distortion. I can talk about heterogeneous distortions in in a  
[07:04:56] second. Yeah. So credit the credit constraint. So you're earning $100 in fuel savings over the next two years, but you don't have access to those future savings. And so today you're only willing to pay 10 $10.  
[07:05:11] >> It's again critically important. >> What's not obvious a credit like that? A credit might be like you know I pay up to 10 bucks whatever fine but more than 10 I  
[07:05:25] have to borrow. I mean this attack >> it's it's absolutely true that the distortion could look a lot of different ways. I think we're just picking the the linear version just to show the intuition. The model is much more  
[07:05:39] flexible. Um and I think the question also is empirically what do we observe and so we'll be able to test that with the RCT. Yeah.  
[07:05:46] >> What's an example? Oh, I mean if you have an interest rate your your So 100% times whatever your private benefit is. Yeah. Exactly. Yeah.  
[07:05:59] Okay. So willingness to pay is dampened but crucially among people who do adopt the externality is still proportional to your private benefit. Right? So the externality curves are going to be the  
[07:06:12] same for these populations. And so if we look here at this market price, who is the marginal adopter? On the left, it's given by this fee and on the right, it's somebody with a larger externality because it's somebody with a larger  
[07:06:25] private benefit as well. And so now we can think what would happen if you introduce a fixed cost subsidy here, right? In the efficient in the in the market where gamma equals 1, this would be the total positive externality  
[07:06:38] generated by the fixed cost subsidy. on the right. This would be the total externality generated. Right? So in a distorted market, the marginal adopters have a larger posit have a larger marginal positive externality. Uh they  
[07:06:51] also have a larger consumer surplus. We'll have a measure of efficiency as well. And also because the demand curve here is more on the right is more elastic. This results in less expenditure. And so we're going to think of subsidy efficiency as the total  
[07:07:06] surplus generated per dollar of subsidy spending. And we'll have a few different versions of that, but it's that ratio that you can visually see here is going to be higher when the demand curve is distorted than uh when it is not.  
[07:07:19] Um so in the full model, we'll allow for a fixed cost subsidy as well as a marginal cost subsidy as well as a marginal cost tax.  
[07:07:27] And then we'll have as I mentioned just now an individual preference parameter.  
[07:07:31] How easy is it to switch? How much do you like it etc. Um we also allow for people to perceive the private benefit.  
[07:07:37] So, this is coming back to the the earlier point that people might be inattentive to the price of cooking, let's say. Um, I'm not going to get to that today. Will allow agents to make two decisions. First, do you buy the technology or not? And second, what fraction of your total, let's say,  
[07:07:52] cooking or your total driving do you do with the dirty technology? And what fraction do you do with the green technology? So, we allow uh mixing in this sense. And then we have some standard bounds that are pretty standard in the literature um to be able to  
[07:08:04] derive some of our key results. So you can start with a social planner whose objective is just to maximize total uh welfare without any distortion.  
[07:08:14] You get the standard piguvium result where you want to set your marginal price equal to the externality with the distortion. You want to introduce a fixed cost subsidy, right? And you want to keep uh the this the uh marginal  
[07:08:26] pricing uh still the same. Um this is a relatively straightforward result. We think the more interesting result is a constrained principle. So first of all, the constraint principle is not going to have the authority to tax, right? They  
[07:08:40] can just either subsidize the price or subsidize uh usage. We think there's a wide list of reasons for this, right?  
[07:08:48] Taxing charcoal in Kenya is infeasible because the sector is largely informal.  
[07:08:53] It is infeasible because politically that is infeasible. Um there are institutional constraints to this. Um uh political considerations. let's say even Canada passed a carbon tax and then  
[07:09:06] uh uh reversed that because of political considerations uh and then also a lot of financing towards uh clean energy is from the private sector right or offsets programs or UNFC or things like this where they do they do not have the  
[07:09:20] authority to tax they're only subsidizing through one of these two channels so their objective is just going to be to maximize the value that's generated by the set of adopters uh and we're We're uh characterizing adopters  
[07:09:33] by that preference parameter alpha subject to some budget constraint of the total amount that you're spending on fixed cost subsidies and marginal cost subsidies.  
[07:09:44] And we allow for two different principles. I think different people depending on if you're a public economist or an environmental economist, you'll be different interested in different things, right? A social principle uh will value the private fuel savings for the recipients and also  
[07:09:58] value the cost of making the stoves. an environmental principle. Say somebody who's trying to buy offsets is just interested in the environmental gain.  
[07:10:07] So, we're just going to derive our results for both of those depending on which of these uh you're more interested in. I'm not going to go through all the math in detail. I just want to show the intuition here that a fixed cost subsidy only affects the extensive margin, but  
[07:10:22] you still have to account for those extensive new adopters also now incurring marginal cost subsidy expenditures, right? Right? And so that's going to be the first order condition here. You also have infram marginal cost when you increase your subsidy. You're paying that also to  
[07:10:35] people who are already adopting the stove. Right? It gets more complicated for the marginal cost subsidy because there this is going to affect and this could affect both the adoption decision as well as the uh usage decision.  
[07:10:48] Thankfully we can simplify this to kind of the main key takeaways. The first is that there's going to be a direct effect of the distortion on the optimal fixed cost subsidy. And then second is that there's going to be some reoptimization where the distortion will also affect  
[07:11:02] the optimal marginal cost subsidy and that will then affect the optimal fixed cost subsidy because you have this budget constraints and so you're directly trading these off.  
[07:11:12] So I'm just going to go through the the main results and then I'll move on to the the RCT. So what we find in the model is that a demand distortion increases the subsidy efficacy of any given fixed cost subsidy. Right? and  
[07:11:25] subsidy efficacy is the welfare gain per dollar of subsidy expenditure.  
[07:11:29] If the principle can reoptimize between marginal and fixed subsidies, then whether the effect of the demand distortion is positive or negative will depend on other parameters. But we can characterize this and we can argue that  
[07:11:42] even when you allow this in many cases demand distortion will increase the subsidy efficacy even when you allow for that type of reoptimization.  
[07:11:52] All right, so that was the theory section. Let me tell you a little bit about the RCT design. So, we're working in Nakuru County, a couple of hours north uh west of Nairobi. Uh about twothirds of households here have an existing electricity connection and  
[07:12:06] about twothirds cook with charcoal or wood and that's pretty representative of Kenya uh as a whole.  
[07:12:12] Um we had some uh eligibility requirements. So people had to use a traditional charcoal stove as their primary cooking technology which which many people in the area did. Um we required spending of at least $4 per  
[07:12:24] week on charcoal. Um just as an indication this is about 5 to 10% of household income. And the this second bullet here if you are a household of at least two people cooking most of your food on charcoal you would probably have  
[07:12:38] met this requirement. And then we also require that they have a prepaid electricity meter. This is also 99% of households in this region would have that that just enables us to to make the electricity transfers.  
[07:12:49] So we're going to randomly assign fixed cost subsidies. We have some mass that is uh distributed AC at different points just for incentive compatibility, but most of the mass is going to be at a price of either $20 or $73. That's 75%  
[07:13:03] and 10% off of the main price of of 82. The payment structure is we're going to randomize people to either have to pay that price upfront or they would pay $12 as a down payment and then they would  
[07:13:15] pay the rest over a 90day loan period. And then we have a marginal cost subsidy that we pay on the electricity that is used by the induction stove. So we can we get data remotely from the induction stove that reports how many kilowatt  
[07:13:30] hours are being used. we can transfer those subsidies uh and our subsidy program in that sense we're sending tokens via SMS and that's actually very similar to how Kenya Power has implemented subsidy programs. So in that sense we're we're relatively policy uh relevant. Um and these three are are  
[07:13:44] fully cross randomized. Um I mentioned that we tried to draw a lot of attention to the marginal price uh subsidy. So this is an example of somebody who is getting a subsidy. So it tells them normally how much would it  
[07:13:57] cost to cook a certain meal on an induction stove. Let's say Ugali would cost five shillings. Rice would cost 10 shillings. And that it'll say for the next six months, which is the duration of the subsidy program. It would cost you, let's say, only three shillings to cook rice because you are in this  
[07:14:12] subsidy program. Right? The actual text here, the numbers would be adjusted depending on what treatment group you're in. We had a 15% holdout group where they didn't get any price data at all.  
[07:14:22] They just said, it just said, you know, you can cook all of these meals on the induction stove. Uh, and we had an identical flyer for the charcoal stove as well, so people could compare how much does it usually cost me, you know, how much charcoal do I need to buy to cook rice versus how much does it cost  
[07:14:37] me to in electricity costs to cook rice on an induction stove.  
[07:14:43] Okay, so we implemented this from August uh to to March. We had an enrollment survey, a main survey where we did the BDM elicitation. Thank you to Conrad so I don't have to explain BDM uh again. uh and then we immediately performed the  
[07:14:57] stove transaction for people who won the BDM, conducted an endline survey and then uh downloaded the the stove usage data uh later on. And so we have that um charcoal stove temperature measurement for that entire period. We have ongoing  
[07:15:11] highfrequency induction stove usage measurement and then the the subsidies went on for a six-month period.  
[07:15:18] Okay. So in terms of the main visit, we start by collecting everybody's Kenya Power meter number. We didn't want to have differential attrition on that by treatment group. We then informed people about what treatments they were in. Uh we then implemented the BDM mechanism  
[07:15:32] which you know all are experts on at this point. Um 98% of people who won the BDM actually purchased the stove and paid the price. And so we're we're happy with that. Um and then people who did buy the stove, they then received a demo  
[07:15:46] on how to how to boil water on the stove, how to use the the timer, the the clock, the the lock function, etc. and they were also added to a customer service WhatsApp group where if they had questions they could ask their local customer service agents.  
[07:16:00] Okay, so let me start by just presenting the impacts of adoption. So this is from the temperature sensors that were attached to the charcoal stove and we've just used a basic algorithm to convert that into how many minutes per day or how many hours per day were you cooking.  
[07:16:15] And so we see a clear uh and immediate and pretty stable drop in how much you use your charcoal stove from before to after the main visit among people who adopted the stove relative to who didn't. This is just the OS, but of course we'll we'll present the IV  
[07:16:28] treatment effect and that's about a 30inut reduction uh per day. People are still using their charcoal stoves again to cook things like rice or ugali where they might prefer the taste over the induction stove.  
[07:16:40] In terms of energy expenditures, charcoal goes down by 10%, sorry, by $10. Uh, electricity goes up by about $2. And so the total reduction in energy spending is about $9. Uh, and again,  
[07:16:53] that's about $184 over the lifetime of the stove, which is about 112% uh ROI.  
[07:17:00] And again, this is from the IV regression. Yeah.  
[07:17:02] >> Just so how frequent were power outages? >> We collected the data and I have not had a chance to look at it yet. I've been on parental leave. That's my excuse.  
[07:17:13] >> Single burner, right? >> Yeah.  
[07:17:14] >> I mean, you might want to cook two things at the same time. Like, it's possible that if you gave two burners, they would just not use a charcoal stove at all, right?  
[07:17:21] >> Yes, this is absolutely true. So, it's certainly true that you might want to.  
[07:17:24] So, another thing that we haven't looked at yet is how often do people cook, let's say, simultaneously, and your total daily cooking time now, you can cook more efficiently. So, that's something.  
[07:17:35] >> Yeah, exactly. Um and and because we have high frequency, so we deployed uh sensors that measure power outages and voltage fluctuations as well because there's a lot of that. And so we can actually see when there was an outage, did you then at that exact moment switch to your charcoal stove or did you do  
[07:17:49] that later on and we can do that type of those types of exercises. Um maybe too much for this paper, but I think there's a lot to be written about the cooking transition in general where questions like that I think are going to be important.  
[07:18:00] Yeah, great. Um okay so we can then convert these expenditures using local market prices to kilograms uh etc of actual energy usage. So this is in tons of CO2. We see a reduction of 2.5 tons  
[07:18:14] of CO2 of charcoal and an increase of 0.02 tons of CO2 of electricity. Partly this is because Kenya's grid is about 90% renewable. It's largely hydro and geothermal and a little bit of wind. But  
[07:18:28] actually even when you use the US emissions grid average it's like let's say that triples then it's 0.06 instead of 0.02 right? So actually it's just because induction stoves are so much more efficient at converting uh  
[07:18:42] electricity to heat and and power plants are so much more efficient than charcoal stoves uh that there's this huge uh difference in how much they emit and so yeah >> yes exactly  
[07:18:55] >> yeah it's renewable matter >> yes so we assume that 30% so we assume based on a very large spatial model that has done this for the whole world that 30% of the wood being used that is used in urban areas in Kenya comes from wood  
[07:19:10] that is uh that's the non-renewable fraction. So we only assume that that and we have a whole methodology section about how to use all of those but we're trying to use the the the best available here.  
[07:19:22] Okay. So we can then uh convert this to to subsidy expenditures required per ton of CO2 abated. So here we're doing all of this among the control group, right?  
[07:19:32] which is kind of representative of what Kenya would be uh what what this would be like if you were to implement this in Kenya. And so subsidies cost $65 uh in subsidy expenditure per additional adoption. Each additional adoption  
[07:19:44] induces $5, sorry, five tons of CO2 abatement per adoption. And so if you put those numbers together, that's $13 in subsidy spending for every ton of CO2 that's abated. Uh you might also be interested in the resource cost, right?  
[07:19:59] That's kind of more of an efficiency measure. Uh there's a stove cost of $82, but then the fuel savings of $200. So every ton of CO2 abaded actually saves $24 in resources. So this is kind of the McKenzie curve for those of you who remember that. This is the negative  
[07:20:13] part. You're actually saving money for every ton of CO2 that you're abating. Uh and we can also think about a welfare gain. So this is observable welfare gain. So, of course, there could be other attributes that people are liking or disliking, but at least from what's observable, when we factor this in,  
[07:20:27] every dollar of uh subsidy spending generates either $1.8 of welfare um if you only value the private benefits or this would be $2.8 if you also value value the cash transfer. Um, and if you  
[07:20:41] assume a social cost of carbon of of $120 per ton, then that gets you uh $11 of social welfare gain for every $1 of uh subsidy spending.  
[07:20:53] So, the second main impact finding is that uh credit distortions in this market, lower willingness to pay or access to a credit, access to credit almost doubles um uh willingness to pay  
[07:21:04] from $20 up to $35. We see no difference in willingness to pay depending or by the marginal treatment subsidy, the marginal cost subsidy, excuse me. Right? So the we're  
[07:21:19] separating here the credit treatment group and the credit control group by whether they got the usage subsidy and we see no differences even among uh those groups and we can rule out uh that people value a dollar of marginal cost  
[07:21:32] subsidy by more than 16 cents. Our point estimate is basically zero.  
[07:21:37] On the intensive margin, we also see that people are very inelastic. So, this is from the induction stoves themselves.  
[07:21:43] They report power usage on an ongoing basis. And we see no differences both for the 25% and the 75%. And again, we're still trying to figure out whether this is an intention story or a preferences story.  
[07:21:57] Okay. So, now we can start to think about the two channels, right? Where we have the theory about Yeah. Yeah.  
[07:22:10] It's just like if I just fix or not.  
[07:22:19] >> Yeah, it's basically a um Yeah, it's I mean it's kind of like a cash transfer. It's kind of making the stove cheaper, but people are not actually internalizing that in their willings to pay, but it is just it's it's just a marginal transfer.  
[07:22:33] That's right. >> Yeah. Uh but people don't realize it. It doesn't change their willingness to pay. So it doesn't change the adoption rate in that sense. Yeah.  
[07:22:46] >> Okay. So recall that. >> Can I just follow up on that? Like so I >> Yeah, you can people don't realize it. I I thought it was just about the credit constraint like that it was a flow over time. So even if you realized it. So, but you're saying that they must not  
[07:22:58] have realized the the NPV of those savings.  
[07:23:03] >> So, this is the um response to the electricity subsidy that we provided.  
[07:23:10] >> And sorry. >> Yeah.  
[07:23:12] >> And so, it's very possible that that response is different from how people are making decisions when they're faced with their regular. So, I don't know.  
[07:23:19] It's it's certainly true that people are um responding also to the actual differences in like how much money are they saving, right? So, they're aware of the savings in some sense because they have some willingness to pay for these these savings, but at the same time, they're not responding here. It could  
[07:23:33] just be the way that we're presenting it or um the fact that maybe they don't believe it or they're Yeah, I Yeah.  
[07:23:42] >> Yeah. Yeah, that's what that's my question which is you sort of inferred that they must know about these >> then this would differ by credit treatment status.  
[07:23:48] >> Then if we're offering credit you should see a response.  
[07:23:55] >> Then for for people where we've relaxed the credit constraint you you should see a response but we don't see that here either.  
[07:24:02] >> Yeah. But you just that many marginal people like on the the first stage on the takeup between those is not huge, right? Or the credit and nonredit.  
[07:24:10] >> That's the result. The credit and nonredit is the blue versus the gray. So those have a large difference. But within the credit group, we don't even within the credit group, even among people whose credit constraints have been relaxed, we don't see a difference in willingness to pay by whether you're in the usage, subsidy,  
[07:24:25] treatment, or control group. >> Yeah. I guess I'm just saying the credit the credit relaxation like didn't relax credit sufficiently for everyone. That's all.  
[07:24:34] >> No, that's totally possible. If you had offered more of a loan, then it might have mattered.  
[07:24:41] Okay. Um, so we're interested in understanding whether the marginal adopters that are induced by a subsidy in a world where people have larger demand distortions are different, right?  
[07:24:54] And so what's nice here is that marginal adopters, we're going to define that as people having willingness to pay between 20 and $73. And that's also going to match to the compliers in instrumental variables regression that we ran earlier.  
[07:25:08] And so here we see that people who bought the stove by paying upfront and who had a willingness to pay in that who are marginal to the subsidy use the stoves more than marginal adopters who  
[07:25:21] bought the stove with a loan. Right? So that's the first piece of evidence that they have larger uh usage or larger private benefits potentially. We can run that in a regression. Here I'm showing both the treatment effects individually  
[07:25:34] that's the even columns and then also the difference between them. That's the odd columns. So we see that people who bought the stove who are marginal to the subsidy with credit have less of a reduction than people who are more  
[07:25:48] constrained. They have a larger reduction here. Right? And that's the case for charcoal specifically as well as in terms of total abatement. And so here we're seeing on the far right column that people who are more constrained have a reduction of 2.6 tons of CO2  
[07:26:03] whereas people whose constraints were somewhat relaxed have a reduction of 2.2. 2 uh tons of CO2 and that statist that difference is statistically significant at the 10% level.  
[07:26:14] So to summarize those results we see that people who are more distorted whose demand is more distorted have a larger uh private benefit as well as a larger marginal externality.  
[07:26:26] The second channel is on subsidy expenditures to induce additional adoptions.  
[07:26:39] people who are paying up front. >> Yeah. Because they do not have access to the loan. Those people have larger private benefits and are also generating  
[07:26:51] a larger positive externality. And then in terms of subsidy expenditure, what matters is the demand elasticity, right? And we're seeing that that's different as well. that there's  
[07:27:03] uh an demand elicity of negative 2.3 among the group that has more distorted and that uh demand elasticity of negative0.9 for people who have access to a loan. Right? So the distortion here also increases demand elasticity and  
[07:27:16] that's going to lower uh the cost of inducing additional adoptions. Right?  
[07:27:21] You can see that very clearly here. you can look at for the two demand curves that we estimate. What are the uh expenditures inframarginal uh payments to inframarginal buyers, right? People that were already buying the technology infram formarginal payments to marginal  
[07:27:36] buyers. So you could have given them a lower subsidy but we're giving them all the same subsidy. And then marginal payments to marginal buyers, those are the ones that are actually inducing uh a change here.  
[07:27:47] Okay. Uh so again um the group that had a larger demand distortion those are people who are paying up front they have a larger demand elasticity this causes fewer of the subsidy expenditures to go to infram marginal participants and this  
[07:28:01] lowers the subsidy expenditure that's required to incentivize one additional marginal adoption. So this makes it cheaper to incentivize adoptions. And so if we combine those two results, the first channel is that the marginal  
[07:28:14] adopter when people are more distorted has larger benefits, they have lower costs. And we combine those and we get those final numbers that we saw earlier, right? Where they generate a larger in these more distorted markets, these subsidy dollars generate a larger  
[07:28:27] welfare gain and it costs less in subsidy expenditure to abate one ton of CO2.  
[07:28:34] Of course, we could ask what if the distortion was even lower, right? kind of perceimas point. We only reduced we only increased the gamma here to 0.19.  
[07:28:41] It could be 0.5. It could be higher. And so this is when we estimate the model to try and say something about this. And so we assume there's some distribution of private benefit with um the two these two omegas. There's some distribution of  
[07:28:54] social benefit that has some correlation with the private benefit. And we have the omega 3 here to set that correlation that could be very correlated or or not very correlated. And I'll I'll show some results on that as well. We allow for a  
[07:29:08] heterogeneous demand wedge here, right? Again, that's going to be set by omega 4 and five. And those are the five parameters that we have. And then we compute everybody's willingness to pay as their private benefit times their uh assigned demand wedge.  
[07:29:22] We then target those seven moments to estimate those five parameters. And so for six of them, we're going to match them to both the control group and the treatment group based on the the gamas that they have. All right. Um here I'm  
[07:29:36] presenting the model fit with the empirical moments. We see relatively close uh fit here. We we I can show this also for other kind of functional form assumptions here. And the fit is is relatively similar. It's not very  
[07:29:50] sensitive to those uh choices. And so once we have this model, we can then say what would happen if the average distortion was let's say 0.5, 0.75 or one. And so these are the the results  
[07:30:03] here. So uh columns two and three are what we observe in our experiment. And so we see here again that the observed and simulated um costs and and welfare gains are are similar. And if we extrapolate this up, let's say when  
[07:30:17] people are uh fully willing to pay their private benefit, so when gamma is equal to one, the welfare gain per subsidy dollar drops down to one and the subsidy cost to abate one ton of CO2 increases  
[07:30:31] by more than 10 to $137. Right? So that 10-fold increase you can think of that as uh what is the efficiency gain of spending green technology subsidies in countries where there are larger demand distortions  
[07:30:46] versus when uh people are willing to pay more of the the private benefit.  
[07:30:52] So here I'm only varying the demand distortion. We can also think about the correlation between private and social benefits. And this goes actually a little bit. Yeah.  
[07:31:04] >> Credit margin for subs. >> Yeah.  
[07:31:07] >> On the one hand, you're basically saying if I don't fix the credit margin, then the marginal people are really, you know, really directly are really awesome, right? But if I but on the other hand how do I is that what else I need to do to do the  
[07:31:22] policy calculation of like should given but of course there's elasticity people I get from from the credit distortion itself.  
[07:31:30] >> So you're thinking what if there was a policy instrument to target the credit distortion >> you're saying the crossartial I think of you know d squ subsidy depending on whether it's credit or not. Yeah.  
[07:31:42] >> Doesn't tell me the policy question I'm interested in which is like if I have a given dollar and I'm the government should I subsidize should I do a direct price subsidy or should I you know subsidize the interest rate.  
[07:31:55] >> Yeah. So you can ask the question of what if you spend subsidy dollars on interest rates. So um we yeah so we can do that calculation and when you factor in default rates even with the digital collateral technology that they're using here when you allow for default rates  
[07:32:10] that means you then need to increase your interest rate to compensate for that but when you increase your interest rate that increases your default rate and actually if you do that calculation we don't find any interest rate where you can so in terms of subsidizing the  
[07:32:22] interest rate um the you could think about like $1 of um $1 of credit subsidy actually it's you still incur the  
[07:32:34] economic cost right which are um the the cost of actually providing the credit the the moral hazard the adverse election and so actually the subsidy becomes more efficient we have I'm not explaining it very clearly we have thought about this point and yeah  
[07:32:49] >> like theoretical exercise here credit there's no feasible policy that looks like the credit because of this issue you point out of So that benefit >> exactly because the the interest rate that they were charging was 200%. Where  
[07:33:03] and so for every dollar you might as well just subs subsidize a dollar itself as opposed to Yeah.  
[07:33:10] >> Yeah. So does that mean if there was a functioning credit market, it would always be more efficient to to to fix to subsidize the interest rate? Like you kind of had a non-existence of a a  
[07:33:23] market story. And so does that mean >> Yeah. the market existed, it would always dominate.  
[07:33:28] >> I think it's an absolutely interesting question. So, these results hinge on the current state of credit markets [snorts] in low and middle- inome countries when we're just saying if you're interested in environmental policy given what capital markets look like now, these are the results. It's certainly possible  
[07:33:43] that these results would change or like there is Yeah.  
[07:33:46] >> Well, I think it's about we didn't have a world where we know that this point you make about the default rates are going to change due to average different places. Yeah. So suppose we  
[07:34:01] had a world where that was not a big those those curves are pretty small. So I could just subsidize the the >> tell us perspective of sort of you know compare  
[07:34:15] the first of subsidize the interest rate versus subsidize the the the price uh as opposed to just comparing kind of the second >> yeah I think that's right yeah I think that's right and I think we can you can  
[07:34:28] think of the interest rate here the 224% as being kind of a sufficient statistic for thinking about what is the efficacy of your policy across different so if you have another location where you have a 50% interest rate or 100% um what does that do to the opt optimal policy. I  
[07:34:42] think um that's Yeah. No, we should think about this more. That's useful.  
[07:34:49] Good. Yeah. [laughter] I think this is the beginning  
[07:35:04] of this um you know willingness to pay and fix cost  
[07:35:17] and I guess from behavioral economics and so on we know a number of places where people's willingness to pay for a good changes with the price they paid.  
[07:35:31] learning or signaling things and so on. I guess I'm curious what you thought through what this do to the results.  
[07:35:40] >> Yeah. So, we can look at people who have a willingness to pay above the high price and say among people who paid the high price versus people who paid the low price, do they use it differentially? Um I think the sample is  
[07:35:54] of people with such high willist bay is very low and so I think when we looked at that it was just largely noisy. Um but in prior work when we did this for our previous paper um we didn't see any heterogeneity and charcoal usage. I think it's just a very inelastic object  
[07:36:08] as well. Yeah. Yeah. Yeah. >> Back to the point about the sub electricity. Uh it's also possible that people are just locked in the new kind of show and it's pos  
[07:36:22] or other way like people but they want they don't have any reason to you know keep it on like which is kind of different from vehicles like you can still drive a little bit more like so it  
[07:36:35] might be interesting to see like what is like kind of you see for LPG or charcoal is different but and then see like >> yeah that makes sense. Yeah, we we have  
[07:36:48] some reductions in LPG and wood usage but we can definitely look at that.  
[07:36:51] Yeah. Okay. So um I just want to emphasize here that um throughout the paper throughout the talk today I've talked about a positive correlation between private and social benefits. You can also imagine there being a negative correlation which  
[07:37:05] generates adverse selection here. And so then this would flip the effect of the distortion where then the marginal externality that's generated would actually be larger when there is no distortion than when there is. Right?  
[07:37:18] And so this first result the first channel depends on that correlation the second channel doesn't. And actually in this setting the second channel dominates. So actually regardless of what the distortion was or sorry regardless of what the correlation is here um it would always be the case that  
[07:37:32] the larger distortion lowers uh the subsidy cost. Okay. Okay, with the final kind of minute that I have, let me just say something briefly about the policy implications here. So, if you're only interested in the cooking sector, we think this is an important result in and  
[07:37:45] of itself. Um, there's large uh changes in uh Asia of people switching away from biomass, but not so in Africa. Um, but you can still think even for for other technologies, where might we find the  
[07:37:58] lowest point on the abatement cost curve, right? We can use the model to think through some of these different parameters to think how that might uh drive um you know uh in in which types of contexts might you see that the  
[07:38:12] demand distortions lower uh the the cost of the subsidy right and we think that these results can explain partly why in many low and middle inome countries we're seeing lower subsidy costs than some of these more expensive um subsidy  
[07:38:26] costs that we're seeing from let's say the inflation reduction act and other uh subsidy expenditure All right, so I think my my time is almost up. Thanks so much for the comments and questions. I look forward to talking more.  
[07:38:45] Do it. Thank you everyone. We're we're all uh done for for this year's edition.  
[07:38:49] So, thank you to all the wonderful presenters, the audience for their engagement, the committee again for uh the work to prepare it, and I look forward to seeing you next year.  
[07:39:05] >> Oh, yeah. >> This year.