YT YouTube2Text

YouTube2Text → Transcripts → Lecture 8: Barriers to technology adoption: what we know from micro empirics

Lecture 8: Barriers to technology adoption: what we know from micro empirics — Transcript

by CEPR & VideoVox Economics · 18,342 words · 2,645 segments · language en · Watch on YouTube

Open in transcriber Summarize with AI Download SRT/VTT

Full transcript

0:00welcome back to steg's uh virtual course
0:02on key Concepts in macro development
0:05steg uh stands for structural
0:06transformation and economic growth it's
0:08a c Le research program funded as part
0:13of the UK uh Aid effort by the foreign
0:16Commonwealth and Development Office of
0:18the United
0:19Kingdom uh we're thankful for financial
0:21support for them from them um we are
0:25also thankful to have uh Lauren buris
0:28and Elanor wisman uh here today with us
0:32um if you've been following the course
0:35uh last week um David lagakos talked
0:38about the agricultural productivity Gap
0:41and um Chris tetti gave us an overview
0:44of macro models of Technology diffusion
0:47um Lauren's going
0:49to uh push toward micro empirics what do
0:52we know about technology adoption in
0:54agriculture in
0:55particular um Eleanor is uh the TA
1:00for the class and she will be giving a
1:02session next Wednesday and she'll also
1:05be uh answering questions in the Q&A um
1:09as as usual um if you're joining in the
1:12front row as a panelist uh please mute
1:15your microphone but feel free to unmute
1:18if you have a question or clarifying
1:19question especially and um Lauren will
1:23leave some time 15 minutes at the end
1:25for questions uh those of you who are
1:27joining through the webinar you can
1:29enter questions into the Q&A uh Elanor
1:32will be uh helping to answer those Doug
1:34and I may also take a crack and if
1:37there's questions that we find you know
1:40either more difficult or their broader
1:41questions uh that we'd like to bring
1:44them them up to Lauren we'll do so at
1:45the end um I want to thank uh Mandy Chan
1:50uh from CPR also kiry mcneel and Lauren
1:53wearing for helping organize the course
1:56as always and especially thank you to
1:58Lauren who's really doing this again all
2:00everyone who's teaching this is teaching
2:01it out of the goodness of their heart um
2:04so uh with that Lauren why don't you get
2:06started and share
2:08screen great um so can you all see my
2:11screen great perfect um yeah so thanks
2:14so much to Joe and Doug and Mandy for
2:16organizing this great course and thanks
2:18very much for inviting me to participate
2:20in this um it's really um really an
2:23outstanding course and and I think the
2:25the level of Interest by the number of
2:26students has been pretty amazing so
2:28great to see you all here
2:30so yeah so today I'll be talking about
2:32uh barriers to technology adoption
2:34specifically and agriculture um and what
2:36we know from micro empirics so um before
2:40we dive in let me give you a big picture
2:41on what we'll be talking about for this
2:43lecture um you may be wondering you know
2:46why are we talking about micro evidence
2:47in a class on on macro topics so I'd
2:50like to sort of motivate this um in in
2:52in two ways but first is that um
2:55agricultural technology adoption is an
2:57area in which uh especially over the
2:59past couple decades we've really
3:00generated um a very large amount of of
3:03well identified evidence coming from uh
3:06natural experiments coming from rcts um
3:08on the constraints or barriers to
3:10technology adoption so it's really um a
3:12place we've got a lot of micro evidence
3:14I think this is potentially great fodder
3:16for um for new macro models so you know
3:19bringing these constraints into macro
3:21models potentially even um building on
3:23or using these micro estimates in macro
3:25models there's a lot of material here to
3:27work with um the second is that there
3:30are you know conversely just as micro
3:32can help macro there's also I think um
3:34uh synergies in other direction right
3:36there's limits to what can be studied by
3:37micro approaches Alone um and I'll try
3:40to be clear about how micro approaches
3:41are trying to move towards macro um at
3:43towards the end of the lecture but I
3:45think there's a lot that can be brought
3:46to the table through bringing a macro
3:48lens so um I think there's a lot of
3:49synergies here and hopefully you see
3:51that come through um throughout the
3:53lecture so goals for today's class um I
3:56will review um admittedly a very small
3:59subset of what is a very extensive and
4:00large um micro literature on
4:03agricultural technology adoption um and
4:05if we have time at the end I'll try to
4:07highlight some of the ways um that I
4:09think macro and micro can can speak to
4:11each other or build off each other in
4:13this area let me give one plug for um
4:16Elanor wisman who's the TA for this
4:18section um who as as Joe mentioned is
4:21here answering your questions via chat
4:23um Elanor is going to hold a session uh
4:25on on Wednesday up next week talking
4:27about um uh you know reviewing questions
4:29from this lecture but also talking about
4:32some of the emerging evidence on
4:33intermediation in agriculture um which
4:35is sort of an adjacent field that is um
4:38a lot newer but I think also very right
4:40for sort of this micro macro interaction
4:43it's also an area in which Elanor is a
4:44real expert so I hope you all attend
4:47attend that
4:48session okay so quick motivation to why
4:51we're focused on agricultural technology
4:53adoption so this is the middle of your
4:55agriculture module I'm sure I don't have
4:57to convince you all that agriculture uh
4:59is key plays a key role um in the
5:02economy of many developing countries um
5:04but it tends to have fairly low
5:05productivity so just as an example that
5:07agriculture accounts for 64% uh of the
5:10labor force in Subs Africa but only 34%
5:14of GDP uh in Asia those figures are 40
5:17uh 43% of Labor Force but only 20% of
5:20GDP um and just as a point of comparison
5:23in the US uh agriculture accounts for
5:25about 2% of labor force and 2% of
5:28GDP and the reason we're focused on
5:30agricultural technology adoption um is
5:33that the lack of adoption of
5:35agricultural Technologies um is seemed
5:37to play a key role in shaping these
5:39patterns potentially these these low
5:40productivity
5:42numbers so uh what do I mean when I talk
5:44about Technologies well the literature
5:46uses this term quite broadly so it could
5:48apply to um uh things as broad as you
5:51know production practices or post
5:54Harvest handling but the most common
5:56thing that that people typically mean
5:57when they talk talk about agricultural
5:59tech Technologies um is these these
6:01Green Revolution Technologies so the
6:03Green Revolution was a push in the 1950s
6:05and 60s to develop um uh new new forms
6:09of of high yielding varieties of seeds
6:12um of inorganic fertilizers insecticides
6:15pesticides uh and
6:17irrigation and uh this is credited these
6:19Technologies are credited with uh
6:20increasing production worldwide
6:22especially in uh in developing countries
6:24okay so Doug has a nice piece uh
6:26recently out in the JP on on exactly
6:28what is the causal effect uh these Green
6:30Revolution Technologies on yields and it
6:32seems to be absolutely
6:34massive um and this is credited with
6:36saving over a billion people from
6:38starvation so Norman Boro who is
6:40considered the the father of the Green
6:41Revolution uh received the Nobel Peace
6:44Prize for this in 1970s it's really sort
6:46of one of the great success stories uh
6:48of the past
6:50Century but that said the Green
6:52Revolution did not transform uh
6:54agricultural production everywhere okay
6:56this is particularly notable um in the
6:58subsaharan African cont text so just to
7:01see this these these figures are showing
7:03you on the left hand side um serial
7:05yields um over the course of 1960 to n
7:08to to 2000 and on the right hand graph
7:11you see uh fertilizer use from 1960 to
7:142000 and what you see is in East Asia
7:16and the Pacific right really rapid
7:18growth um in yields and Rapid adoption
7:22of
7:23fertilizer you see Latin America and
7:25South Asia being somewhere in the middle
7:27both in terms of yields and fertilizer
7:29used
7:30um and Subs Africa really is an outlier
7:32here so very little growth uh in yields
7:35pretty stagnant over this period um and
7:37also very low adoption of of
7:40fertilizer so what's going on here so
7:42why would people not adopt adopt these
7:44new seemingly profitable Technologies so
7:46there's sort of two big buckets um of
7:49explanations and I'll break them down
7:50further in just a bit the first big
7:53bucket is that you know maybe in fact
7:54these Technologies are not all that
7:56profitable so uh yes they do seem to
7:59increase yields but they also may be
8:00increased cost right both the cost of
8:02buying the technology itself um but also
8:05in particular thinking about things like
8:07complimentary labor cost so many of
8:08these Technologies require Farmers to
8:10also put in more labor when you value
8:13that maybe these Technologies are not
8:15profitable um and moreover there's
8:17probably variation across Farmers right
8:19so for some Farmers it might be
8:21profitable and others not um tabn has a
8:24really nice piece on this that I put in
8:25your syllabus on looking at Farmer
8:27heterogeneity and how important that is
8:29to take into account here okay so that's
8:31one sort of set of explanations the
8:33other set of explanations is you know
8:35maybe these Technologies would be
8:37profitable um in a world with perfect
8:39markets but they're not adopted at least
8:41not adopted universally um because of
8:43one or more market
8:46inefficiencies so what do we mean by
8:48these marketting efficiencies well this
8:50uh categorization comes from a nice
8:52review piece by Kelsey Jack uh this
8:55piece was done uh it's white paper that
8:57was done in conjunction uh with
8:59agricultural technology adoption
9:01initiative a Thai which is an initiative
9:03out of um UC Berkeley sea UC Berkeley
9:06and Jal at MIT and they funded a lot of
9:08work in this space and so this is Jack's
9:11sort of uh taxonomy of of market
9:14inefficiencies so um one possible
9:16inefficiency is externalities so some
9:19technologies May create benefits that
9:21spill over to other people and if
9:23Farmers aren't internalizing those
9:24benefits they might under adopt of
9:27course you can also have this in the
9:28other direction of adoption but we're
9:30thinking here mostly about about under
9:33adoption uh second input and output
9:35market inefficiency so sort of most
9:37basically you can think here about um uh
9:39infrastructure things like roads um you
9:41can also think about Supply chains and
9:43other uh other Contracting problems um
9:46uh Contracting problems or or um uh
9:50Market power issues along those Supply
9:52chains things that are making it more
9:53costly for either Farmers to uh purchase
9:56inputs um or or um to sell their
10:00outputs there might also be Landmark
10:03inefficiencies so in setting where
10:05settings where land tenure rights are
10:07weak um and property rights are weak um
10:09Farmers might be hesitant to invest in
10:12their land because they fear that
10:14they'll make an investment and that
10:15Harvest might be expropriated um and
10:18they choose not to invest that's one
10:21possibility labor market inefficiencies
10:23so um many of these new technologies as
10:26I mentioned require complimentary labor
10:28um it's often also more complicated than
10:30just saying more general labor it often
10:33depends on the type of Labor skilled or
10:35unskilled um the timing of of uh of the
10:38labor can matter quite a bit for these
10:39agricultural cycles and so anything that
10:42places uh a restriction on labor
10:45Mobility uh or just raises the cost to
10:48uh to labor um May interfere with
10:51adoption of these
10:53Technologies uh credit marketing
10:54efficiency so many of these Technologies
10:56most in fact require sort of upfront
10:58costs for future benefits so imagine
11:00you're investing in fertilizer you know
11:02at the beginning of the the planting
11:04season the return is only going to come
11:06later on at Harvest Time so this Gap in
11:08time uh gives room for potentially
11:10having uh credit market and efficiencies
11:12play a role so maybe Farmers don't have
11:13the money up front to make what would
11:15otherwise be a profitable uh
11:18investment risk market inefficiency so
11:20uh many of these Technologies are both
11:22higher return but higher risk um and
11:25most have at least you know some small
11:26risk of loss and so uh it might be that
11:29if you're in a setting where there's not
11:31uh you know good insurance markets to
11:33offset this risk that farmers um who are
11:35risk adverse are not willing to adopt
11:37these new technologies and you can
11:39imagine this could be particularly true
11:41for for very poor farmers who are right
11:43on the the sort of margin of subsistence
11:45or
11:45not and then finally informational
11:48inefficiencies and this can mean
11:49something as simple as you know Farmers
11:51don't know if this technology exists um
11:54it can also mean not sort of
11:56understanding the production function
11:57more deeply so not knowing about the
11:58benefits that you're going to face in
12:00particular not knowing how to use a
12:02technology uh effectively has Encompass
12:04sort of a wide a wide range of
12:06learning so these are some of the
12:08barriers um uh that that can potentially
12:11limit technology adoption and I want to
12:13map this to some of the research that's
12:15been done uh in this space recently so
12:18we have um pretty good data I would say
12:20on the last three constraints on credit
12:23Market inefficiencies risk market
12:25inefficiencies and informational
12:27inefficiencies and I'll describe those
12:29um uh uh today as we go through um as we
12:32go through the papers that are on your
12:34syllabus we have relatively um less
12:37evidence on uh input and output
12:39marketing efficiencies so this has been
12:40growing in uh in recent years and this
12:42is part of what um Elanor is going to
12:44talk about in her in her ta session
12:46thinking about the supply chains uh for
12:48both inputs and
12:50outputs and then finally we have um very
12:53little evidence on on these three
12:55remaining uh categories of
12:56inefficiencies on externalities
12:59um land market efficiencies and labor
13:01markeet efficiencies and I want to note
13:03you know these are all places that I
13:05think macro models can be helpful but
13:06you might imagine for these three in
13:08particular these are things that are
13:09both you know harder to manipulate uh
13:11experimentally which is where a lot of
13:13the micro evidence comes from and it's
13:14also places where I think bringing in um
13:16some some uh macro approaches could be
13:18particularly
13:21useful um so this is just a quick plug
13:23for uh for a tie which I mentioned
13:25before um I've put here uh their website
13:28in case you're interested learning more
13:30um they've funded since their um uh
13:32since it was founded in 2009 they have
13:35funded uh almost 50 evaluations in 15
13:38countries they've got a 100 Affiliated
13:40researchers working on all these
13:42different constraints that I just
13:43mentioned um and so if you want to learn
13:45more about any one of these constraints
13:47and and sort of the research that's out
13:48there aai is a major fund and their
13:50website has a lot of the a lot of the
13:52work that's been done on
13:55there okay so that's on the intro let me
13:58go over an outline of what we'll uh
14:00discuss today so um we'll start by
14:03talking about a model of agricultural
14:05technology adoption um coming from a
14:07review piece by Jeremy
14:09McGruder um we'll then talk about uh
14:12those three uh buckets of inefficiencies
14:14where we have pretty good micro evidence
14:16right so credit and risk
14:18inefficiencies um as well as
14:20informational inefficiencies and these
14:22are places that I really see you know
14:23there's good evidence here for for macro
14:25to build on um and then I'll talk a
14:27little bit about where we have sort of
14:29less micro evidence um but where macro I
14:31think can be particularly helpful um and
14:34I'll talk about a piece uh one of my
14:36papers with um Marshall Burke and Ted
14:38Miguel on externalities uh and GE
14:41effects and then if there's time at the
14:43end I I'll make a few quick comments on
14:45sort of synergies between micro and
14:47macro and try to highlight a few papers
14:49um that have uh tried to work um uh in
14:52this intersection yes is there a
14:57question okay
15:00um so let me dive into uh uh this model
15:03I will say I think um maybe for the sake
15:05of leaving uh 15 minutes at the end I
15:07will um skim over this model and then
15:10Elanor I might ask you to review it in
15:12Greater detail um uh in your section but
15:15let me give you at least sort of the
15:16intuitive flavor of this model so this
15:19model is uh comes from that review piece
15:21that I mentioned um which is again a
15:23nice place to look if you're looking for
15:25um a review of what's been done in the
15:27literature so far
15:29this is an extension of the model that's
15:31in the paper on your syllabus today uh
15:33the Carlin at all all paper and so this
15:36is a model of Technology adoption under
15:38those three constraints um where we have
15:40a lot of good evidence so credit
15:41constraints risk and imperfect
15:45information so the basic setup of this
15:47model is that um some input X must be
15:50committed before the state of the world
15:51s is realized okay and S is going to
15:54determine the farmer's return from that
15:55input and urm probability pi and so uh
16:00the the the sort of intuition here is
16:02that farmers again take that fertilizer
16:03example farmers are going to have to
16:06invest in fertilizer at say planting
16:08time they're not going to get the return
16:10until later on at Harvest Time and that
16:13return is going to depend on um on the
16:15state of the world so maybe you know
16:17it's going to depend on whether it's
16:18been a particularly uh rainy or dry uh
16:21growing season and so you can already
16:24see again here how credit and risk are
16:26going to come into play so credit right
16:28you're investing in input up front
16:30you're not getting the return till later
16:31on that's where credit constraints can
16:33potentially play a role and risk because
16:35you don't know what state of the world
16:36you're going to be
16:38in we're going to add to this incomplete
16:40information so a farmer is also going to
16:42be uncertain about how uh her input
16:45choices will perform in each possible
16:47state so we're going to add an
16:49additional layer of uncertainty here so
16:52Productions now also going to depend on
16:54the technological production function
16:55that's realized so again what's the
16:57intuition well a far you know not only
16:59do they not know is it going to be a dry
17:02or rainy uh growing season they also
17:05don't know even if it is rainy um you
17:07know how profitable fertilizer going to
17:09be on my particular field than this year
17:12so there's uncertainty there coming from
17:14a lack of information as well so again I
17:17think um I'll skip over this for the
17:19sake of time to leave time for questions
17:20at the end but hopefully Elanor can go
17:22over this as a very simple model which
17:24farmers are picking uh inputs and the
17:26amount to either save or borrow as well
17:29okay so let me just go over the
17:30implications here sort of intuitively so
17:32the idea here is that credit constraints
17:34are going to reduce input usage right
17:36you might want to invest in fertilizer
17:37but not have that cash up front to do so
17:41risk is going to reduce input usage um
17:43you're going to be less willing to adopt
17:45if there's some risk and there's no way
17:46to ensure that risk um and you'll notice
17:49in this particular model informational
17:51failures are modeled much like Risk so
17:53it has a similar implication where
17:55information failures um uh might also
17:58reduce your for uh your
18:00investment this is just a a very basic
18:02sort of framework for how to think about
18:04these three constraints and how they
18:06might limit uh input usage let me now
18:09get into the meat of of uh this lecture
18:12which is really about the empirical
18:13results on these constraints so we'll
18:16talk first about uh credit and risk
18:19inefficiencies coming from this Carlin
18:21uh at all
18:22piece so they're going to focus here
18:24actually on on two constraints right so
18:26on credit constraints which we talked
18:27about many uh invest might require
18:29upfront cost um and produce gains later
18:32on in time and they're also look at risk
18:35so again the profitability of many of
18:37these Technologies depends on a variety
18:40of factors that are going to vary so the
18:41weather the macroeconomic conditions
18:43which might shape the price at which you
18:45can sell your output
18:46Etc um a few things to note the first is
18:49that these risks are are mostly
18:51aggregate risks and therefore pretty
18:52hard to insure informally so informal
18:55Insurance means imagine that Joe and I
18:57live in the same Village if my kid is
18:59sick Joe might lend me some money um
19:01when I'm facing that idiosyncratic shock
19:04I might lend Joe money when his kid gets
19:05sick you know a month or two from now
19:07and so these are sort of the the
19:08idiosyncratic shocks that we can help to
19:10informally uh ensure it's very difficult
19:13to do this sort of informal insurance
19:15for um uh for risks that are that are
19:18coming in the aggregate which is mostly
19:20the case for agriculture right so if I
19:21have a bad rainfall on my farm if Joe's
19:24my next door neighbor and he has a bad
19:26he's GNA have a bad rainfall as well
19:28we're not going to be able to insure
19:29each other and so it's very difficult to
19:31have sort of informal versions of this
19:33insurance and again as I mentioned
19:35before this is particularly hard for
19:36poor Farmers right who are right at that
19:38margin um of subsistence or
19:41not so they're going to study these two
19:43constraints in the context of Northern
19:45Ghana uh just a few quick details on
19:47this context um in their setting the
19:49median farmer uses no agricultural input
19:52no no chemical inputs so no um
19:54fertilizer here is what they mean
19:56primarily by chemical inputs and
19:59agriculture is almost exclusively
20:00rainfed and so what that means is they
20:02don't have a ton of irrigation things
20:04are really rainfed um the yields are
20:06going to be very dependent on was there
20:08enough rainfall exactly when it was
20:10necessary during the growing period And
20:13so um you know there's been sort of
20:14previous work here showing that that
20:16weather shocks translate directly to uh
20:19consumption fluctuations okay so a lot
20:20of risk in this
20:23environment so they're going to run um a
20:26pretty complex threeyear multi
20:29rrct so in year in year one they're
20:31going to run two different interventions
20:34so the first intervention is a cash
20:35Grant they're going to give Farmers cash
20:38upfront um and they're say use this
20:40however you want and the idea here is
20:42that if farmers are credit constrained
20:44and and this is the reason they're not
20:45adopting inputs you've now relaxed this
20:48credit constraint Farmers may now choose
20:49to go buy inputs with this
20:52cash the second intervention is a grant
20:54of rainfall index Insurance okay so let
20:57me just describe this this rainfall
20:59indexed Insurance are often called uh
21:00weather index insurance so the most
21:03basic insurance product you can imagine
21:05for Farmers would be um you know a basic
21:08insurance that pays out when your
21:10Harvest fails now the challenge with
21:12that model is that's going to bring up
21:13issues around moral hazard and adverse
21:15selection so moral hazard because if you
21:18exert less effort on your farm your
21:20crops are more likely to fail and you're
21:21more likely to get a payout and aders
21:24selection because the farmer may have
21:25some um asymmetric information about how
21:28risk
21:29uh their their Farm is that the insurer
21:31may not know and you're going to induce
21:33in the sort of riskiest Farmers into an
21:35insurance product so weather index
21:38insurance is a way to try to get around
21:39these moral hazard and aders election
21:41issues and the way that it works is you
21:43make payouts based not on the Harvest
21:46happening on this particular um uh
21:48Farmer's plot but based on uh the
21:51weather reading at weather stations uh
21:53in the area and the idea here is this is
21:55going to reduce moral hazard because um
21:58you know your payouts are going to just
21:59depend on do we record bad rainfall at
22:02your local uh weather station it's not
22:04going to if you exert more or less
22:06effort on your farm it's not going to
22:08affect your your
22:09payouts and it should reduce aage
22:11selection because the insurer should
22:13know should sort of be able to see the
22:14history of rainfall and should have an
22:16accurate sense of the risk of of payouts
22:19uh for each farmer it does however bring
22:21up a basis risk which is means you know
22:24it's not always the case that the um you
22:26you could have totally not your own
22:28fault you could still have bad yields on
22:30your particular particular farm and if
22:32the weather station you know isn't right
22:34next to your farm or isn't capturing
22:36exactly what's happening on your farm um
22:38the weather station might say no payout
22:40when really you need to pay out that
22:41year so there is some risk to this
22:43approach okay but this has been a really
22:45exciting new um in and of itself a new
22:47technology um but one that's meant to
22:50reduce risk and therefore encourage
22:51Farmers to invest in other Technologies
22:54so they're going to look at Grants of
22:56this rainfall index Insurance the other
22:58thing that's nice here is in the first
23:00year they're going to give this
23:01insurance as a grant and the value here
23:03is they're going to now be able to
23:05observe the effect of essentially
23:06removing risk or reducing risk um for
23:09the full population of farmers so rather
23:11than just seeing um uh you know among
23:14those who buy it um what is their
23:16reduction risk and getting some local
23:17average treatment effect here their
23:19their treatment effect in year one is
23:21really going to be sort of the the
23:22global treatment
23:25effect or the global average treatment
23:27effect okay year two they're going to um
23:30oh one more thing I should say about
23:31this is a two by two design so what that
23:33means is some farmers get cash some
23:36farmers get insurance some get both and
23:38some get nothing okay so they've g a two
23:40by two design
23:41here in year two they're going to um
23:44keep the cash grants uh as a treatment
23:46arm but they're going to now uh sell
23:48rainfall index insurance at randomized
23:50prices ranging from8 to the full actuar
23:53fair price and that's going to allow
23:55them to then estimate the demand curve
23:57uh for this insurance and sort of nicely
23:59to measure the local average treatment
24:01effect among different populations to
24:03select into um uh into this product uh
24:06at each price and then finally in year
24:09three they're going to continue this
24:10insurance pricing
24:12experiment so there's a ton of results
24:14from this paper I'm just going to show
24:15you two um the first is on that the
24:17question about uh demand for insurance
24:19in year two right so when they actually
24:21offered randomized prices so one thing
24:23to note is that in this particular
24:25context they found that at the actuar
24:27fair price which is around I think it's
24:29somewhere between six and N here um they
24:31found that 40 to 50% of farmers demand
24:34index insurance this is pretty high
24:36compared to the rest of the literature
24:38so there's been sort of a whole sub
24:39literature um uh documenting and then
24:42trying to figure out why um uh takeup
24:44seems to be much lower in most other
24:47contexts right um so there's a lot of
24:48work on on why farmers don't seem to be
24:50adopting this um seemingly useful
24:53product um again sort of a technology
24:56question adoption question in its own
24:57right um one of the hypothesis is
24:59Farmers might not either understand the
25:01product it's pretty complicated um or
25:03they might not trust they're going to
25:05actually get payouts um uh uh when um
25:08when they need it and so here uh they're
25:11able to document that experience does
25:12seem to help so they find that demand is
25:15higher among farmers who uh either
25:17themselves got a payout in year one um
25:20or who knew someone in their Network who
25:22got to pay out in year
25:24one now let's turn to uh impacts on Farm
25:27investment right the primary outcomes
25:28they care about here when you reduce
25:30risk and you reduce credit constraints
25:32do farmers actually invest more in their
25:34land so let's first look first at this
25:37first um uh this first row here on
25:39insured so they're going to instrument
25:40here for being insured with uh with that
25:43varying price and they're going to find
25:45that farmers who are exogenously uh
25:47induced to to be insured are going to uh
25:50invest more in land preparation so this
25:52is primarily um uh renting tractors for
25:55example to prepare your land they're
25:57going to actually cultivate more more uh
25:59acres of land they're going to use more
26:01chemical inputs this is um mostly mostly
26:04fertilizer um the wage effects are going
26:07to be in terms of hiring um uh paid
26:09labor and family labor they're they're
26:11positive although not statistically
26:13significant um and in that when you add
26:15this all up total cost or total
26:16investment in the farm goes up so it
26:18looks pretty clearly like in their
26:19context when you reduce risk you
26:21increase people's um investment a
26:23farmers investment in their
26:25Farm Capital grants which are which are
26:28right this this third row here so here
26:30we're seeing more mixed evidence um and
26:32and on the whole they're going to take
26:33this to be basically a null so you know
26:36a small positive um uh point estim on
26:38land preparation but not statistically
26:40significant no effect on the Acres
26:43cultivated there there is an increase in
26:45fertilizer this is the one thing that's
26:46going to come through for the the cash
26:48Grant folks um but sort of mixed
26:50evidence to no effect on on labor and
26:52when you add this all up they're going
26:53to find basically a null and total cost
26:56so their takeaway here is that in their
26:57setting it doesn't look like credit
26:59constraints are the major thing right so
27:00you give people money they don't seem to
27:02be investing a lot more in their farm
27:04and also note the fact that the
27:06insurance does produce a bigger
27:08investment you know for the insurance
27:09product they're not giving money upfront
27:12and people are still able to find a way
27:13to invest in their Farm they're still
27:15able to find the cash needed to invest
27:18so their takeaway here is risk is really
27:19important maybe credit a little less so
27:22with the exception here of of fertilizer
27:24and then let me just say for the
27:25interaction term uh it's it's
27:27significant for fertilizer and sort of
27:29positive but not significant for a lot
27:31of other
27:32outcomes I will also note one other
27:34thing which is if you turn to this last
27:35column here value of harvest right so
27:37they've increased the investment or the
27:39sort of cost put into the farm does it
27:41actually produce higher returns for
27:43farmers and do those returns exceed the
27:45cost um and they actually find no right
27:48so they find that there's not a
27:49statistically significant increase in
27:51the value of harvest um and it's
27:53certainly not bigger than the total cost
27:56now one hand this might make you think
27:57sort of back to that first bucket of
27:58explanations maybe these these
28:00Investments are just not profitable um
28:02but I do want to add sort of two notes
28:04of caution here about some of these
28:05these interpretations which the authors
28:07themselves note in the paper so the
28:09first is that harvests are traditionally
28:10very difficult to measure and you'll
28:12find actually a huge number of papers in
28:14this overall uh literature even when
28:17they find impacts on Farm investment um
28:20uh find noisy effects on on Harvest and
28:23so harvests are very difficult to
28:25measure um in sort of the real world
28:27Farm farmers are often asked to recall
28:29how much they harvest which is a very
28:31difficult thing to remember the exact
28:32number there's been advances made in
28:34terms of things like satellite remote
28:36sensing uh that I think are potentially
28:38really exciting in this space and the
28:40second thing is um labor is just very
28:43very difficult to um uh to measure here
28:46so you know how do you think about
28:47valuing the opportunity cost of family
28:49labor um this is sort of a challenge
28:51that you'll see throughout uh throughout
28:53this
28:55literature okay so um this is just one
28:58example of um again a really large
29:00literature on credit and risk I want to
29:02just quickly um summarize the takeaways
29:04from a few other papers um just give you
29:06a sense of the flavor of the findings
29:08here overall so there's been lots of
29:10other papers that have tried to give um
29:12uh credit rather than cash grants
29:14actually give loans um what they tend to
29:17find is that takeup is pretty far from
29:19Universal so in this sort of you know a
29:21third or less or fewer uh Farmers take
29:23up um fertilizer use does generally go
29:27up with credit access so it seems like
29:28for some folks getting access to credit
29:30does encourage them to adopt more more
29:33agricultural uh Technologies more
29:35fertilizer but the increase in input
29:37usage accounts for only a small fraction
29:39of the total loan size so a lot of the
29:41loans are going to something else um it
29:43doesn't seem like this is the biggest
29:44thing that farmers could possibly invest
29:46in with the highest return um it doesn't
29:48seem like if you gave loans to everyone
29:50that everyone's going to use the full
29:52loans to uh invest in agricultural
29:54technology so one I think um uh balance
29:57perspective perspective on this is it
29:58seems like there's a substantial
30:00minority of farmers um who would consume
30:03more uh inputs of credit constraints
30:04were
30:05lifted on the risk side um pretty
30:08clearly it seems like across the board
30:10that reducing risk does seem uh to lead
30:11Farmers to invest more um in inputs and
30:14in highrisk high return Technologies um
30:16but as I mentioned with the exception of
30:18this paper um most other papers tend to
30:20find pretty low adoption um again
30:23potentially because there's uh there's
30:25trust issues there's uh understand
30:28issues there's a paper right a nice
30:30paper here on the fact that you have to
30:31pay the premium up front um uh to get
30:34the return from your insurance later on
30:36there's time issues as well one thing
30:38that's particularly promising here is
30:40there's been some papers for example
30:41like this em paper here that has looked
30:43at rather than trying to get people to
30:46adopt a second technology of weather
30:48index Insurance to reduce risk to adopt
30:50you know the First Technology can we
30:52just embed lower risk in the first
30:54technology so concretely what they're
30:56doing here is they're looking at um
30:58flood resistant rice variety so can we
31:01actually change the seeds themselves to
31:03be less risky and that seems to be very
31:05very
31:07promising uh questions on that before I
31:09go
31:10on ask a question about the value of
31:13harvest so you said you said it's hard
31:15to measure and I wonder whether it's
31:16measured in terms of quantity or in
31:18terms of market price because if you in
31:20ter if you measur in terms of market
31:21price then the general General
31:23equilibrium effect will come in because
31:24if there's higher Harvest then the price
31:26May decrease because of that
31:28yeah absolutely and we'll talk about um
31:31uh uh one of the papers of mine that
31:32actually looked at this exact issue
31:34later on so here I they're looking at
31:36the values this does include the price
31:38um the the the price tends to be uh
31:41easier to measure so it's easier to run
31:43Market surveys or just ask farmers what
31:45price did you um uh did you get uh in
31:48terms of the measurement issues I was
31:49primarily referring to the quantity here
31:51of getting the actual um quantity that's
31:53Harvest and that's where satellite
31:55imaging can help um the price my
31:57takeaway the price is a little easier to
31:59measure especially because you can do
32:00your own uh if you're going collecting
32:02your own data you can collect that
32:03yourself um directly but uh Harvest
32:06often requireing farmer require farmer
32:08recall and that's where I think a lot of
32:09the problems come
32:11up thank you yeah great question okay so
32:15let's turn now to uh informational
32:17inefficiencies so I want to just very
32:19briefly discuss um sort of a very basic
32:22learning model that I think um is common
32:24to many of these papers that are
32:25thinking about uh about learning uh
32:28about the sort of returns to
32:29Technologies so again the constraint
32:31here is maybe Farmers don't know these
32:33Technologies are profitable and that's
32:35why they haven't adopted so the um the
32:37very basic learning model here you know
32:39think about an agent who's trying to
32:41learn about the return to a certain
32:43technology um and we're going to call uh
32:45Theta star the true
32:47return now this agent this farmer is
32:49going to have some prior uh Theta not
32:52which is centered around the True Value
32:54uh but has some variance and we're going
32:56to Index this variance by or
32:58parameterize its variance by one over
33:00over H knot so the higher H knot is the
33:02more precise that or the stronger that
33:04that prior
33:05is in each period she's going to learn
33:08so she's going to get a signal uh which
33:10we'll call A A feda Tilda and again
33:13that's going to be centered around the
33:14True Value um uh with a variance one
33:17over hu so the higher hu is the more
33:20precise that signal is um so both the
33:23prior and the signal are centered at the
33:25right value the relative Precision is a
33:27H uh hu over HH not um and the agent's
33:31going to be updating her beliefs using
33:33basian updating so let's just see sort
33:35of what's going to happen so uh we can
33:37show that the posterior belief right so
33:39your new belief after you've received
33:41the signal uh uh Theta Theta Tilda is
33:44going to be um centered around the the
33:46weighted average of your signal and uh
33:51and your and your prior um and the
33:53weights are going to be how precise were
33:55your prior and your signal
33:58now after te periods you can iterate and
34:00show that essentially beliefs are going
34:01to converge to the True Value as te
34:04grows so people will eventually learn
34:06the sort of true productivity um uh of
34:09of um of this technology but it's going
34:12to take time right um and learning is
34:14going to be faster depending on um uh
34:17for example the Precision of your of
34:18your signals I will just note this is
34:21one particular example there are other
34:22examples in which even over time the
34:24beliefs don't converge to the true
34:26belief so if you're interested in that
34:27you can look for example at this uh
34:29paper by abaj bannery on on herd
34:31Behavior but let's take the case where
34:32we're going assume that farmers will
34:34eventually overtime figure out whether
34:37these Technologies are profitable but we
34:39want to think about um what might
34:41inhibit or speed up um uh the the speed
34:44of that
34:45learning so talk about two papers the
34:48first is this bold at all paper so this
34:50paper is set in Uganda and one of the
34:53first uh stylized facts that they
34:54document in this paper is it seems like
34:56in Uganda there's substantial variation
34:58and input quality so they run a pretty
35:01neat um mystery shopper experiment where
35:04they have their enumerators pretend to
35:06go be actual Shoppers at at retailers um
35:09and they go out and they buy actual
35:11fertilizer and uh hybrid seeds from
35:13retailers and then bold all take that
35:15those those samples to a lab and test
35:18them and so what they found was that 30%
35:21of the nutrients um were missing in
35:23their samples of fertilizer and less
35:25than 50% of their hybrid seeds were
35:27authentic so really high rates of um
35:30either intentional or unintentional um
35:32adulteration or just um you know
35:34deprecation of the um or depreciation of
35:36the the quality of these seeds over time
35:40I should note by the way this um
35:41particular finding about the really high
35:43rates of inauthentic um inputs is is
35:46higher than or more problematic in this
35:48setting than there than they found in
35:49other settings there's been um attempts
35:52to look at this uh in Tanzania and some
35:54other locations and they haven't found
35:56as high rates of adult filteration but
35:58in this setting it seems to be to be
35:59quite important so really high variation
36:02in in input quality um but interestingly
36:05not much variation in price okay so it
36:08looks like um you know this is not the
36:10case of having some differentiated
36:12product where the high quality is
36:14selling for a high price and the low
36:15quality for a low price really across
36:18the Quality Distribution there's just
36:19sort of a single price for these um uh
36:21these inputs and so this suggests that
36:23consumers maybe can't fully uh observe
36:26the quality of what they're what they're
36:28purchasing right they're not able to um
36:30uh demand a a lower price for lower
36:32quality for
36:34example the second step in this paper is
36:36that buold that all um run agronomic
36:39trials um they actually have plots of
36:41land and they apply to it uh the
36:44authentic uh seeds and authentic uh
36:46fertilizer and they have other plots
36:48where they use the sort of Market level
36:50um you know uh partially inauthentic
36:52seeds and fertilizer and they confirm uh
36:55that the returns to authentic input
36:57surprise are much higher but they
36:59actually have quantitative estimates of
37:00how much higher the authentic returns
37:02are uh I should note even the market
37:05rates while you know there's still
37:06positive returns uh to using these
37:08inputs they find um but they're much
37:10higher for authentic inputs and so
37:12you'll want to ask you know how can
37:13Farmers um uh start learning about which
37:17inputs are authentic right and then they
37:19can adopt those inputs and that might be
37:21really profitable for them so what are
37:23the challenges here to Farmers um uh
37:26distinguishing between authentic and
37:27inauthentic
37:29inputs they note here Farmers an infant
37:32problem that the farmers are facing and
37:34I want to note this is true for um uh
37:36for separating authentic and inauthentic
37:38inputs it's also true in general it's an
37:40infants problem for Farmers learning
37:42about the profitability of some
37:44technology and the infants problem is
37:46that yields are going to vary due to a
37:47number of factors uh other than just the
37:50quality of the inputs so they might find
37:53um uh that at the end of the Harvest
37:55that their Harvest failed and they don't
37:57know is that because my inputs were
37:58inauthentic or because you know the
38:00weather was bad or the um for some
38:04reason my soil wasn't as good this year
38:05a lot of other things are changing
38:07things that are either unobservable or
38:10observable but the farmer doesn't know
38:11the exact production function and so
38:13it's hard for the farmer to attribute
38:15you know was this happening because my
38:17inputs were not authentic um or because
38:19of some other
38:20Factor so they develop a really nice um
38:23uh uh basian learning model which
38:26they're going to calibrate us
38:28um using moments from their experiments
38:29and their data collection which is
38:30particularly nice so this is the result
38:33of their model let me um there's a lot
38:35going on this figure so let me break
38:37this down so on the y- axis here they're
38:39showing the log odds ratio in favor of
38:42the profit of the fertilizer being
38:45unprofitable so a log odds ratio of zero
38:48means uh you're sort of on the fence
38:50you're 50/50 about whether this
38:52fertilizer that you're using is
38:53profitable and higher than that means uh
38:56your best guess is the fertilizer is
38:58unprofitable for you and lower than that
39:00means you think the fertilizer probably
39:02is profitable okay so zero is sort of
39:04that 50/50 point they're showing this as
39:07a function of the dilution rate of
39:09fertilizer so a dilution rate of one
39:12means basically it's a totally fake
39:13fertilizer nothing's real and an
39:16dilution rate of zero means there's no
39:18dilution it's totally authentic so this
39:20is sort of good quality to bad
39:22quality and they find in their agronomic
39:24trials that as long as the dilution rate
39:26is less than
39:28to I think it's around uh 35% I may 30
39:3238% um as long as the dilution rate is
39:35less than that then uh they find in
39:37their trials this is all profitable
39:40fertilizer they want to ask what are
39:42farmers belief about whether or not it's
39:44it's profitable depending on the
39:46dilution rate and depending on how many
39:48times they've tried to use this
39:49fertilizer so how fast is the learning
39:52so what these lines are showing you this
39:53this dark black line is showing you um
39:55uh Farmers posterior belief after one
39:58Harvest round so after they've had one
40:00experience with the
40:01fertilizer this uh this long dash line
40:04is showing you after five Harvest so
40:06five rounds of experience with
40:08fertilizer um and after 10 Harvest so
40:10how quickly are they learning so a
40:12couple takeaways here so note that from
40:14a dilution rate of 0.5 to one so the
40:17really terrible fertilizer that's been
40:18really diluted there's strong learning
40:21here right so even after one round of
40:23trying one round of harvest um farmers
40:26are still above zero in terms of their
40:28their uh posterior log odds ratio
40:30they're pretty sure it's unprofitable so
40:32they can detect really bad fertilizer
40:34pretty quickly this is why um you might
40:37imagine firms have a strong incentive to
40:38not totally dilute right farmers can
40:40quickly pick that up here in moderate
40:43learning though oh here in um uh in the
40:45zero to um uh uh you know 1.8 or so
40:49right this is called moderate learning
40:51so this is the really good stuff right
40:52this is the really good fertilizer
40:54that's been barely if at all diluted and
40:57still you see here only moderate
40:58learning so it does take a while for the
41:01really good stuff for Farmers to figure
41:02out that what they're working with is
41:04the really good stuff right so even
41:05after one Harvest their best guess is
41:07that still is probably not profitable
41:09and it takes somewhere between one and
41:11five uh five rounds of trying to figure
41:14out that actually this is profitable for
41:15them to get below
41:17zero now what's particularly concerning
41:19is this area here of weak learning right
41:21this is right around that switching
41:22point of of it actually being profitable
41:25or not um and you see it takes a really
41:27long time for Farmers to figure that out
41:29so everything sort of in this range over
41:31here below I think it was
41:3338% is actually profitable um in their
41:36studies but it takes a really long time
41:38for Farmers to figure that out so even
41:40after 10 harvests this dotted line here
41:42we're still above zero right Farmers
41:43best guess is it's still unprofitable to
41:46use so there's really slow learning here
41:48um when the actual quality of fertilizer
41:50is not so far off from that Tipping
41:52Point of being profitable and so you'll
41:55want to ask you know for the actual
41:56fertilizer are available in the market
41:59where does it lie in terms of these
42:00these different buckets and most is here
42:03in this weak learning bucket okay so 60%
42:05of the samples um are here in this
42:07learning uh in this weak learning range
42:09so this is sort of an explanation behind
42:11why it's difficult for Farmers to
42:13observe the quality and to learn um uh
42:16about the quality of what they're what
42:18they're growing with right and that
42:19might be one of the reasons that they're
42:21they're they're very slow to uh to
42:24adopt learning on the last
42:28on the one prior to this so is it the
42:31case that these three lines do not
42:33intersect at the kind of switching point
42:36between profitability and
42:37unprofitability
42:38or or because that's what seems
42:40intuitive to me at least but is it the
42:43case that they actually intersect lower
42:44than that yeah they're intersected lower
42:46than that so the the cut off here is uh
42:48is 338 so uh one thing that's important
42:52for this model is thinking about
42:53people's priors so people you know and
42:55you can make different assumptions about
42:57people's prior so they're going to
42:58assume that people's priors are evenly
43:00distributed um zero to to one uniformly
43:03distributed Z to one and so uh but the
43:07actual level that's profitable is 38 so
43:09that's why for example um even after one
43:12trial even after one Harvest Time um uh
43:15you know people think it's probably not
43:16profitable it's because their prior is
43:18centered around you know 05 for example
43:21and the true the true value is 38 you
43:23could think about having different
43:24distributions there so I don't think
43:25there's any mathematical reason about
43:27why these should cross at certain at
43:28certain rates here um because the I
43:31wonder what happened if you had the
43:32distribution actually be centered at the
43:34the True Value um or so the true
43:37profitable value um uh but yeah there's
43:39no reason why these things have to cross
43:40at any particular
43:43Point other questions on
43:48this okay so this whole model though is
43:50talking about sort of how long can take
43:52you to learn so you know even that 10
43:54Harvest you still might not know um that
43:56can be a very very long time if you're
43:58thinking about only learning from your
43:59own experience but there's a thought
44:01that potentially we could speed up this
44:03learning process for this case and for
44:05other Technologies if you can learn from
44:07your peers so this is really this idea
44:09of peer learning potentially speeding up
44:11the learning process um has been a big
44:13motivating factor behind some of the
44:15models that are common to see um uh in
44:18extension services worldwide so
44:19extension services or extension agents
44:22uh are people who are hired by um uh the
44:25government to go out to community and
44:27try to disseminate or share information
44:29about say new technologies or planting
44:32processes and one of the models that
44:34many um especially developing country uh
44:37governments employ is that of lead or
44:39seed Farmers they call them and the idea
44:42here is extension agents will go out and
44:44they'll pick a prominent person in the
44:45community and they will uh dub them to
44:48be the the lead or the seed farmer and
44:50they will give that seed farmer the
44:52technology they will set up a demo plot
44:54with that seed Farmer and the goal is
44:56that seed seed farmer once trained will
44:58then go out and disseminate that
45:00knowledge to their peers and to try to
45:02Kickstart this sort of peer learning
45:04process and this paper by be and all is
45:07going to ask you know how do we pick
45:08those seed farmers and is there a way to
45:11uh use targeting or in their case
45:13Network Theory um to more optimally pick
45:16these seed farmers and speed up peer
45:17learning so if we're thinking about peer
45:19learning how do we sort of optimally
45:20seed this
45:22network so they're going to take
45:23seriously um two different models of
45:26contagion or this case sort of you know
45:27technology
45:28diffusion so the first model they're
45:30going to consider is one of simple
45:32contagion and in this model an agent's
45:34going to adopt as long as one of uh her
45:37peers adopts so uh you can think I think
45:39of this as being the case where maybe
45:41you don't know a technology exists and
45:43all you need to do is hear it once from
45:44one person and then you'll think oh
45:46great idea let me go do this complex
45:49contagion is when an agent's going to
45:51adopt as long as Lambda of their peers
45:53adopt and typically Lambda is greater
45:55than or equal to two right so this is
45:57the case where you have to hear from you
45:58have to see multi multiple of your peers
46:00adopt before you're willing to and I
46:02think that this is the case more where
46:04you know Farmers might not know well how
46:06profitable are what are the returns to
46:08this new technology I'd like to see
46:10multiple my peers give it a go before
46:12I'm willing to actually dive
46:15in so um what are the implications of
46:18these two of these two uh models of
46:20contas well imagine that you are um
46:23running an extension service and you're
46:26thinking about training up two seed
46:27farmers in every Network let's say every
46:30village who you're going to pick the
46:32optimal entry points are going to depend
46:34on which contagion model is accurately
46:36describing um how diffusion happens so
46:39in a simple model the simple contagion
46:41model if you had two seeds you put the
46:44first seed in the dense part of the
46:45network right you want them to reach
46:46sort of the most people but the second
46:48seed you wouldn't bother putting that
46:50second seed right on top of the first
46:52right there those people who are
46:53connected to that first seed are already
46:55going to adopt you want to put that
46:57second seed further out in the network
46:59to get to other people who are in the
47:00more distant periphery of the
47:03network however if complex contagion is
47:05the right model here then you want to
47:07choose Bose seeds an essential part of
47:09the network right because those people
47:10are going to really need to see um at
47:12least two people adopt in order to be
47:14willing to adopt themselves so this is
47:16going to change who you would pick
47:17optimally in terms of where how to
47:19seed so they're going to apply these two
47:22different models of contagion to um a
47:24context in Malawi so looking at the
47:26adoption of a pit planting technology so
47:29this is basically it's a very simple
47:30technology it's just changes how you
47:32plant you basically create a little
47:34divot or hole and you put the seed into
47:36it and that can capture rainfall and
47:37other nutrients better than just
47:39scattering the seeds on the ground this
47:41is a very common um planting uh approach
47:44in West Africa but it was new to East
47:45Africa it hadn't been used there before
47:47so look at the spread of this new
47:49technology one really nice feature of
47:51this paper is they're going to collect
47:53complete network data in 200 Villages
47:56now this is incredibly expensive and
47:58cost they're going to go to each
47:59individual person and they're going to
48:01ask you know who do you talk to about
48:03agricultural decision making and they're
48:05going to get the full network data then
48:07because they have this full network data
48:09they can then simulate various models of
48:11well who in this network would you pick
48:13under simple contagion who would you
48:15pick under complex for example and then
48:17they're going to randomize those
48:18Villages into actually picking based on
48:21different models right so some some
48:22Villages are going to get uh their seed
48:24farmers who are selected via simple
48:26contagion some via complex
48:29contagion they're also the third arm I
48:31think is particularly nice which is kind
48:33of acknowledging that um in reality
48:35collecting complete network data while
48:37very useful for an academic study is not
48:39feasible in terms of at scale
48:41implementation so we don't expect the
48:43government of Malawi to go out and do
48:44full Network surveys in all Villages um
48:47and so what might they do as sort of a
48:49rough proxy they might look at
48:50geographically where are people located
48:52and assume that folks who are located in
48:54kind of dense parts of the village
48:56geographically are in dense parts of the
48:58network okay so the third thing they're
49:00going to do is basically pick via
49:02complex contagion but using the
49:04geographic Network um as a as a proxy
49:07for uh the the actual underlying
49:10Network and then they're going to
49:11compare all these to a control which is
49:13going to be selection made via status
49:15quo in which the extension officers just
49:17pick who they think are the best seed
49:20Farmers so um just to show you this
49:22really changes who they pick as seed
49:24Farmers so these are just uh three
49:26examples let me show you just one here
49:28in Village uh number 288 so you see here
49:31for the simple um simple contagion mod
49:34module which is um our model which is uh
49:36here in green they've picked the first
49:38seed again right in the middle of the
49:40network the next optimal second seed is
49:42going to be out here right these folks
49:44are all getting hit by this first guy
49:46and so you want to try to reach some of
49:47these folks who are out in the periphery
49:49with this guy here in
49:50green however if you want to apply a
49:52complex model then you're much better
49:54off picking both folks in the center of
49:56the network work um because no one's
49:58going to adopt unless they see uh two or
50:00more people adopting and there's some
50:02overlap but it's nice because they can
50:03actually say counterfactually who would
50:05have been picked in every location
50:07because they had the full network
50:09data a bunch of results again I'm just
50:11going to show you twoo um the first is
50:13looking at an individual level on uh
50:16diffusion within the
50:18village and so uh what they're going to
50:20find is that um uh both in terms of
50:24whether you've heard of pit planting
50:26whether you know how to pit plant and
50:27whether you actually adopt pit planting
50:29being connected to one seed doesn't seem
50:32to change um uh your knowledge of or
50:35your adoption of pit planting it's
50:37really being connected to two seeds that
50:38makes the big difference here okay so
50:40you're really seeing increases in both
50:41knowledge and ultimately adoption when
50:44you're connected to two
50:45seeds um so it looks like this is
50:47evidence that that that um that complex
50:50contagion model is probably the right
50:52one to describe how information is
50:53diffusing in this network one thing just
50:55to note um um uh a really important
50:58thing to do here and they they note this
50:59in their in their footnote three is
51:01whenever you're looking at sort of um
51:03connections to say treated individuals
51:05you are you want to control for um
51:07underlying Network right so you want to
51:09say you don't want to compare people who
51:10are in the center of um uh of of of an
51:14area to people who are in the periphery
51:16they're going to have different
51:16connections to seeds but also be
51:17different in a bunch of ways right so
51:19they're going to control for all the
51:20underlying Network characteristics and
51:22then say based on the randomization did
51:24you randomly happen to be connected to a
51:26seed controlling for kind of all of your
51:28connections more generally so they're
51:30using just that treatment induced
51:31exogenous
51:33variation okay so it looks like complex
51:35contagion is the right way to um uh or
51:38sort of the the right model here and so
51:40is it the case that in villages now
51:41we're turning to Village level
51:43regressions in villages where they ran
51:45the complex contagion do you see more
51:47adoption than in villages where they
51:49picked seed Farmers based on simple
51:51contagion and it seems like that's the
51:53case so it looks like especially on the
51:55extensive margin of do you have any um
51:59non-seed adopters here it looks like um
52:01the complex contagion model really does
52:03a lot better than the status quo okay so
52:05it's a lot better at getting at least
52:07some people in in the village to adopt
52:10um simple seems to do better than a
52:11status quo of having uh extension agents
52:13pick uh but not quite as well as
52:16complex and then finally Geographic um
52:19is sort of similar where um uh it's not
52:21a perfect substitute for having full
52:24network data which is I think
52:25disappointing when you think about how
52:26this would be scaled right it's not as
52:28good to just say let's proxy for the
52:30network based on people's location but
52:32it does seem to do a little bit better
52:34than uh the status quo of having an
52:36extension agent pick so this is all to
52:39say you know if we can take seriously
52:40some of these Network theories can we do
52:42a better job at um speeding up the
52:45knowledge diffusion process right by how
52:47we pick uh pick these seed these seed
52:50Farmers okay so there's a lot uh more on
52:53on info constraints I think um uh for
52:56the sake of time I'll skip over this but
52:58just to say you know there's a lot of
52:59work out there um uh and I really
53:01encourage you all to uh to take a look
53:03at this okay let me now uh in the
53:06remaining time um talk about uh this
53:08last paper on um externalities and GE
53:11effects um and this is a place where we
53:13have uh relatively less um uh micro
53:16evidence but it's a place where I really
53:17think that macro can can help so this is
53:20a paper of of mine joint with um
53:22Marshall Burke and Ted Miguel so the
53:24context of our paper is looking at um uh
53:27this trend that's common in East Africa
53:30which I'm showing to you here but really
53:31across the developing World which that
53:33there tend to be um uh fairly large and
53:36fairly regular uh fluctuations in the
53:38price for staple Commodities over the
53:41course of the season so this is showing
53:43you uh major markets from East Africa
53:46and what you see here is the price over
53:48the course of the season with this blue
53:50box showing you uh the timing of uh of
53:52harvest in this case we're looking at
53:54maze so what you see here is that really
53:56sort of across the board what's
53:58happening is right after harvest um
54:00prices are are dropping pretty
54:02precipitously um only to uh increase and
54:05Rise as you go into uh what's often
54:07called the lean season when prices are
54:09high again and they fall the following
54:11Harvest so there's a cycle of falling
54:13prices right after harvest and growing
54:15over time so in major markets these
54:17price fluctuations are pretty large
54:19they're about 25 to 40% in major markets
54:23um they tend to be much larger in um in
54:26more rural areas so for example on our
54:28study site which is here in in Western
54:31Kenya um the price fluctuations over the
54:33time period that we were doing the study
54:35which is about four years ranged from uh
54:3840% to 125% so really large fluctuations
54:42in price um over the seasons this is a
54:45point of comparison in US Corn markets
54:48right the price fluctuation is more like
54:495% so much larger fluctuations uh in in
54:53developing
54:54countries okay so the establish that
54:56staple food prices um are not fixed
54:59within the
55:00season um now one important thing to
55:03know is these staple foods are storable
55:05right so storable should allow um uh the
55:08movement of grain intertemporally and
55:11sort in the simplest model an
55:12unconstrained intertemporal profit
55:14maximizer should store a unit of grain
55:17um if the discounted expected future
55:19price is greater than uh the price today
55:22plus the cost of storage so you might
55:24think that uh farming household should
55:26be using storage to uh to buy at a low
55:29price and sell at a high
55:31price instead it looks like they're
55:33doing the exact opposite um so this is
55:36uh this is data from from our pilot um
55:38and what you see is that farming
55:39households appear to be either selling
55:41low uh at Harvest time or buying back at
55:45a high price later in the season and
55:47often they're doing both so many farming
55:49households will essentially sell right
55:50after harvest time at a low price
55:53they'll run out of uh their inventories
55:55and then return to the market to buy
55:57back at a high price later on in the
56:00season and so again this data here is
56:02showing you that uh inventories um in
56:04the post in the weeks post Harvest are
56:06rapidly depleted so this is the mean and
56:08median farmer um and so what you see
56:11here is the median farmer um runs out of
56:13their own inventories by about 20 weeks
56:15post
56:16Harvest we can also look at the
56:18quantities sold minus uh quantities
56:20purchase and you see that farmers are
56:22selling um primarily selling in this
56:24post Harvest period and actually buying
56:25back uh later on in the lean season um
56:28so they're selling when prices are low
56:29and they're buying back when prices are
56:31high this really matters when it comes
56:33to welfare so the the modal household in
56:35our sample um is giving up about one to
56:38two months of agricultural wages by
56:39selling low and buying High instead of
56:41the
56:42reverse okay so what's going on here why
56:44are farmers not using storage to
56:46Arbitrage these price
56:48fluctuations so we did a lot of focus
56:50group uh work with these Farmers prior
56:52to uh to running this intervention um
56:54and the most common answer by far across
56:57the board that we got from Farmers is
56:58that their credit constraint so as
57:00Farmers say you know they know they
57:02could make more money by waiting uh till
57:04later in the season to sell but they
57:06have a lot of bills that come due at
57:08Harvest Time so in particular school
57:10fees are due right after harvest and the
57:12issue here is they don't have access to
57:14other sources of Finance to pay for
57:16example those school fees and that's
57:18going to then tie together in time the
57:20timing of of expenditure needs and sales
57:23right so these high Harvest Time
57:24expenditure needs have to be funded by
57:26High Harvest Time sales if there aren't
57:28financial markets that allow them to
57:29move across time um uh move money across
57:33time so that's the hypothesis coming
57:35from these focus groups we're going to
57:37test this by offering um a Harvest Time
57:39loan to small holder Farmers worth about
57:42$100 now what's particularly key here is
57:44we're going to randomize this at two
57:46different levels so we are going to do
57:48the basic thing which is randomizing the
57:49loan offer across farmers and that's
57:52going to allow us to look at outcomes on
57:54things like um inventories and net
57:56revenues and
57:58consumption we're also going to
57:59randomize the density of treatment okay
58:01so in some areas we're going to randomly
58:03treat more Farmers than in other areas
58:06and that's going to allow us to look at
58:07outcomes on things like Market level
58:09prices I'll describe that experimental
58:11design in a little more detail in just
58:13in just one
58:14slide we'll have three rounds of data so
58:16we collect really high frequency data
58:18this is a context in which the timing of
58:19things really matters so we'll have
58:21three rounds of household surveys
58:22throughout the year and we collect on a
58:24monthly basis price data at 52 uh local
58:28markets we ran this intervention for two
58:30years so we ran it for one year um it
58:33seemed promising we went back and
58:34randomized and ran it again to get a
58:36replication uh analysis embedded here
58:39and then we also ran a longr run
58:40follow-up survey one to two years after
58:42the loan
58:44ended so a little more in this
58:45experimental design so this is what's
58:47known as a randomized saturation design
58:49and there's a couple ways to do it but
58:50here's how we did it so we um uh took
58:53our our entire study area and looked at
58:56uh the sublocation level which is an
58:58administrative District in Kenya um that
59:00accounts for you know maybe five or so
59:02Villages and we randomized those
59:04sublocations into high intensity and low
59:07intensity and then we in the
59:09high-intensity sublocations we um
59:11sampled 80% of farmers and in low
59:13intensity just
59:1540% then within those uh within those
59:18groups we then randomized into treatment
59:20and control and so as a result of this
59:22design we're going to have double the
59:24number of treated farmers in the high
59:26intensity as we have in the low so this
59:29is becoming a new and growing sort of
59:31approach within uh the micro literature
59:33to think about estimating um uh ge
59:36effects experimentally so we're going to
59:37be able to ask what happens when more
59:39people are treated what happens to for
59:40example market
59:43prices so let me first before I show you
59:45market prices let me first show you uh
59:47what happens to uh at an individual
59:49level just comparing basic treatment
59:51control uh
59:52Farmers so this is showing you um across
59:55the across the season so November
59:57December is sort of that immediate post
59:59Harvest period June July August is the
1:00:01lean season when prices are quite high
1:00:04and so what you see here in black is the
1:00:05point estimate on the treatment effect
1:00:07and the 90 and 95% confidence intervals
1:00:10so you see that treated farmers are
1:00:12storing more they have higher
1:00:13inventories basically all the way out
1:00:15until the lean season starts and then
1:00:17they start selling off their their
1:00:20inventories in terms of net revenues net
1:00:22revenues are initially actually lower
1:00:24for our treated guys or our treated
1:00:25farmers
1:00:26um uh that's because they're holding
1:00:28back on on selling to the market right
1:00:30so they actually lower net revenues at
1:00:32first a few people are actually buying
1:00:34um uh with their loan and then later on
1:00:37the season they're releasing uh that
1:00:39stored Maze and at a higher price and so
1:00:41their revenu is go up and I'll show you
1:00:43in a moment when you add all this up you
1:00:44get a very strongly um positive total
1:00:47effect on
1:00:48revenues household consumption um we see
1:00:50similar patterns but it's not going to
1:00:52be statistically significant um sort of
1:00:54like Harvest levels consumptions often
1:00:57also quite difficult to measure uh with
1:01:00Precision okay so this is showing you at
1:01:02an individual level the results um so we
1:01:04look at results broken down by those
1:01:06three rounds of of um uh of surveys so
1:01:09you can see what's happening over time
1:01:11but let's I think for the sake of time
1:01:13just focus here on the overall effects
1:01:14so consistent what we saw in the
1:01:16previous graph overall inventories go up
1:01:20um overall revenues go up and
1:01:22consumption goes up though again not
1:01:23quite statistically significantly so
1:01:26okay so it looks like this intervention
1:01:27Works to encourage Farmers to store more
1:01:29and as a result they earn higher
1:01:31revenues now what's happening at a
1:01:33market level so by encouraging greater
1:01:36storage this loan is essentially
1:01:37shifting Supply at different points in
1:01:39the season and so you want to ask does
1:01:41this affect local market prices so what
1:01:44should we expect to happen here well
1:01:46when more individuals are storing or
1:01:48engaging in Arbitrage we would predict
1:01:50that prices will be um higher
1:01:53immediately post Harvest Right this is
1:01:54when prices are typically low so
1:01:56there'll be higher immediately post
1:01:57Harvest as maze is in stores rather than
1:01:59on the market um and lower later on as
1:02:02that stored maze is released okay so
1:02:04higher post Harvest lower later
1:02:06on and that's in fact exactly what we
1:02:08see okay so in this post Harvest period
1:02:10prices are significantly higher in those
1:02:13high intensity in high intensity
1:02:15locations um and they're lower later on
1:02:18in the lean season though this is not
1:02:19statistically significant so it looks
1:02:21like there is an effect on local on
1:02:23local market prices
1:02:26just to note you know one thing that we
1:02:27don't go into in this paper but I think
1:02:29would be a great place for um either a
1:02:31macro model or trade model to think
1:02:32about is you know what is this imply
1:02:34about how integrated these markets are
1:02:36so we're really shifting Supply at a
1:02:38very local level and we're seeing a
1:02:40local price response that to me
1:02:42indicates these markets are not terribly
1:02:43well integrated and I think you could
1:02:45think about applying models to think
1:02:46about what are you learning about the
1:02:48kind of degree of integration of these
1:02:49markets um based on the local price
1:02:52response to local Supply
1:02:54shocks okay finally we turn to uh this
1:02:57question externalities right so this is
1:02:58why I put this in the externality bucket
1:03:01um so what do these GE effects mean for
1:03:03the individual returns to the loan so
1:03:05just to summarize where we're at this
1:03:07loan is designed to enable Farmers to
1:03:09Arbitrage these seasonal price
1:03:12fluctuations high density areas are
1:03:14seeing uh significantly higher post
1:03:16Harvest prices and weekly lower lean
1:03:18season prices so we've sort of
1:03:19diminished that Gap you typically see in
1:03:21prices and so by diminishing that Gap in
1:03:24prices we may be diminishing the
1:03:26profitability of of storage right we've
1:03:28reduced sort of this Arbitrage
1:03:29opportunity and that might reduce the
1:03:31benefit to those who are
1:03:32treated now note that you know this
1:03:34might reduce the benefit to treated
1:03:36individuals but for untreated
1:03:38individuals right folks who aren't
1:03:39storing this might be a beneficial thing
1:03:42right so the folks who are still selling
1:03:43at Harvest Time are now getting a higher
1:03:45price and the folks who are buying back
1:03:47in the lean season are now able to buy
1:03:49at a lower price okay so it's changing
1:03:51the distribution of those of those
1:03:53benefits so we're explore this using the
1:03:55following a specification we're going to
1:03:57regress uh those three main outcomes on
1:03:59inventories revenues and consumption on
1:04:02a dummy for treatment a dummy for being
1:04:04in a high-intensity sublocation and the
1:04:06interaction term and so now beta one is
1:04:09giving you the treatment effect for
1:04:11treated individuals and low intensity
1:04:13areas right folks who don't have many
1:04:14people around them who are also
1:04:16storing beta 1 plus beta 3 is giving you
1:04:19the treatment effect for treated
1:04:20individuals in high-intensity areas
1:04:22right where many people around them are
1:04:23also engaging in Arbitrage and beta 2 is
1:04:26going to give you if there any
1:04:27spillovers to uh untreated individuals
1:04:30who are located in high-intensity
1:04:33areas so here's what we find so let's
1:04:35focus on on this net revenue column to
1:04:38so uh the the treatment effect for
1:04:40treated individuals in high intensity
1:04:42areas is really sorry in low intensity
1:04:44areas is really large much larger than
1:04:45we found before in the pooled
1:04:47specification okay so a lot of these
1:04:48benefits that we found overall are
1:04:50really concentrated among treated
1:04:52individuals in low intensity areas where
1:04:54not many people around them are storing
1:04:57in contrast for treated individuals in
1:04:59high-intensity areas right for them you
1:05:00have to add up beta 1 and beta 3 for
1:05:03them the effect is still positive but
1:05:05it's it's much smaller and it's no
1:05:07longer statistically
1:05:09significant so it looks like these gains
1:05:11are really um much larger when you're
1:05:13thinking about folks who are in these
1:05:14low intensity areas where not many
1:05:16people around them are storing the gain
1:05:17seem to be eroding when other people
1:05:19around them are engaging Arbitrage too
1:05:22and then finally this is not
1:05:22statistically significant but suggestive
1:05:24there might be some small me to uh to
1:05:26untreated individuals uh in
1:05:28high-intensity
1:05:29areas so the takeaways here are you know
1:05:32these Revenue effects are concentrated
1:05:34among low intensity individuals um you
1:05:36the takeaway here is that Arbitrage is
1:05:38most profitable when you're the only one
1:05:39doing it when others start engaging in
1:05:41this this is going to erode uh the
1:05:43profitability of you engaging in
1:05:46Arbitrage methodologically uh one of the
1:05:48contributions of this paper is just to
1:05:50note that the impacts that are captured
1:05:52in small Pilots might not correspond to
1:05:54impacts scale so so note that if we had
1:05:56just estimated this intervention among
1:05:58the low intensity folks right we would
1:06:00have really sort of overestimated the
1:06:02benefits You' probably see if you scale
1:06:03this up to most farmers in the area so
1:06:05we would have said hey this is fantastic
1:06:07for net revenues this is going to really
1:06:09benefit the borrowers when in reality at
1:06:11scale borrowers are seeing much lower
1:06:13gains you know still positive but much
1:06:15lower um that said there are potentially
1:06:18dispill over gains to others and so we
1:06:20try to think about what these GE effects
1:06:22mean for the distribution of gains right
1:06:24so these small potentially um indirect
1:06:27gains per person um we do um a very
1:06:29small model to show this can be a large
1:06:31part of the of the total gains um and in
1:06:34this context it looks like high
1:06:35saturation so sort of scaling up these
1:06:37loans does provide larger social gains
1:06:40from smoother prices um but the gains
1:06:42the actual borrowers themselves are
1:06:44lower okay so lower private gains and
1:06:46that might um explain some of the under
1:06:48under
1:06:50adoption okay so let me um any questions
1:06:53about that before I go on
1:06:57um hello yeah um I was just thinking
1:07:00maybe if you looked at um the effects of
1:07:04maybe risk preferences of people and how
1:07:06that influences whether they they store
1:07:09their Harvest or they sell it yeah it's
1:07:12a great question because so just to to
1:07:14flush it out because it's a really great
1:07:16question so um prices rise on average
1:07:18over time but there is some variability
1:07:21across you know across the years and
1:07:22Farmers don't know exactly by how much
1:07:24prices are going to rise R so there's
1:07:26uncertainty you can think of storage
1:07:28itself as the as the technology right
1:07:29there's uncertainty about the returns to
1:07:31storage um and so we did look at that we
1:07:34do have measures of you the typical way
1:07:36you often ask about risk preferences um
1:07:39uh in these surveys is you run a module
1:07:41where you say you know would you take x
1:07:42amount for sure or a lottery in which
1:07:44you get you know y for 50 50% you get Y
1:07:47and 50% you get zero and you can kind of
1:07:49measure people's risk aversion by
1:07:50changing those X and those y terms um we
1:07:53don't find any correlation between um uh
1:07:56people storing and and risk preferences
1:07:59I have to say I'm not totally convinced
1:08:01that these questions are doing a really
1:08:03good job of capturing kind of real risk
1:08:05preferences um and so I think
1:08:07improvements in in doing that whether
1:08:09it's by having you know these are all
1:08:10hypothetical too so having kind of real
1:08:12estate questions um looking at how
1:08:15people choose to take risks in you know
1:08:17in their actual real real life all those
1:08:19things would be I think valuable um
1:08:21methodological improvements to how we
1:08:23capture this and then you might see a
1:08:24stronger correlation
1:08:26um but I will say if you're interested
1:08:27in this topic in particular um uh hope
1:08:30Michaelson um has a nice review piece
1:08:32looking at um variability in these
1:08:34seasonal price fluctuations and she
1:08:36shows that there's you know a minority
1:08:38but a substantial minority of sort of
1:08:40seasons and years where prices don't
1:08:42rise and so this is a risky um albeit on
1:08:44average profitable uh
1:08:47investment great question other
1:08:53questions okay so let me in my final um
1:08:55five minutes just quickly make um a few
1:08:58comments on this potential synergies
1:09:00between micro and macro and just
1:09:01highlight very briefly a few papers that
1:09:03are trying to do this in the
1:09:05agricultural uh space so this is a
1:09:08review piece uh by by Joe that I just
1:09:10absolutely love and really encourage
1:09:11folks to take a look at uh Joe and
1:09:13co-authors um entitled from macro from
1:09:16micro to macro um so uh I'll try to
1:09:19summarize Joe please correct me if I'm
1:09:20getting any of this wrong but this is
1:09:22also sort of you know my take on on
1:09:23things as well so um the the sort of uh
1:09:26premise of this review piece is that you
1:09:28know each micro and macro they each have
1:09:30their strengths right so micro has had
1:09:31undergone this you know causality
1:09:33revolution in which um especially
1:09:35through rcts in which we've gotten
1:09:37really nicely sort of well identified
1:09:39causal estimates um uh of program
1:09:42effects estimates um of different
1:09:44important uh economic parameters coming
1:09:46from
1:09:47rcts and from micro literature tends to
1:09:49be pretty deeply grounded in kind of the
1:09:51reality and the data of developing
1:09:54countries so doing a lot of field work
1:09:55having a lot of um uh Rich admin Micro
1:09:58Data really sort of grounded in what's
1:10:00going on on the other hand macro their
1:10:02strength is really thinking about kind
1:10:03of big picture questions right so what
1:10:04are policies that have an aggregate or
1:10:07economy-wide effect on welfare so those
1:10:09are the strengths and then uh the common
1:10:12though I'll be I'll very clear not
1:10:13inherent critiques of these literatures
1:10:15are you know micro often is not focused
1:10:17enough thinking about you know what
1:10:19would the aggregate effects at scale be
1:10:21what's going to happen in general
1:10:22equilibrium when when when other things
1:10:24adjust um uh a common critique again not
1:10:27inherent but common is it's disconnected
1:10:29from Theory and from the macro side
1:10:31again common though not inherent
1:10:33critique is it's relying on very strong
1:10:35structural assumptions often uh
1:10:37employing calibrated parameters that
1:10:39might be either you know not well
1:10:40identified um period or maybe are well
1:10:42identified but not kind of from the
1:10:44local local context and so you can
1:10:46imagine you know the the strength of one
1:10:47or sort of the the weakness of the other
1:10:49and you can see how these two things can
1:10:50really start to complement each other um
1:10:53and there's been uh movements in macro
1:10:55that maybe you're discussing in this
1:10:56course uh sort of moving towards uh
1:10:58incorporating you know heterogenity
1:11:00other factors that make it tie a little
1:11:02more easily to the micro dat I do want
1:11:04to highlight here from the micro side
1:11:06kind of moving towards towards the macro
1:11:08there's been uh an improvement in
1:11:10methodological tools that are helping to
1:11:12bridge this Gap so we just saw one on
1:11:14randomized saturation so how can we
1:11:16experimentally um in a well identified
1:11:19way estimate these these GE effects for
1:11:21example um or local effects on on prices
1:11:25there's also been attempts to do
1:11:26randomization at scale so again one of
1:11:28the critiques is that these micro many
1:11:30of these micro um uh uh evaluations are
1:11:33often implemented in small scale we saw
1:11:35an example of where the effects that
1:11:36scale might be be different um there's
1:11:38been approaches to even just do the RCT
1:11:41fully at scale um so there's been you
1:11:43know evaluations in which entire
1:11:45programs are randomized across you know
1:11:46the entire State entire state in India
1:11:49for
1:11:49example um and then finally there's been
1:11:52um sort of this merging between um
1:11:54taking uh um uh say say calibrated or
1:11:57estimated parameters from from rcts or
1:11:59natural experiments um and embedding
1:12:01them into structural models so trying to
1:12:03learn more from uh these rcts or natural
1:12:07experiments um and so uh I won't uh
1:12:09probably have time to go over over all
1:12:11these but there's just um there's been I
1:12:13think a lot of work in this kind of
1:12:14merging of of micro macro um coming from
1:12:17the agricultural technology adoption
1:12:19literature so um uh the piece that we
1:12:21just mentioned um uh on storage loans
1:12:24there been kind of a similar piece by uh
1:12:26Gunther think and Kelsey Jack and others
1:12:28looking at offering seasonal loans to
1:12:30Farmers and seeing how labor responds
1:12:32and shifts around and how that affects
1:12:34local labor market wages um uh work on
1:12:37thinking about uh Bridge access and
1:12:39Market access and how that affects
1:12:41Market Choice um uh juletta and namata
1:12:44had this nice piece in which they um are
1:12:46looking at Farmer mechanization and how
1:12:48that affects labor demand and they've
1:12:50actually um uh basically randomly
1:12:53subsidized people to adopt um
1:12:55uh I think it's tractors in their case
1:12:57or some sort of mechanization and
1:12:58they're seeing how that affects labor
1:12:59demand um and Lorenzo cabur and Jack
1:13:02Willis are randomizing subsidies to
1:13:04Farmers in Kenya to rent out their land
1:13:06to look at frictions uh in land markets
1:13:09um and I have another paper um looking
1:13:11sort of generally at how agricultural
1:13:13policies might look at scale um using
1:13:16existing RCT estimates and trying to
1:13:17embed that in a GE model to think about
1:13:19how when you scale up policies and
1:13:21output and input prices change um what
1:13:24that means for changing average and
1:13:25distributional effects so sort of a
1:13:28growing literature here you see most of
1:13:29these are dated pretty recently um uh in
1:13:32sort of this agricultural technology
1:13:33space um and I just want to encourage
1:13:35folks there's I think a lot of room here
1:13:36to do some exciting merging of micro and
1:13:38macro um and we'd love to list your
1:13:40paper next so let me uh it's 12:15 let
1:13:43me stop there Joe and see if there are
1:13:46uh questions from the from the room
1:13:47thanks do we have questions from the
1:13:49front row I have a ton of question I
1:13:52have a question um so on kind of the
1:13:56timing of liquidity constraints in
1:13:58agricultural markets um so in in your
1:14:02paper right you're relieving your
1:14:04liquidity constraint after the Harvest
1:14:06kind of immediately post Harvest other
1:14:09cases right you might think you want to
1:14:11give out credit right before planting
1:14:13which is when people can invest in
1:14:15agriculture in the think at all paper
1:14:17right that you have here they um they're
1:14:19relieving liquidity constraints during
1:14:20the lean season um and I can imagine so
1:14:24actually in most of these cases if it's
1:14:25credit right you're leaving liquidity in
1:14:27one location but you're actually going
1:14:29to like force them to repay the loan in
1:14:31a different time right and so you could
1:14:33even imagine that like uh a loan at
1:14:36planting time right that lets people uh
1:14:40purchase fertilizer it exacerbates the
1:14:42problem that you highlight in your paper
1:14:43if they have to repay the the loan um
1:14:46post Harvest so in your view right is
1:14:49there is there kind of an optimal timing
1:14:51of credit or or you know how do you how
1:14:55do do those do the results in those
1:14:56papers conflict with each other does
1:14:58that kind of make sense yeah know it's a
1:14:59great question to think about kind of
1:15:01mapping out the full um uh over the over
1:15:04the Season the full distribution of kind
1:15:06of the input of when cash is coming in
1:15:09and when expenditures are going out and
1:15:11Kelsey and I have actually talked about
1:15:12doing something like that um sort of
1:15:14with our two papers so um yeah it is a
1:15:16challenge I would say actually so yes I
1:15:18agree that actually some of these loans
1:15:20if they're then coming due at Harvest
1:15:22Time can really be a problem and in fact
1:15:24um
1:15:25uh there was um there was an effort for
1:15:28example by the malawian government uh um
1:15:30Ryan Dylan has a paper on this in which
1:15:32they intentionally move the timing of
1:15:34school fees closer to harvest with the
1:15:35goal of making easier for Farmers to
1:15:38actually pay um school fees you know
1:15:41moving it to the same time as as harvest
1:15:44and Brian Dylan actually finds that this
1:15:45exacerbates this storage problem that
1:15:47I'm mentioning here and so farmers all
1:15:48sell at Harvest and it's basically
1:15:50making everyone sell at the same time
1:15:52and so they're all getting a lower
1:15:53return because Supply is sort of flooded
1:15:55on the market so I think the challenge
1:15:57here is figuring out first kind of what
1:15:59is the the life cycle of when uh
1:16:01expenditures and um uh an income happen
1:16:04and then also thinking about is there a
1:16:05way to um think with the GE effects and
1:16:07not have everyone do things at the exact
1:16:09same time so that's really the problem
1:16:11here is that you know everyone has
1:16:12school fees at the same time um a lot of
1:16:15bills going do at the same time if you
1:16:16could spread those out that would
1:16:17probably be most beneficial um and so
1:16:19actually our loan is not due at a
1:16:21particular time so farmers have
1:16:23flexibility to repay at any point over
1:16:26the next year and so they can choose the
1:16:27timing of when they're going to repay so
1:16:29we're not forcing them to repay for
1:16:30example at planting time which might be
1:16:32harmful if they're trying to invest that
1:16:34point in inputs so I think in general
1:16:36just more more flexibility and
1:16:38disconnection of the timing of
1:16:40expenditures and um uh an income would
1:16:43be useful and having an eye towards the
1:16:45fact that um for any of these things
1:16:47they going to be aggate shocks like for
1:16:48example school fees where everyone has
1:16:50to pay at the same time that could
1:16:52induce aggregate supply decisions that
1:16:54have effects on local market prices
1:16:56thanks I guess one more question then if
1:16:58no one else in the in the row has it so
1:17:01um do you think there needs to be higher
1:17:04kind of standards for rcts to be used in
1:17:08calibrating macro models like this I
1:17:10mean specifically maybe around standard
1:17:13errors um so like uh I think most rcts
1:17:18are kind of barely powered or at least a
1:17:20lot of them are kind of barely powered
1:17:21to reject zero right but not necessarily
1:17:25um be super informative about the exact
1:17:28point estimate um like how do you
1:17:30account for say if you have an RCT
1:17:32estimate like that could it's centered
1:17:34at 10 but it could be one it could be 19
1:17:37right uh yeah can you still use that
1:17:41yeah it's a great question oh and I'm
1:17:42getting a notice that my internet
1:17:44connection is unstable can you all hear
1:17:45me okay yeah okay yeah so this is a
1:17:48great question and I should say I I
1:17:50think this is an important um critique
1:17:52or concern and I actually I don't think
1:17:53it's specific to you using these an in
1:17:55model so I I hear you which is if you're
1:17:58estimating you know certain parameters
1:17:59you you care about the magnitude not
1:18:01just that it's distinct from zero I
1:18:02think that should be true in rcts in
1:18:04general so imagine you have kind of the
1:18:06most basic you know totally just policy
1:18:09oriented you just want to know does this
1:18:10thing work um you want to compare that
1:18:13probably to the cost of the program and
1:18:15so you want to know not just does it
1:18:16produce any benefits but what is this
1:18:18exact size of the benefits and do they
1:18:20justify the cost um and is this the you
1:18:23know for the cost is it the biggest
1:18:25relative to others so I think you know
1:18:26there is um there should be more of an
1:18:28interest in thinking about the magnitude
1:18:30in general and not just distinguishing
1:18:32from zero um but I hear you I think it's
1:18:34particularly important if you want to
1:18:35think about you know bringing this into
1:18:37uh into a model I think there's other
1:18:39things to think about too so for example
1:18:42um uh you know the time duration of
1:18:44changes so a lot of these rcts are also
1:18:46fairly uh short run and so when you
1:18:49think about for example key elasticities
1:18:50you might want to bring into a model um
1:18:53you know do you mean a short run
1:18:55elasticity or do you mean a long run
1:18:56elasticity and many of these models are
1:18:58assuming unless they're sort of
1:18:59intentionally Dynamic models many of
1:19:01these models are thinking about kind of
1:19:02long run um uh changes or that ways that
1:19:05people are responding to things um many
1:19:08these rcts only run for a year or two um
1:19:10and so if you think there's transition
1:19:12cost are you getting sort of the full
1:19:14effect within that time period And
1:19:15there's been you know pushes on that to
1:19:17try to for example um uh do longer run
1:19:20rcts but a lot of ways in which I think
1:19:22um uh you know I think thinking about
1:19:25putting them into models would help to
1:19:26raise these issues uh more clearly and I
1:19:29think these are issues that should be
1:19:30relevant even to folks who are just
1:19:31caring about kind of basic impact
1:19:33evaluation totally on a micro level um
1:19:37uh anyway and so I think you know these
1:19:39kind of questions are really crucial and
1:19:40and could be pushing forward these
1:19:42questions could be a nice side of side
1:19:43benefit from the the merging of these
1:19:45two of these two
1:19:47literatures please okay can you hear me
1:19:49yes okay I I'm I've got a little
1:19:52challenge with rcts basically
1:19:55uh because I think they tend to
1:19:56undermine um macro problems I mean the
1:20:02the I mean the some macroeconomic
1:20:05challenges under which many of the uh
1:20:09under which they carry out the uh
1:20:12experiment that's number one and I also
1:20:15have a challenge with the scalability of
1:20:16the output o of of of you know of such
1:20:20studies basically I mean if I go to a
1:20:23rural Farm in Malawi and I do a simple
1:20:26experiment and I give 100 $100 and
1:20:30things like that definitely we observe
1:20:32certain effect but that not taken into
1:20:36consideration look for agricultural
1:20:38output basically I would have imagined
1:20:40that the role of the government in
1:20:43taking up excess uh output of the
1:20:46farmers as as a way or providing uh some
1:20:50form of um
1:20:52um storage facility
1:20:55where uh the government takes up excess
1:20:59produce from the market thereby
1:21:01stabilizing the price in the market and
1:21:03things like that and what about the
1:21:05demand side of the the
1:21:10uh purchasing power of the people to
1:21:13also be able to compete in that market I
1:21:15mean so some of the studies honestly
1:21:17they leave me more confused than
1:21:20uh so I don't know how do we address
1:21:22such problem yeah yeah so so great
1:21:25questions let me take the I think two
1:21:26two branches of of your comments which
1:21:28are both fantastic so so one is sort of
1:21:30the limits of thinking about what are
1:21:31you learning in an RCT and how might
1:21:33that be different from the actual impact
1:21:35of the policy at scale um and so there's
1:21:37a nice review piece by um caric merlet
1:21:40Haren and and Paul neous and and maybe
1:21:42others um looking at how things can
1:21:45change at scale and they have kind of
1:21:46three um main factors they're worried
1:21:49about things changing at scale so one is
1:21:51the issue you raised I think about
1:21:52implementation so um yeah people might
1:21:55respond differently if there's sort of a
1:21:57researcher coming in and offering a
1:21:59program the other uh critique that's
1:22:01often brought up that's very similar is
1:22:03you know the the types of NOS for
1:22:05example or even types of governments
1:22:07that we as researchers choose to work
1:22:09with for these rcts are typically some
1:22:11of sort of the you know you're only
1:22:12going to run this RCT if you feel like
1:22:14the NGO you're working with is
1:22:16implementing the program properly um uh
1:22:19has the time and the managerial capacity
1:22:21to work with you on r on randomization
1:22:23understands randomization
1:22:25all that and so there's a select group
1:22:26of NGS or even governments that are
1:22:28willing to work with you on these rcts
1:22:30and so we're getting evidence from maybe
1:22:32the best ngos what would this look like
1:22:34if it were scaled up by other NOS or
1:22:36different governments so implementation
1:22:38can be quite different uh at scale and
1:22:40the two other things which you didn't
1:22:41mention but I think are worth
1:22:42highlighting is you know one many these
1:22:45micro studies are done in one particular
1:22:47area and so there can be uh
1:22:48heterogeneity and treatment effects um
1:22:51where your results from one area might
1:22:53not apply to other areas
1:22:55um again this is where a place where I
1:22:56think if you have a model of um you know
1:22:59what explains that heterogeneity and if
1:23:01you can sort of explain the
1:23:02heterogeneity within your study area you
1:23:04might be able to apply that to other
1:23:06areas to predict how those effects would
1:23:08change and the Third Kind of Camp is a
1:23:10is um uh ge effects thinking about for
1:23:13example um uh changing input or or
1:23:15output prices or other shifts that
1:23:18happen um this is where I think some of
1:23:20these randomized saturation designs can
1:23:21be can be quite useful for trying to to
1:23:24disect locally and then you might want
1:23:25to embed that for example in a a GE
1:23:28model so that's on scaling so there have
1:23:30been I think I think people are taking
1:23:31really seriously some of these concerns
1:23:33you're raising and are trying to move
1:23:34towards that um both in sort of
1:23:37qualitatively talking about how these
1:23:38can be different and then trying to
1:23:39quantitatively develop tools that can
1:23:41address these the first issue of
1:23:43undermining is really important and
1:23:45again this is where I think you're right
1:23:47that the micr literature is often focus
1:23:49on solving one problem and not thinking
1:23:50about kind of are We creating new
1:23:52problems in other areas or how do these
1:23:53problems interact um in particular for
1:23:56storage so I will just say with that
1:23:58particular example so again we work here
1:24:00in in Kenya and the national cereals and
1:24:02produce board in Kenya um uh does do
1:24:05some storage for the purpose of
1:24:06stabilization that you're mentioning um
1:24:08it's not been terribly successful so you
1:24:10still see prices fluctuating quite a bit
1:24:12over the season and they've personally
1:24:14had issues around um uh you know just
1:24:17sort of uh management issues and
1:24:20Corruption issues and a bunch of issues
1:24:21over time um so it hasn't been terribly
1:24:24successful but there certainly could be
1:24:25a role for that um and you're right this
1:24:27is really kind of going to the the micro
1:24:29level of why aren't farmer restoring you
1:24:31might ask an intermediate question of
1:24:32why aren't Traders or intermediaries
1:24:34storing who don't have these kind of
1:24:36credit constraints um and happy to chat
1:24:38about that um as well um offline
1:24:42maybe Lauren let me let me take uh two
1:24:44questions from the Q&A uh one they're
1:24:47both excellent questions one you will
1:24:48have thought a ton about and one maybe
1:24:50not so much about um the first one is by
1:24:53Renee D daily own um in the context of
1:24:58the crop fluctuation problem how local
1:25:00are the markets for these crops to
1:25:03consider General equium should we think
1:25:04at the municipal level the national
1:25:06level or the regional level that's the
1:25:08first question that's the one I think
1:25:10you thought about a bunch about um the
1:25:13second one is from Emmanuel marray
1:25:16laclair I was wondering if the speaker
1:25:18had some insights on the use of
1:25:19observational plant data which are
1:25:22abundant in manufacturing Pete cleano
1:25:25had a long list of such publicly
1:25:26available data sets to learn about
1:25:28technology diffusion
1:25:30empirically so I guess that's a little
1:25:33bit of a swing because it's thinking
1:25:34about plant data and observational data
1:25:37um go ahead yeah so um on the first
1:25:40point yeah about how integrated are
1:25:41these markets so um you know we didn't
1:25:44we haven't done this exercise but this
1:25:46is where I was saying I really think
1:25:47this would be sort of a neat exercise
1:25:49you could do a studies like this so we
1:25:50find you know local price effects um
1:25:53from local storage that's suggest that
1:25:54these markets are pretty poorly
1:25:56integrated meaning that sort of the
1:25:57right context to think of a market is a
1:25:59very local level um there's been um
1:26:02there's been other work actually Ted has
1:26:04another paper um in a in a slightly
1:26:05different part of Kenya where he's
1:26:07looking at Price effects for different
1:26:08Goods um uh looking at doing a
1:26:11randomized saturation design with give
1:26:13directly of giving cash transfers and
1:26:15there they don't find big price effects
1:26:16um of local sort of infusion of cash or
1:26:19or which can generate in some cases
1:26:21Supply shocks so um there's some
1:26:23conflicting evidence there and I think
1:26:24thinking about developing a model to
1:26:26actually infer from the price shocks
1:26:28that are seen what does this mean for
1:26:30different goods for example um uh uh in
1:26:33different locations what does this mean
1:26:35about how integrated markets are I think
1:26:36there's there's lessons that could be
1:26:37taken um if you could apply a model to
1:26:39thinking about that um for now when
1:26:42you're actually in practice doing
1:26:43randomized saturation designs this is a
1:26:45huge issue and frankly a lot of it's a
1:26:46little bit you know um ad hoc based on
1:26:49your particular knowledge of the
1:26:50particular context um at what level you
1:26:53want to randomize
1:26:54I will say um this last paper I've
1:26:56mentioned here which is looking sort of
1:26:58at how do things change at scale um this
1:27:00is trying to basically model kind of
1:27:02price effects and how they're getting
1:27:03transmitted across across locations and
1:27:05taking price data and trying to back up
1:27:07um trade costs and the like um our hope
1:27:10is that eventually uh and this is still
1:27:11a bit of a work in progress but um um
1:27:14almost ready and our hope is that this
1:27:15could be a model you could use to almost
1:27:17do like power calculations for um you
1:27:20know when you make certain assumptions
1:27:22about um how integrated markets are what
1:27:24trade costs look like Etc then uh what's
1:27:27the right level at which to do your
1:27:28randomized saturation design um so again
1:27:32power calculations are always sort of
1:27:33guess ahead of time you make a guess
1:27:34about what you're going to see um but it
1:27:36could maybe help inform that decision on
1:27:38a practical note um and I think then
1:27:40conversely taking models to kind of
1:27:42apply it to existing results would be
1:27:43very interesting so that's the first
1:27:45point um the second question was about
1:27:47sort of observational data um yeah so I
1:27:49have mainly shown you rcts here just
1:27:51because there's been such a huge
1:27:52abundance of them um but of course
1:27:54there's also ton of observational uh
1:27:56data looking at um uh you know
1:28:00misallocation for example Doug and Chris
1:28:02udre have a paper on misallocation uh
1:28:04looking at uh input use uh at at a
1:28:06actually plot By plot level um so one
1:28:10nice feature of this is you have of
1:28:11these many of these uh datas you'll have
1:28:13not just for individual households we
1:28:15have data but you'll have data for
1:28:16different plots the same household is
1:28:17growing and that's actually quite quite
1:28:19an interesting data set to look at um
1:28:21Chris udy also has work looking at
1:28:23diffusion
1:28:24of agricultural Technologies based on um
1:28:28uh based on your network using
1:28:29observational data um so there is a lot
1:28:32you can do with observational
1:28:33observational data it's just it's
1:28:35particularly nice if you have um uh
1:28:38variation that's um uh that's exogenous
1:28:40coming from RCT plant data there's
1:28:42there's there's a lot I didn't go over
1:28:43technology adoption um I should say in
1:28:46manufacturing but there's a lot of work
1:28:47there so for example David Akin has a
1:28:49work on um with with sever other
1:28:51co-authors on uh on soccer ball um
1:28:54soccer balls in Pakistan and Adoption of
1:28:56a new cutting technology for those
1:28:57soccer balls um uh that's also an RCT uh
1:29:01I'm trying to think about um
1:29:03observational data but I think there's a
1:29:05lot you can do if you think about kind
1:29:06of knowing the network and how these
1:29:07firms are are are um are connected but
1:29:11then you have to find some sort of
1:29:12plausibly exogenous shock um which can
1:29:14come from natural experiments as well as
1:29:16as well as
1:29:19rcts okay thank you very
1:29:21much um have some question
1:29:25unfortunately
1:29:26um somebody I'm sorry did I cut I have
1:29:31some
1:29:31question okay go ahead I we'll take one
1:29:34last one okay thank you uh my name is
1:29:38Emmanuel so um most of the studies tend
1:29:42to focus on uh a single intervention
1:29:46that address adoption constraints uh for
1:29:50example giving um subsidies to Farmers
1:29:54and yet we know that actually farmers
1:29:56face so many production risks uh they
1:29:59also face so many markeet risks uh an
1:30:03example uh is input quality that is
1:30:06always a problem in developing countries
1:30:09but also output prices to
1:30:12fluctuate uh then Farmers also face um
1:30:15weather uncertainty for example now many
1:30:18researchers tend to focus on a single
1:30:21intervention uh yet farmers face
1:30:25multiplicity of so many challenges uh
1:30:28doesn't this require a bunding of
1:30:31interventions that address different
1:30:34problems uh so that uh adoption
1:30:36constraints can actually be resolved and
1:30:39then if we focus on one intervention are
1:30:41we surprised then to see low adoption of
1:30:44Technologies in developing countries if
1:30:46we tend to offer only one um
1:30:49intervention to address so many
1:30:51challenges yeah it's a great question so
1:30:53there could be could be interplays
1:30:54between these uh between these these
1:30:57these um uh constraints and it might be
1:30:59you have to address all these Contra
1:31:01constraints simultaneously or at least
1:31:03some in combination to get the effect um
1:31:05and in some ways I think um there's been
1:31:08a bunch of reasons that we haven't had
1:31:09as much of that you know having to do
1:31:11with both the fact that it's difficult
1:31:12um you know you're limited for power and
1:31:14so you have to choose certain arms
1:31:16there's also I think some unfortunate
1:31:18kind of academic incentives in terms of
1:31:20publishing about really wanting to to
1:31:21nail down what is the barrier um that
1:31:23that stop people from doing kind of
1:31:25overall bundled interventions um there's
1:31:28been some attempts to try to look at
1:31:29interaction terms right so for example
1:31:31the Carin at all piece that we looked at
1:31:32does try to look at risk um the
1:31:34interaction of risk and credit
1:31:36constraints so what happens when you
1:31:37alleviate both and it looked like for
1:31:39fertilizer that does seem to have a
1:31:40bigger effect um and for others you
1:31:43didn't see a significant interaction
1:31:44term so there's been some attempts but
1:31:46sort of uh if you want to kind of parse
1:31:48out both the different uh barriers as
1:31:51well as the interaction that's just
1:31:52going to add you know it's going to be
1:31:54more difficult in terms of in terms of
1:31:55power I will Note One Piece if you're
1:31:57interested in kind of thinking about um
1:32:00uh alleviating multiple constraints at
1:32:01once there's a nice piece by um Maya
1:32:04Duru and her her co-authors um looking
1:32:06at actually the same NGO that we looked
1:32:09at in our storage paper um and they
1:32:11offer um a variety of services that try
1:32:13to address actually all three of those
1:32:15constraints so credit um uh uh risk and
1:32:19information all at once and they find
1:32:20very big effects from doing all three at
1:32:22once so there have been I think some
1:32:24that's a very new paper and there have
1:32:25been new attempts to move in that
1:32:26direction um but I agree it's an
1:32:28understudied area of how do these
1:32:30constraints uh
1:32:32interact that was a good last question
1:32:34I'm glad we took it um unfortunately
1:32:36we're out of time uh I want to thank um
1:32:39Lauren and Eleanor and um tomorrow we do
1:32:43not have let's let's unmute and clap
1:32:46yeah there we
1:32:49go um tomorrow we don't have class
1:32:52respecting the Easter weekend for the UK
1:32:55um but we will return um next week uh
1:32:59Wednesday Eleanor will have her session
1:33:02and then um Thursday and Friday we'll
1:33:04have more sessions so uh also another
1:33:08thank you to um Mandy and kiry and
1:33:12Lauren um we'll see you all next week
1:33:15bye bye thanks
1:33:19everyone thanks Lauren great
1:33:21presentation just wonderful she there
1:33:24still Maybe not maybe
1:33:27not and thanks
1:33:30eler thank you

About this transcript

This page contains the full transcript of Lecture 8: Barriers to technology adoption: what we know from micro empirics by CEPR & VideoVox Economics, generated from the public captions YouTube serves with the video. The transcript has 18,342 words across 2,645 segments, with the original timestamps preserved so you can click any line to jump to that moment in the embedded player.

What you can do with it

Use the transcript to take notes, quote the speaker, build a study guide, generate a summary with ChatGPT or Claude via the YouTube Summary tool, or export it as a timed subtitle file with YouTube to SRT. You can also re-open it in the transcriber to choose a caption track available on the video.

Free YouTube transcript tool

YouTube2Text is a free YouTube transcript generator — no signup, no daily limit. Paste any YouTube link and get the full transcript instantly, with timestamps, click-to-jump, caption language selection, AI prompts for ChatGPT, Claude, and Gemini, and exports to TXT, SRT, VTT, or Markdown.