Good, good morning, uh good afternoon,

good evening everyone. Thank you for uh

joining this side event on small air

estimation

bringing theory to practice. Uh we will

have uh the next an hour and uh probably

half an hour and 30 minutes for uh uh

covering the the topics in our side

event is one of the 57 statical

commission side events which is being

held uh virtually uh we've um

distinguished guests here speakers uh

presenting uh regional aspects of the

the work and country examples and also

more from the uh uh the the the World

Bank and UNSD in terms of organizing and

managing uh this project so far. So uh

without further ado, I will um I'll pass

the mic to uh our colleague Miss Hawen

who is leading this uh activity from

United Nations Division to do the

opening and overview of the session. uh

we have a very tight uh uh program uh

and lots of participants will try to

make the most out of it. How please uh

go ahead.

>> Thank you so much Daniel. Thank you.

Welcome colleagues. My name is um Jan

coordinator of the intersector working

group working here division. So really

be welcome to this event on smarts that

jointly organized by UNC

um the world by and only from Africa

Asia Pacific and Latin America policies

for the noise behind

um so as countries start to deliver on

the sustainable development goals and

leave no one behind the demand for

disagregated timely and policy relevant

data.

Many NSO face constraints in producing

reliable estimate at subn national lower

geographic areas and using different

survey code.

My mission really offer a practical and

methodologically robust pathway to

bridge this gap by integrating survey

data with administrative data geospatial

and other nontraditional data filters to

produce more granular insight while

maintaining traditional results. G from

the UNSD site have been partnering with

our colleagues um here if you you will

be listening hearing from them advance

the use of smartation

offices really grateful for their

partnership

banking

materials guidelines in learning courses

in workshop

will be hearing their work and also from

countries

and the students are literally

countries that done great work in apply

as

um we look forward to hearing uh from

them and then we also look forward to

better serving our communities

they use as well estimation over to

thank you so much

>> thank you how for uh setting the stage I

think uh we have a sound issue from you

a little bit but I think uh most most of

uh the the uh topics have been I mean

been mean covered uh just making sure

that our partners are in in the house

the regional commission the world bank

who is really pushing this forward uh

for enabling NSO uh to to get to uh work

with small area estimation where where

it makes sense so uh without further ado

I'll Switching to uh the wallet bank

who's uh been the anchor for this

program. Uh our senior economist David

will will take you through the uh uh why

small area estimation and what the bank

is doing in supporting uh countries and

some some of the outcomes of the the

work so far uh they have been uh working

on. So I will uh pass it to David. David

uh floor is yours. Thank you very much,

Daniel. Um, let's give me a moment

perhaps to get the presentation up. Um,

can everyone see the presentation?

Let me also go into slideshow mode.

>> Wonderful.

>> Um, so I hope everyone can hear me well.

If not, I can uh speak louder and thank

you very much um Daniel and Howy for

organizing this uh presentation. It's

really an honor and a pleasure to be

here um with you and with uh so many

colleagues from national statistics

offices. I'm going to be talking about

geospatial small area estimation with a

focus on a recent application that we're

currently working on for Nigeria. Um

this is a nice compliment in my view for

a lot of the work that the bank has done

traditionally on smaller estimation with

survey and census data including the

relatively recently produced guidelines.

um and uh but we are also looking to

extend this um to geospatial

applications.

um so I think we mostly understand the

benefits of data integration. Surveys

measure very important socioeconomic

indicators but they're expensive and

because of that they're small and

because of that we cannot be very

granular with them. So most surveys

cannot go below state or district

levels. And of course we would love more

granular information. And so to do that

we can combine surveys with

comprehensive auxiliary data to increase

accuracy and precision of estimates. Um

and this is uh potentially useful for

many things. We can use it to target

social assistance programs, monitor and

evaluate programs and do quality control

for sample surveys. Um there are

concerns that sample surveys may not

always be representative. For example,

if potential areas in a country are

affected by conflict. Um there may be

some places where enumerators can't go

and data integration actually offers a

method for measuring that or estimating

that. Um traditionally, as I mentioned,

we've used survey census data for

smaller estimation. There's a long

history of this going back to Fay and

Harriet in the late 70s in the US

estimating average income for counties.

Um then in the bank Albert's Loa and Loa

kind of popularized the use of survey

and census smaller estimation around the

world. Uh Jen and Lehi and Molina and

Ralph also made improvements to the

methods. Um and so this is great but

censuses are very expensive. They're

conducted typically once every 10 years

but in some countries not always that

frequently. Um and so there's been more

recently a large and literature

combining survey and geospatial data to

try to uh see if we can do smaller

estimation more frequently. Amazingly

this started in 1988 with a paper that

was 30 years ahead of its time. Um then

everybody forgot about the potential

geospatial data until 2016 and then a a

relatively recent literature as I

mentioned has really pushed forward the

boundaries on it. And there are a couple

uh recent reviews um one of which I I

did that basically show that this this

stuff works um that the the geospatial

data that is available is predictive of

of many important socioeconomic

indicators. Um there's a lot of interest

in big data. I'm focused on geospatial

data in particular because it is

publicly available and we don't have to

worry about selection bias. Um I still

have some concerns with mobile phone

data. I I believe it can be useful for

some applications, but still not

everybody has a mobile phone. Also,

there's been a really revolutionary

increase in the access to publicly

available imagery and indicators spurred

on by platforms like Google Earth Engine

and Microsoft Planetary Computer. Um,

and now a lot of research as I mentioned

kind of that shows that geospatial

indicators really excel at predicting

spatial variation and urbanization. And

if you think about the things that it

can measure like buildings and nighttime

lights and vegetation and land

classification, these are all kind of uh

good proxies for urbanization and how

urban places and that in turn is

correlated to many important social

economic indicators. For example, almost

everywhere on average uh rural places

are poorer than urban than urban places.

Um so uh one area in which this has been

applied quite frequently is poverty

estimation. Um we've done a number of

tests relative to survey data and about

eight countries and shown that the

increase in precision uh is equivalent

to expanding the survey data by about a

factor of 2.5 to nine depending on the

context uh and the measure. Surveys cost

a huge amount of money to implement as

you all know a million dollars and more.

So if we can expand the precision of

these by 2.5 to nine for using a

procedure that is essentially free uh in

my book that qualifies as a huge win. Um

although much of the research so far has

applied this to poverty estimation I

believe it can be applied to many other

indicators and in fact we are doing so.

Not all indicators are sufficiently

precisely measured to do this. Not all

indicators are correlated enough with

population density to do this but many

are. Um so just to give you an example

of the best case scenario and this is

from older work done in Tanzania

um and uh we use geospatial smaller

estimation to go basically for from the

district to the commune level. Um that's

the the lower level is on the right. The

higher level is on the left. Um and

what's notable about this that we don't

show here is that the average precision

is measured by the coefficients of

variation is about the same in both of

these pictures. And so the the first one

did not use geospatial data. The one on

the right does. And it just shows in

this case, which was kind of a best

example, best case scenario, we're able

to uh go a level lower in terms of

administrative units with no loss in

precision.

Um so this is great. And then you might

think why isn't everybody using it? Um

uh but there are sort of some obstacles

I think that have held back adoption.

One of them is that there is a variety

and sophistication of methods um that

have been used for this and that can be

confusing and people have used linear

mixed models which are relatively simple

but even somewhat complicated like

empirical best predictor models. These

have been used uh traditionally for

survey and census data. They can also be

used for geospatial data. There's also

treebased machine learning like extreme

gradient boosting and now more recently

mixed effects gradient boosting. um you

so th those are uh maybe one step up

from linear mix models in terms of uh

complication sophistication but then a

lot of the literature has been focused

on deep learning approaches especially

convolutional neural networks and then

more recently AI with using foundational

transformer models. Um these are kind of

AI big AI models that have been trained

on many uh many thousands of terabytes

of imagery um to recognize features and

then they can be fine-tuned with data um

to develop predictions um tailored to

data that's so that's one issue is that

the methods can be sophisticated um a

second issue is that the EA location uh

can be sensitive notably the demographic

and health surveys would publish jitter

EA location information, but uh not all

surveys come with EA location

information. Um and so for for public

researchers, these can sometimes be

difficult to obtain. Um the this can be

somewhat comp complicated to implement

and requires considerable computing

capacity and uh memory. This is

especially true for the more

sophisticated deep learning and AI

methods. And of course, these also

require very strong technical skills.

the tools and the documentation are

still evolving and developing. We're

trying um but we're frankly not gotten

as far as we would have liked in terms

of getting the tools and documentation

out there. Um so these are all obstacles

that are kind of holding back adoption.

The hope is that through um events like

these and finishing tools and

documentation and doing workshops of the

sort that uh we've organized in the in

the recent past that the word will get

out and people will start uh taking

advantage of the benefits of publicly

available geospatial data. So what we're

trying to do is uh work on applications

of generating smaller estimates for

indicators for countries doing

additional research and evaluation

testing methods using geoloccated census

data. There are a few countries where

we've been able to obtain geoloccated

census data and those are great for

testing methods and then developing

tools and we're working on tool two R

packages in particular one called

Geolink and one called Podmap. Um the

first one helps download publicly

available geospatial indicators. The

second one helps integrate it with

survey data. We we've also used to learn

to utilize Google Earth Engine which is

a very powerful platform for obtaining

publicly available geospatial data. Um

and of course now with the development

of AI and the AI chat bots uh you can

just ask it to write code for you.

Momentum is accelerating here but a lot

of work remains. Um so I'm going to talk

here mostly about an application I'm

working on currently um for our country

team in Nigeria um where I'm uh doing

geospatial smaller estimation for 10

indicators. Um the 10 indicators are

listed there. I think what's neat about

this is almost all of the work and sweat

goes into uh downloading the geospatial

indicators. Once you have that, the cost

of adding new survey indicators is

essentially negligible. it's just you

know switching in a different dependent

variable in the same code. Um so uh you

know going to 20 indicators would not be

difficult um if they're if they can be

predicted by geospatial data. And so

just to give you a sense of what we're

doing uh we created the shape file of

about a million one square kilometer

grids covering all of Nigeria and then

we use that to obtain publicly available

geospatial features starting with

population estimates. Um there are some

relatively recent population estimates

produced by World Pop that we used and

when we use that we find that there are

about half of the grids in Nigeria are

populated nearly 500,000 and so we can

obtain geospatial data for those

populated grids. Merge those with the

survey data using area centrids obtained

from the survey. We were able to get

those then estimate a model using the

survey data use that model to predict

outcomes for each grid. weight it by the

estimated population and aggregate it to

the desired geographic level. Um we're

doing both wards and local government

areas. The most complicated part of this

is using bootstrap techniques to

estimate confidence intervals. Um that

can be difficult. Um uh but the our aim

is to get sort of documentation out on

that on how to do it. Um, so just to

give you a sense of the geospatial

features being used, the modeled

population estimates come from World

Pop. Nighttime lights um can be

downloaded from the Colorado School of

Minds website. There's building data

from 2018 and 2023.

Um, land cover, crop land, open- source

cell tower location, which may not be

super reliable, um, but it we can use

it. Um, it's also not the most

predictive variable. Um, as I'll show

you, vegetation index, average rainfall,

aerosol and ozone index, which are

measures of pollution that are local,

and then conflict events. Um, we also

experimented with these new and

interesting Google deep mind embeddings,

which are available on Google Earth

Engine. They work well on their own, but

when we have all these other variables

in in the model, they the embeddings add

very little. Um so you can estimate a

model at the grid level. P hat here is

our survey estimate of poverty in a grid

that's been uh matched to the

enumeration area. Um we use as

predictors predictors at the grid level

and predictors at the target area level

which is a ward or an LGA. And there's

an error term uh both at the ward level

and at the grid level. Um we use an arc

sign transformation because we're

estimating proportions and we want to

keep the uh estimates bounded between

zero and one.

Um we also in this case we're using

extreme gradient boosting. Um I find

that's widely used and very mature

software. It's been very well tested.

We've tested it in house. It slightly

outperforms linear models in many cases.

Um not always of course. Um but it has a

very flexible functional form that

handles nonlinearities and interactions

quite well and you can measure the

importance of predictors through sharply

decompositions without too much

difficulty. So it's a little more

flexible than linear models and a little

more interpretable than the AI and deep

learning models. Perhaps not as accurate

as mixed effects gradient boosting, but

that's a little more complicated and we

haven't tested it. And uh I you know

probably not as accurate as

convolutional neural networks or

transformers. Um but other times that

we've tested that they use we use

different features than we do here. And

so I I think with the extra features we

have here we can get very solid

estimates even with extreme gradient

boosting. Um and indeed the models are

pretty predictive. If you look at R

squar these are in sample R squares but

there is some regularization. So it's

probably not overfitting. Um when it

comes to XG boost and just predicting

these indicators um the R squares are

quite high. Um they do vary. Um so then

I mentioned you can do sharply

decomposition to look at what measures

are important and when we look at

poverty um it looks like it's buildings

uh pollution and land classification and

nighttime lights that's doing most of

the work. Um and I think this makes

sense. These are all correlated with

urbanization kind of as we'd expect. Um,

and fortunately I think they're pretty

solid in terms of the quality of the

data. Um, obviously not every building

is going to be correct, but on average

um, there's very useful information. And

I think what's nice is that this, as

we've seen in the past, leads to large

increases in precision relative to the

direct survey estimates. Um, these are

the mean width of the confidence

intervals of estimates. Um and you know

they they fall by over a half. Um and uh

you know that's really impressive I

think because you know this the mean

confidence interval widths are basically

proportional to the standard error and

the standard error is the square root of

the variance and the variance is

inversely proportional to the size of

the sample. So this is uh an efficiency

gain approximately equal to expanding

the sample by a factor of eight or six

um by this measure. Um so that that I

think that's again a big win. They are

less precise than the the quote

representative state level estimates. Um

so there there's a average confidence

interval width of 15 for the estimates

that are currently published. We would

go up to 26 even for LGA estimates if

you use this method. And so I think this

this leads to kind of some philosophical

questions of how precise is precise

enough to publish. Um so I just want to

give you a sense of these are

preliminary estimates but these are what

the maps look like. Um they show quite a

bit of poverty in the northern half of

Nigeria. Um but there are pockets um and

you can see where there are pockets

where it's redder than other places and

even in the northern and the northeast

there are pockets where it's notably

less poor than other places. Um and also

in the south you can see there there

there's quite a bit of variation even

within states. Um these have been

benchmarked to match the survey

estimates at the state level which I

think um is important. Um you can see

somewhat similar patterns but also some

differences in multi-dimensional poverty

which is I I think interesting. Um and

also some pockets um where you can see

particular words. Um we can look at

other indicators like the secondary

enrollment rate for education. Um again

uh different pockets sort of towards the

middle of the country where that's

weaker. Um so these are interesting. I

think for someone who knows Nigeria

better than I do they would be very

interesting but I want to again caution

that these are preliminary. They need to

be reviewed within the bank and by the

government. Um so uh you know I I

they're not ready to be used quite yet.

Um anyway to conclude um I feel that the

geospatial SAE is practical and

expensive and useful. Um all of this is

using publicly available data. Um so

it's relatively cheap and for those

reasons I think it should be used

routinely. Um getting to kind of routine

use requires tools, knowledge and

computing power. Um and uh I think we

can help certainly with the tools and

the knowledge. Um re the research on

methods and there's been a lot of

research on methods. I think it's useful

but it shouldn't stop applications at

this stage and in many cases where we've

compared methods the differences in

accuracy are really not that major. So

as an extreme example we did some

there's a recent paper that's coming out

soon that compares uh different methods

for combining census and survey data and

we found that the prefer when we

simulated a targeting program the

preferred method beat kind of the less

preferred method slightly and that

translated to a 0.01 01 percentage

points in simulated poverty after a

targeting program. So in my view that's

kind of small enough that it shouldn't

cause a lot of debate or hold back

adoption. Um that you know there can be

a lot of discussion about methods but in

practical terms the differences are are

probably not that major in many cases.

Um, so meanwhile the methods offer

trade-offs on interpretability and

simplicity to estimate parsimony in

terms of how easy it is to communicate

model parameters. And so if the

differences in accuracy in a practical

sense are not that huge, maybe one does

want to go for a method that offers more

interpretability or is simpler to

explain. Um, there can be bias in this

kind of technique. The model based

estimates, they reduce sampling error.

The reduction in sampling error

outweighs the introduction of model

error. So it's a net win. The estimates

are more accurate on average, but there

is model error. And so um we have to

think about how to deal with that. And

maybe this involves some sort of redress

mechanisms of being open to uh a

procedure for handling complaints if

there are complaints. It is possible

that a place looks less poor in terms of

its urbanization than it is. Um and even

though everything is kind of more

accurate on average, um these cases can

be important. And then I think we need

to think a little harder about to decide

when estimates are sufficiently precise

to publish. Um so survey based estimates

often uh there's a threshold based on

the coefficients of variation adopted by

many national statistics offices.

Actually I feel this is problematic for

proportions because it's it you know it

varies whether you use a measure or its

complement. So it's quite possible we

could publish the non-poverty rate but

not publish the poverty rate and that

doesn't really make sense to me. or you

could publish the inchool rate but not

the out of school rate. Um so I'm not a

huge fan of coefficients of variation

for proportions. They can be useful for

other things. Um and then in general,

you know, these estimates could be

useful for policy even if they're not

entirely precise. So uh thinking about

you know what what the right uh

precision measure if any for publishing

these I think is is important. Um of

course quality control regardless is

crucial. Um so whenever we do this there

does need to be some sort of process of

uh evaluating the estimates making sure

they make sense um etc. But I do believe

that uh it is not useful to suppress

useful estimates um and that these

techniques um can be very widely applied

and really help uh provide more policy

relevance and useful data um worldwide.

Thank you very much.

Thank you David I think for walking us

through this uh impressive example on

Nigeria. I think I'm sure everybody is

looking forward to see the the published

results which would be uh something that

uh you know the NSO would be very much

interested in picking up uh uh the

pieces in on their side.

uh it's there are a couple of questions

but I will we'll come to the questions

at the end so that we can answer them

along the way. I'm jotting some of them

in the chat uh some of them in the Q&A

and we have some from uh the the the

submission and registration. So I will

come back to that but really thank you

for uh putting perspective putting the

research into practice. I mean that's

that's what everybody is looking uh

forward to. uh and saying that the

regional commissions are at the center

of this making sure this these come to

practice by uh you know doing the

interlocation work that brings NSOs in

the region to to to this uh advanced

research methodologies use and

implementation and I will uh transition

to our regional colleagues we have ECA

and ELA um sorry escap on this session

we'll start with ECA uh I'll share your

presentation Angela Angela is our uh

lead from UNCCA who's pushing this with

this work with uh the region the African

region and a number of activities have

been happening in the past couple of

years uh on ACE I think it would be uh

uh great to see uh where they are at and

what they're uh uh pushing towards so

I'll I'll try to put your presentation

Angela in the meantime you can introduce

yourself please

>> thank you so much Daniel um as he has

indicated my name is Angela Chicho and

I'm a statistician at the African center

of statistics uh of the United Nation

Economic Commission for Africa I'm happy

to share with us uh because we've we've

each been given five minutes so it's a

really short presentation we're about

slide. Yeah. But uh it's really to

report on uh what we've been able to do

in the year 2025. But maybe as a way of

background I would like to share that uh

we've been doing this since uh 2023

and we have seen of course improvements

based on uh the experiences that uh we

encounter as uh you know this um course

is um uh

as this course is done by the

participants. Could you kindly go to the

next slide please? Yeah. So like I

mentioned um we started in 2023 and uh

in terms of uh trends uh we have seen an

increase in the number of countries uh

that have been uh um sending nominees to

participate in this uh uh e-learning

small area estimation course. Um I think

probably we could share the link in the

the chat for those that don't know about

it to maybe go there and see but uh it

is a self-paced uh course e-learning

course and uh it lasts uh about 7 weeks

it has several modules and so on and so

forth in terms of materials. Um so in

the case of Africa for example in 2023

we had only nine countries but of course

also the approach at the time was

different because the call was just open

and it was uh whoever saw the call that

you know self-enrolled and uh at least

the case of Africa we decided that we'll

pick like 30 participants

um for a manageable class in the next

year the approach was different we

reached out to the heads of national

statistical offices and requested to

nominate uh but this that time it was

really focusing on the anglophone

countries um and not all of them at that

time though. Uh so we had about 12

countries and this time round we put out

a call to all the 54 countries including

the franophone and uh we had a response

of about 30 countries. Now specifically

in 2025 uh we got about 70 nominees as

uh people proposed by the heads of NSO

because they understand their people

that is at least the the basis for

requesting or having the heads of NSO as

the entry for us and uh we requested

them to um register themselves at least

for the Anglo one and 41 did but of

course we can see the numbers keep

reducing and then of course committing

to uh you know have to do the course

over the 7 week period. You can see the

numbers again also keep producing and

ultimately those that actually do the

assignment because if you do not do the

assignment and pass it then uh you

cannot say that you have completed or

done the course or you know uh qualified

to take a certificate that you at least

have an idea. Yeah. So that has been uh

or that was the case with the English or

the anglophone. uh in the case of the

franophhone we had nine countries that u

sent nominees uh of which uh there were

about 27 uh 17 registered and uh

unfortunately in the case of the

franophone were not able to have the

cost done as per the model that is used

because there's an issue with some of

the materials but it is still planned

that once this is completed then uh

they will be benefiting from uh um from

the course. So the course is done in

such a way that uh for those that enroll

uh over the seven weeks at least on a

weekly basis they'll have sessions with

uh a facilitator a course facilitator

who takes them through uh what you call

a synchronous class and uh uh that is of

course still virtual. uh but during this

session um the discussions about the

challenges uh that the participants may

be experiencing especially with the

videos or the materials they would have

interacted with in the course of the

week and uh um as you can see I think it

still comes up as an issue a challenge

the the the attrition uh but we need to

see what would be a solution for it.

Now, we also did have uh an in-person um

training um for sorry, someone is trying

to call me. Yes. So for an inerson

training for the the company that uh

David has just uh taken us through and

of course the prerequisite was that uh

the the the participants for the in

person and this was uh only for for a

select um set of anglophone countries um

had to have completed the the e-learning

course the small area mission e-learning

course with proof that they had

completed it uh because it was forming

being part of you know the base for what

was going to be learned in person though

of course this one is skewed more to uh

the geospatial uh side as you can see it

is on earth observation data so it was a

5day more or less really intensive

workshop um in terms of the material

being uh offered and so on and uh we are

really grateful to the partnership from

the world bank the east African

community as well as uh UNSD um that

supported uh uh for for for the workshop

to to happen this training workshop. So

in a nutshell we had about uh 15

participating countries. Could you just

go back one sorry? Yeah. So we had 15

participating countries and those are

they and uh a total of 26 participants.

We also ensured at least for the host

country which is Kenya we invited um the

data person that uh is in the office of

the UN RCO that is the coordination or

the coordinator's office. Next please as

I conclude I just thought I'd highlight

some of the challenges that uh continue

to persist. Um I think of course people

who start with R and other uh um

packages usually beg your pardon other

packages besides R usually have a

challenge because this course is run in

R and uh so if one doesn't uh do that

foundational bit of uh learning R they

have a challenge um doing the course and

then of course uh this has really

persisted especially for the e-learning

the The completion rate continues to

remain a challenge but uh we hope that

along the way even during this webinar

we can have maybe ideas on how we can

overcome this one but they do exist

opportunities at least in the case of

Africa. We strongly see um support from

the top management and by top management

I mean the heads of the national uh

statistics offices and this is evidenced

by their responsiveness to a request for

nominees. Um the other is that uh um the

interest or for small area estimation uh

work is also evidenced in the

application especially by those that

have you know gained the skills of

course later during this webinar.

Colleague from Ghana will be sharing the

Ghana experience. Um David alluded to

the Nigeria case but then also as

individuals some uh uh uh uh

participants from course have gone ahead

to do their own papers. An example is

some colleagues from Kenya and uh we

still think that uh at least not to

leave people behind. So it would be

great to have the materials you know

expanded to other packages that people

are more familiar and more comfortable

with uh be it STA or Python. So I'd like

to thank us for uh thank you all for

your attention and uh also appreciate

our partners uh once again and uh thank

you so much for listening to me.

>> Thank thank you Angela. Thank you uh

taking us to the story of the region in

the past year and I think uh I was um uh

part of the the in-person training and I

have witnessed you know the the progress

even from countries and also the the you

know the innovative ways of participants

uh pulling other I think you were

mentioning about other tools like strata

and python we had a number of

participants who were proficient in

stata who were able to even pull their

data from Stata into uh uh this this

workshop and I think it's it's something

that we need to think about putting some

more examples in that in that area.

Thank you for that. We'll come back to

questions later on uh at the end. Now

I'll pass it to uh and you know and

saving some time I'll pass it to uh Roth

from uh Bangkok. Uh he's also leading

the regional uh work in Asia and

Pacific. Um, back to you. Uh, please go

ahead.

>> Thanks so much, Daniel. Um, so give me a

few second while I share my screen. Um,

hopefully it's showing up all right.

Seems to be okay on my side. Can you let

me know if you see the screen?

>> Okay, perfect. Uh well thanks thanks

again Daniel and thanks uh David and

Angela for um sharing the work that

you've been doing. Uh good morning good

afternoon uh evening everyone. So my

name is Sana Rod. I'm an associate

statistician from the statistics

division here at ESCAP. Um and it's a

pleasure to be with you today and share

uh ASCAP um well uh go beyond the

training uh to to share a bit more on

the capacity development program um once

that uh that we've uh conducted back in

20 uh 25.

Um I should first start off by saying

that uh ESCAP has been working closely

with uh UNSD, ECA uh and UN agencies

such as UNICEF and development partners

namely the World Bank to implement the

capacity development program. uh so

their support was uh very much crucial

to the success of the 2025 um activities

and even before uh getting into the

activities I should also mention that um

our SAPE program built on the decisions

of our previous committee uh I mean back

since uh the seventh session uh which

was in 2020 to prioritize data

integration and innovation integrate big

data into official statistics and

promote a whole of society approach um

to implement the um our uh declaration

on navigating policy with data to to

leave no one behind. So there's history

to to that and with that foundation we

car we carry out the following

activities

um back in 2025. Well, first of all, uh

in collaboration with uh UNSD, ECA and

the World Bank, we published the uh

how-to guide on um geospatial uh SAPE in

all. Um it's a practical guide with uh

runnable codes and real data uh for

practitioners and policy makers alike

interested in uh geospatial uh

uh the guide is available in both HTML

and GitHub versions as you can um should

be able to see in a bit. Uh I'll share

my screen to uh sort of give you a quick

demo of the guy. Um so this is the HTML

version. So you can see it's very

interactive. Uh we also prepare uh a

short sort of demo for you as well if

you would like to know how to navigate

um the guy. I won't play it because we

we're running out of time. Uh but do but

do please uh you know watch it and and

you know navigate the the guy um you

know in in your spare time. I'll share

the link um with you in the chat as

well. Uh there's sub uh components to

each chapter that you can go through. um

you know there's uh chapters on setting

up all as Angela mentioned some

participants in the workshop that I'm

going to tell you about later on um had

you know uh trouble uh getting used to

working on all and we've uh incorporated

that in here as well um but um overall

this guy is is uh essentially a

practical walk through where uh if

you're interested in in geospatial essay

E um you can just simply go through the

chapters um and you'll be able to uh

you'll be able to um you know not only

understand the importance and the

usability of geospatial uh SAE but

you'll also be able to uh run the codes

and um you know uh follow along the

examples and eventually um use uh all

this knowledge and codes uh for your own

uh country context or indicator of of

interest. Um, so let me get back to the

uh slide,

right? Um,

yes. So, um, that's that's the uh how-to

guide. Uh, and I highly encourage

everyone to visit it. Uh, the link will

be shared with you in the chat later on.

Uh second with support from UNSD and the

World Bank we organized um our 2025

Asia-Pacific capacity capacity

development program on SE uh and we uh

combine virtual e-learning uh

facilitated by um you know an expert

from the World Bank and uh with

in-person uh regional workshop in

Bangkok on geospatial SAE which I will

uh briefly tell you uh later on as Well,

so the guided e session started um on

the 2nd of October um and it concluded

on the 14th of January because um it it

includes seven weeks of facilitated

online classes um and participant went

through um you know uh sessions where we

have um the expert uh uh conducted um uh

tutorials uh essentially uh through the

e-learning

uh course uh that we developed and links

will be shared again uh in the chat box

uh on the Cup um the Cup platform. uh

CIP is our stat statistical institute

for Asia-Pacific. Um and through that

link uh participants went through uh all

of the uh materials. Um and all of the

participants uh able to complete uh the

course including the graded assignments

uh and they were able to um uh attend

the workshop which was similar to what

Angela mentioned was the main

requirement for participants to uh

participate in the in-person workshops.

Um there were 20 of them for the 2025

cohorts and from uh 10 uh countries. So

two per countries. Uh and let me get the

list for you. They're from Bangladesh,

India, Indonesia, Malaysia, Pakistan,

Philippines, Sri Lanka, Tajikistan,

Thailand, and Vietnam. We're lucky to

have uh one participant from Malaysia,

Miss Fisa uh here with us who will be

able to share her experience later on.

Um but yeah so this this uh course um

enable participants to be able to

prepare themsel for the inerson workshop

and contribute um as as much as they

can. Um speaking of the workshop uh it

happened on from the 24th to the 28th of

November. Um and uh it was attended by

all uh 28 participants. The focus of the

workshop was on providing hands-on

capacity uh building support uh to NSO

staff essentially on using uh or a

little bit Python to conduct geospatial

SAPE for their indicators of interest.

Uh we had um Dr. Josh Murfield some of

you might have uh worked or known him.

uh he was our facilitator and he taught

participants on uh working with shape

files, rusted data uh packages uh like

the one that um David mentioned earlier

uh geol uh titer puff map uh and

estimating models uh such as obviously

you know pop uh mentioned earlier and

then uh uh it's gradient uh boosting uh

as well um participant did bring uh

their own well many of them brought

their own data sources uh and they were

able to estimate their indicator of

interest using the um the skills uh

learned from the workshop. Um uh if

you're interested you can click on the

link or I will share in the chat box

again um to learn more about uh the

workshop itself. Um and the link at the

bottom there is another link uh in

reference to the the how-to guy uh

mentioned earlier. Last but not least,

uh we wanted to extend our capacity

building initiatives even further. Uh so

we organized our Asia-Pacific STA cafe

on geospatial SAPE on the 27th uh of

January 2026. Um, we invited

participants from our workshop and

e-learning program to share their

experiences uh and lessons learned in a

panel uh with expert reflections from

Josh and a re resource person uh from uh

BPS uh Indonesia.

Uh we also took the opportunity to

showcase the how-to guide again on

geospatial sea. As you can see, we're

very very proud of that uh with with uh

in collaboration with a colleague from

ECA uh USD and World Bank. Um and we've

got uh quite a number of participant uh

registered 60 plus uh and uh uh the uh

the panel discussion was lively and then

everyone was was very engaged uh and

more information on that can be found

again in the link that will be shared uh

in the chat uh later on. So that's all

from ASCAP in 2025 and look we look

forward to uh to uh 2026 and and what uh

the year will bring. Um thanks again for

the opportunity Daniel. Uh back to you.

>> Thank thank you Ro. I mean it's it's

really good to see uh what's happening

in in the African region and in the Asia

Pacific and uh I'm probably um

participants have picked up this is

happening uh for people who are

motivated because it's not it's not uh

uh other kind of courses you just go for

a week. Uh we also have a prerequisite

of people finishing the 7week module

that really uh puts uh the groundwork

ready for uh the 5-day workshop physical

workshop at the end. Uh we will be

sharing all this uh information later on

all the right links. Um I see a couple

of questions coming which which link

where it where the e-learning courses

and things like that they will be shared

and the recording of this meeting along

with the presentation will be shared

with all the participants who registered

here. So now with with all the regional

commission's backing and the UNSD and

our partners and in the world bank

countries have been uh participating in

this uh uh taking the course on online

and then also doing the physical uh

course for a week. uh it it is time now

to switch to countries and see what's

happening from uh their uh their side of

uh the the the the work on SAE. Without

further ado, I will pass this to our

colleague from Chile from the National

Institute of Statistics in Chile who

will be showcasing their uh SAPE

experience and they've been doing a lot

of experimental statistics using SAE and

we'll be uh yeah following up our

presentation. Miss Aier will be

presenting uh uh the activities in in

NA. U the floor is yours Mr.

Hello. Uh so well good morning everyone.

Thank you for the invitation. C can you

see my presentation?

>> We saw it and then it's gone. Can you

reshare again?

>> Oh

me

put it on presentation mode and I think

it's not working.

>> Okay.

>> Huh?

>> Yeah we can see it. Okay,

>> thank you.

>> Okay, thank you.

So, well, my name is Javier Torres. I'm

from Chile and we're going to show you

the results of the implementation of

small area estimations uh in the

national victimization survey.

So a little context uh our victimization

survey is called national

and it's one of the longest

victimization surveys in Latin America.

It's collected annually since 2005

and it has a sample of about 24,000

household providing national and

regional representativeness.

So in 2023 we have a major redesign of

the survey. We redesigned the

questionnaire and the sampling and with

that uh we collected the first survey

with a communal representativeness for

the European areas of 136 communes. Uh

this came from a growing demand from our

government for better geographical

disagregations given the characteristics

of the phenomenon. Uh however uh having

uh coming out a survey every year it's

uh expensive and and it's not

sustainable on the long run. So for 2024

we use SAI to obtain reliable estimates

of the proportion of households

victimized by violent crimes which is

the main estimate of the survey.

uh this for the 136 communes of the

design and the results were just

published we published in January and

you can see it on our website.

So uh regarding SAI we started working

back in 20 uh 2018 actually with the

first uh capacity of building phase uh

this came from assistance from EKLAC. So

we started working in 2018 with the

survey uh from that moment which is like

the old version of the survey and we

work with the proportion of households

victimized by high social impact crimes.

Uh this uh was a working paper that it

was published in 2024 and it gave us uh

a lot of lessons mostly that uh it was

needed to have a like a strong

theoretical framework and establish a

strong criteria for evaluating the

results and that these were consistent

with the phenomenon.

So uh well from 2018 to 2022 we mostly

work in the capacity building in NSO and

then for 2024 we establish uh SAI

estimations as official estimations of

the survey. So uh in 2023 we had an

exercise with the communal uh survey uh

which allow us to understand better the

coariates and the needs of the model.

So for uh 2024 as I told you we use uh

the survey we use SAI as an official uh

estimates or mostly it was uh

so we use uh uh some methodological

framework we use aid uh proposed

framework uh for the specific

specification phase

uh we evaluate the user needs uh the

data availability and the SA AI methods

available. So for the user needs, we

defined that our estimator was going to

be the proportion of households

victimized by violent crimes which is um

oh here it's a aggregate of seven

different crimes such as robbery or

assault. And for the data ability uh we

put efforts on creating a theoretical

framework which guided uh the search for

the coariate. So we establish some

dimensions like you know socioeconomic

or socio demographic characteristics but

we also look for data regarding uh crime

and victimization like police records

and infrastructure and environment such

as satellite images, national uh

community information and

and the system and informal settlements

data.

Uh this uh this phase we also chose our

target of estimation which is commune.

Uh commune is the smallest

administrative submission in Chile and

it's equivalent to a municipality. And

finally we decided on using EVAP uh is

based on the ferret model as it has been

applied to poverty estimations in Chile

and is quite consolidated as a

methodology.

Uh after that uh we went on the analysis

and adaptation phase. uh so to reduce

the volatility associated with the

estimates uh the sampling variance is

modeling using a generalized variance

function and this gave us a more stable

and robust measure of variance which is

subsequently used as an input for the

model. We also establish a domain

inclusion criteria which give us uh how

many communes will have a purely

synthetic estimation or which ones was

going to be uh direct and synthe

synthetic.

So um part of this criteria were the

degrees of freedom the number of

observations and the sign effect for

each domain and then we search uh for

our final model.

So uh for our model we have a model

selection algorithm algorithm uh

statistical criteria and conceptual

validation. I'm going to explain that

very shortly each one of them.

So uh we use a baseline model uh

including regional dummy variables. Uh

we use a stepwise selection and we did

the exploration of uh combinational

variable subsets. So we uh throw all the

coariates that we have and try to

simulate different models.

Uh we also well we look for the

statistical significance of the core

variates the AIC and the BIC. Uh we also

look for the residual diagnostics and

multicolinearity

checks and the prevent benchmarking

check. And finally for the conceptual

validation we selected coariates that

are expected to cover keymatic

dimensions related to victimization.

Uh so uh

for the evaluation finally after we have

our model so we evaluate its precision

and certainty and to verify the

assumption of linearity we check the

residuals looking that there's no

influal domains. Uh we also look for

consistency with the regional estimates

looking the SAI fall within the

confidence intervals of the direct

regional estimates.

uh in this case uh region is the first

level administrative division in Chile

and the commun are part of the region.

So we expected that you know no uh

domain were uh above the um confidence

intervals.

Uh we also look for error measures. Uh

particularly we compare the root mean

square uh error of the SAI within the

direct direct estimations of the survey.

And this show us uh this is the the

graphic we have here and this show us

that um the SAI is significantly more

efficient than direct estimations

especially in areas with a smaller

sample size sizes. Uh so in the figure

we also sort the communions by sample

size uh which is highly associated with

the reli reliability of the estimation

according to our NSO standards.

So um we have the point estimations this

is for 2024

uh where we can see the elab uh elap

estimations tend to be more conservative

than direct estimations. uh in blue uh

we can see the draw direct estimations

uh in the in the areas with the smaller

sample sizes also as we expected uh for

the communes with bigger sample sizes uh

both estimator estimators are quite

similar.

So this is sorted by sample size. Uh

this is smallest and the uh biggest on

the right.

Uh so uh well this is has been a long

work for us. We have been working on

this for the past six years and there is

a lot of lessons that we have taken from

the different exercises that we have

made. So uh first the quality of uh SAI

estimates depends uh strongly on the

relevance the coverage and consistency

of the auxiliary variables which are not

easy to find. uh regarding the

limitations of the administrative

records use we identified in some cases

coverage limitations. So we worked with

uh 136 communes from around 300 that

there is in Chile and it was really hard

to find data that covered the

136.

uh we also find missing data in the

administrative records and

inconsistencies which uh made it

necessary to perform some imputations or

in some cases to reject the use of the

data.

Uh also uh to use an aggregated

victimization indicator can facilitate

the explanation of the phenomena but

introduces uh challenges in generating

predictive models. As I told you, this

was seven uh different crimes with seven

uh different characteristics. So, it's

it's harder to find uh auxiliary

variables that can

work with the seven of these crimes.

Also, uh models with high predictive

performance are not always easily

communicable. So, for public policy

context, it is essential to balance

statistical precision with

interpretability and transparency.

uh for us also incorporates

incorporating the SAI requires a

safeguarding compatibility with direct

estimates and across periods which uh is

related to

to the last point. Uh the annual

frequency of our survey uh requires

iterating and adjusting and

progressively validating the models

strengthening their institutional use

over time and requires uh distinguish

distinguishing stages of methodological

learning, model testing and

implementation which do not always align

with the timelines of foring results. So

we have a very short timeline here. We

uh gathered the data in the last

trimester of the year. we publish during

the second trimester of the year. So, uh

actually being able to produce SAI

estimations over uh each year, it's uh a

serious challenge for us

and for the future as we already have

our first version of the SAI work. Uh we

are looking to expand uh to the use to

other indicators of interest. for

example uh the dark figure of crime or

the perception of insecurity. So it is

important for us to understand that the

results both on the quality of the

auxiliary information sources sources

and on aspects related to direct

estimations.

Uh we also well as the limitations of

the administrative records uh we

identified in some cases uh you know

limitations as as a missing data. So we

need to we need to search uh we we need

to search data every year for one uh

data that working for 2024 might not be

updated for 20 and 25 and so on. So it's

a constant work

and

finally uh as although it is possible to

make uh reliable uh commute level

estimates for the 20 and 24 period uh

for us uh

uh it's it has been a difficulty that we

had that as we produce uh direct

estimations for 2023

and scientific estimations for 2024. We

cannot rely on the usual statistical

test to compare the estimates with the

previous period and this is going to

happen to us again in 2025 because we

have direct estimations again. So this

is something that we're uh trying to

research uh how to

um how to use this data and avoid uh the

comparations between the two different

methods. So that's kind of what we are

doing right now. Uh thank you

>> many thanks Navier for an excellent

presentation and also showcasing you

know the the this the history of ACE in

in in the National Institute of Chile.

It it it shows clearly you know the

maturity of the work you you're doing.

Uh I will come back for the Q&A later on

but I will pass now to uh the other

continent

uh to Africa and uh Edward from Ghana

sical service will be presenting their

their experience in SAE uh will come

back uh to the questions later on. Uh

Edward the floor is yours.

>> Okay. Thank you. Um good afternoon once

again from Ghana.

Um I'll be presenting on Ghana's

experience using SAE and then what we

have been doing so far so date

screen is back.

>> Okay. Please can you see my screen?

>> Yes. Yes, we do.

>> Okay. Thank you. So

is it rolling or is still the same page?

Okay, sure. So this is going to be the

outline for the presentation. And I have

the introduction and then why we are

into SAPE the entry point where we

started from the capacity building and

training from training to practice what

we have been doing and then what we have

done so far and then what we are doing

as GSS when it comes to SAPE to instit

institutionalize it and then challenges

encountered so far um and then how we

are moving forward and then looking at

what we will do in the future.

So in Ghana we have 16 regions that's

admin 2 regions and then we have 261

municipal metropolitan industrial

assemblies that's for the admin 3 mostly

our um surveys are actually at the admin

level but the demand for data is always

coming in from the admin level because

of some of the local government policy

that they have over there. So what we

adapted is also part of this SAPE

actually we have to estimate for them

and mostly during the past we couldn't

because we didn't have any idea of these

estimates and then nobody will give them

the original and then try to make

assumptions around how the districts

will be like but for now that we have

been trained in SAE from

um

World Bank USD UNA ECA UNFPA and in

other regional commission

we have some ideas of how we can do

these estimations using the SA methods

that we have been trained on. I

particularly was part of the just recent

um ended training that we had in Kenya.

I was present there and then with Daniel

and then the team from World Bank.

So we have we have been through this

training from the past. I remember my

colleagues also were there some time ago

and then I also came and then we are

also training other people in the

office. What we do is that in most work

that we take under SAPE we try to

include a particular person or two to

also be part of the work so that they

also get some of the experiences

in SAPE so that we all work together and

then also get more people with some

skills in SAE to help us work

from that. We the Ghana saskar service

has been able to publish 15 reports

using SAPE and then these reports were

from the Ghana demographic and health

survey um data and then the population

housing census data and then it's it

ranges from the um exclusive

breastfeeding childhood immunization

women's empowerment gender based

violence excessive alcohol consumption

double burden of malnutrition we even

have um um excessive alcohol content and

all that we have a lot they or 15

reports that has been published together

with the UNFPA and the USA DHS. So in

this report we try to use different

methods. We use the ELE methods, we use

the EBP, we use a ferot the the the

reason why we chose a particular method

to use sometimes revolve around how we

interpret the results and then the

assumptions the estimate methods to the

policy makers. So that also embrace the

results that they are seeing. And

sometimes we test these methods to see

which one is giving us the best

estimate. Not just statistical

assumptions, but we look at what we know

from these districts, the data we have

from the past and what the estimates are

giving us. which one is closest to being

true or being real regarding the numbers

we are having and then sometimes we test

the assumptions like the UNFPA method

that we are seeing the ELN normally we

do logistic regression on these

estimates

um on these data sets and then we test

the ROC's that's the area under the

curve assumptions we test a lot of

things to see which one is performing

better before we choose the model to

And then we also trying as Ghana service

to like I said institutionalize

the G um the SAE method in Ghana. So we

release a report not just testing these

estimates or testing these methodologies

but we release the report to policy

makers to also use to the district level

local government policy makers to also

use sometimes to we invite them to these

publications so that they get sense of

whatever we are doing for them. the

numbers we are given to them, how they

were generated, how the estimates came

about so that they know what we are also

going to use the numbers for, what even

went into them and there are still

processes in place to make sure that SAE

methods become core part of whatever we

do in GSS most are incorporating it

apart from the um DHS report that we

have done currently as I'm sitting here

we are also working on the NPI reports

that were released um some months ago I

think last

So we are doing that was also at

regional level and it's like a trend

analysis from 2022 to 2025. We are also

trying to run SAE for all these years

for all the districts that's the 261

districts that we have in Ghana and it's

not just the headcount of poverty but we

also doing it at the indicator level and

also even running the intensity of

poverty for all these districts. So with

all these models there are different

different methods that we are trying to

use. Currently we are done with the hair

counts which was done using the ELLL the

intensity is um I've done for 2023 and

then we have we tested the EBP ferot and

then the base approach the reason is

that we can't estimate intensity of

multi-dimensional poverty for each

household but it has to be an area level

indicator so we need an area level model

to do this so that's why we are trying

to use the ferot the EBP P and then the

base that's the basian approach and then

the current one that we have that has

been accepted is based on the EBP so we

are trying to use the EBP to estimate

intensity for all the years also so GSS

is currently doing a lot and I won't say

we haven't faced some challenges we have

and then that's also the reason why

these are also based on purely sensors

data sets we haven't actually included

um

special data sets yet because sometimes

it comes with data harmonization between

the survey sensors and even the

geospatial data sets and then also

communicating uncertainty and model

based estimates to nontechnical users

are also a challenge sometimes. So we

try to find a way of interpreting them

in simpler terms or in simple languages

for them to understand to also cherish

what we are doing at GSS and then also

choosing the right auxiliary variables

also becomes difficult sometimes. I mean

it's like the NPI that we are doing

sometime we have to also make sure that

the indicators that we used to

estimating the head count poverty

doesn't also end up in the um the

auxiliary variables that we are using

the predators we are using to estimate

poverty so these are all difficult

sometimes you need to bring in other

external data set but because of how you

have to harmonize and then you are not

getting the right variables to me data

sets choosing the right auxiliary

variables become a challenge but well we

are doing our best and Then also the

right model like a lot of challenges

goes through when choosing the right

model. And then we are facing capacity

constraint when it comes to we were let

me put that way facing capacity

constraints when it comes to estimating

models but we try to put a member who

have not had a skill in SAE in the team

so that the person gets something in the

tips when it comes to SAE. So the next

time we can also rely on this person to

help us in SAPE models

and then what help us move forward I've

mentioned one we put members in the team

so that they can do it and then we also

learn from each other so we communicate

sometime we bring people from external

to also discuss what we are doing so

that they also give us their point of

views and then we put them together and

then we also had great leadership from

the part of the GSS management where

they also accept that okay this is not

our traditional way of doing things.

This is something new

but we have accepted it and we are also

trying to put them forward and then we

give them also thanks and then looking

ahead. So currently there are plans for

the DSS data science team to also look

at external data sources and how we can

incorporate that is also solving our

challenges when it comes to determin

harmonization and then I mentioned the

work of poverty multi-dimensional

poverty that we are working on there's

labor statuses coming we are working on

the Ghana living standard survey

currently we also working on the

consumption and the poverty line

estimates which when it's done we are

also going to run estimates for the

district levels because that one is also

representative at the regional level and

then the data science team is working on

um core detail records and then I am not

part of that team so I don't really know

what is in the data set but they are

working on how they can get some

variables to help us estimate these um

um models at the district levels

something that will distinguish one

district from another not just based on

the human beings living in it but

something that is peculiar to a

particular district from the data set so

That is also work that we are also doing

at the Ghana statistical service. Thank

you.

>> Thank you Edward for uh going through

Ghana's experience. I mean it's really

impressive to see uh all this reporting

is using all this SAPE um modeling uh

which which seems to be a little bit of

a challenge for many NSOs but uh going

through these examples and answering the

questions at the you know local level

that's where uh decisions are made where

the impact is much more important uh

that's that's really uh great to to see

Yeah, time is um flying and uh we are

trying to get uh everybody in and we'll

probably add extra minutes at the end

for more Q&A. Uh so without further ado,

I'm I'm now switching to another

continent going to Asia Pacific. We have

our colleague FIA from the Department of

Statistics Malaysia. they will be she

will be presenting the experience of SAE

in their uh uh in the in their uh

department uh FISA the floor is yours

please go ahead

>> okay thank you Mr. David, let me share

my slide first.

Can you see my slide?

>> It's coming.

>> Yes, I can see your slide.

>> Okay. Assalamu alayikum and hello

everyone. I am Faizar Rosanti Taj Aaros

a statistician from department of

statistic Malaysia. Today I will share

my experience attending the SAE capacity

development program under ESCAP UNSD and

World Bank with our team use SAE for

estimate income and poverty at state

legislative assembly in Kang Joho.

This is the geography of Na and the

arrow show the location of the study

which is Kangjoo.

Below are the four maps clearly show the

different boundaries for each

administrative level in Kuang Joho. The

first map is our administrative district

of Kuwang as admin level two. Then the

district is divided into three

parliament as admin level three. Each

parliament actually consists of two

SLAs's with a total of six SLA in admin

4 and we have 11 census district at

admin level five. For this study we use

three set of data. First the population

and housing census 2020 the data up to

enumeration block level in shape file

format. Then we have household income

and expenditure survey in 2022 with the

data up to living quarters level in CSV

format. And we have satellite image data

2024 in format CSV and TIF

by using our census shape file data.

Then we generate the population map

based on the enumeration block. From

this map we can see that the most

population in Kwang administrative

district are in parliament of Kuwang.

Then we create a new shape file by

combine the census and survey data and

the second map showing show showing the

distribution of average household income

across the Kang area

by using the data from Google Earth

Engine. We also plot a point of

nighttime light by using the point of

coordinate and we change the point at

the center into a polygon. This map

display night time light intensity for

clu.

Then we use satellite image to create a

simple map for showing how bright each

area is at night. From the nighttime

light map, we can conclude that the most

population area also the most bright

area in this satellite image data. The

polygon then are colored based on the

average nighttime light and the area

with no data are shown in light gray in

second map. Then we do some treatment

for the EBS with no satellite data. We

replace the missing data with the global

mean.

After finished with the data cleaning,

we managed to use the guideline to SAPE

for poverty mapping published by our

bank to identify which method should be

used for this study. From the decision

tree on method availability, we choose

to use area level model base. This is

because the census and survey data are

not conducted in the same time frame.

the census in 2020 and the data were

received at EBIS level while the survey

data was from 2022 and the data at LQ's

level. So for this study we choose to

use the fair Harriet model. Then we use

LEO and GLM net from package in R for

the variable selection and transform the

predictors to improve the fit of the

model. The lambda value was used for

extracting the variable with nonzero

coeffic coefficient.

For this study, we use income

indicators, poverty line information

and nighttime light from satellite image

as auxiliary information to help

stabilize the estimate for this area. We

cannot use direct survey alone because

of the sampling might be used only a few

household. So the numbers will become

noisy, unstable and sometime misleading.

We use the fairheaded model to improve

the estimate by combine two sources of

the information which is what the survey

tell us and what we know from other data

from this small area. If an area has

strong survey data, the model relies

mainly on the survey. But if the survey

data is weak, the model borrows strength

from the auxiliary information. For this

model, we group household by EBS and

calculate the direct survey poverty rate

for each EB. Then the model improve this

estimate using the auxiliary data

without any transformation.

The fair model without any

transformation can be uneven or noisy.

From the chart, we can see that the

relationship between the direct survey

and the model is less clear and the

results scatter among the point. The

brown test result show that the

correlation between the model predicted

and the direct survey estimate is 0.68.

This correlation result show that the

auxiliary data we use is informative and

support the model in producing better

and more stable poverty estimate.

Next we try fitting the fair model with

log transformation from empty package.

We apply a log transform to stabilize

the modeling scale. After do the

transformation, we convert result back

to the original poverty rate scale using

a bias corrected model or BC_SM.

So they are interpretable. The output

now include improved estimate for every

EB plus their uncertainty or MSE with

bias corrected back transformation to

the original scale. As a result, the

relationship become clearer, the data

points are more orderly and the model

become more consistent.

The brand test result show that the

correlation between the models with

transformation and the direct survey

estimate is more better from 0.68 to

0.79.

The main model accuracy indicator here

are the MSE for how confident we are

with the estimate.

In summary, the fair model give more

stable and reliable poverty estimate for

each small area. This model improve the

direct survey result by reducing noise

and borrow strength from auxiliary

information to make the estimate more

accurate. Limitation and way forward for

this study. For improving the data

quity, we need stronger administrative

and geospatial data such as housing

condition, land use, school attendance,

clinic visit or we can also incorporate

with other data source such as mobile

phone data, digital primary records and

additional satellite indicators to

strengthen the model. Beside that, all

data set must include geographic

boundaries information so the coordinate

more accurate and the model become more

perfect. And last but not least to have

better partnership by collaboration with

local authorities, welfare department,

utility providers and other facilities

provider is essential to access micro

data and these data sources can

significantly enhance the accuracy of

small area poverty estimate. That's all

from me. Thank you very much for your

attention.

>> Thank you Fisa. That's it's impressive.

Even I was thinking I'm going into the

course myself. So it's uh it's really

refreshing to see you know the the the

sweat that goes into uh running all this

uh modelings. Um as as they say in

Malaysia statics plume in harmony that's

uh that's the motto. So uh we've now

come to the conclusion of the

presentations. We have uh an extra 5 to

10 minutes for Q&A. Uh I have seen uh uh

most of the uh uh questions are being

answered in the chat very much thanks to

David and Howie Adit. But I will

probably um open one question for all

the three country presenters from for

Edward Fisa and uh Javier. This question

is about uh I think uh you've alluded it

to your uh presentation

question from uh in the Q&A we have a

question from uh Sanchez. uh he's asking

about uh if you can say a few words

about how policy makers and government

is accepting USA results uh when you

present it. So what are uh what do you

consider what are the things uh that uh

is you know important to communicate

this to the policy makers. If you have

some experience if you can share with

that with us that would be good

and you can uh you can jump in and

answering FA Edward or Mavia V.

>> Okay. Thank you. Um for Ghana I know

that this um estimate that we are

producing is a great deal for policy

makers especially when we go to the

local government they use some of the

estimate that we give them in their um

district common fund estimates. So when

they are preparing their um district

common fund um formula that's how they

call it here they use some of these

estimates to guide them in that. So yes,

it's a great deal for them. When we

produce them, they come, they sit in and

then we release the numbers to them.

Yes. So I know they are using it because

that's one of the things that they do.

And I think the government also Yeah.

Because I've seen the um Ministry of

Sanitation um some of the other relying

on these numbers that we put out for the

districts

to do some of the policies. I know they

are revising some of the documents

because of some of the estimates that we

put out there. Yeah.

Thank you Edward. uh any any input from

uh FISA or miss Javier?

Okay, for Malaysia we are from Malaysia

we are currently serious looking for the

small area statistic. We start publish

our small area um like parliament and

SLA statistic and also in our current

strategic plan for the department. We

also include the small area estimation

for our future statistic. That's all

from me.

>> Thank you. Thank you. And now to Chile.

Yes. So promise uh well uh poverty

estimations with SAI models are already

in place. So we we do not produce those

models though. But uh they have been

working since around two or three years

maybe a little bit more. And for us with

victimization it has been um

uh it has been a lot of work uh to uh

consider these estimations. As we told

you, we started working in 2018 and just

by 2024, we were allowed to produce this

uh estimations within the um the

official results of the survey. So uh

and it's still uh we still publish as uh

experimental uh statistics. So it's

still not like official official. So but

we are working on it. So we respecting

after two or three iterations uh we

might publish the result as official 66.

>> Thank you. Thank you very much. I I'll

just switch back to for extra few

minutes back to David. I've seen a

number of questions answered there.

Maybe if you can uh give us a quick

highlight of uh what if you want to

broadcast this to the the whole group on

some of the questions. most of them are

on technicalities and type of models

used but I think it would be good to uh

also um I've seen a few questions

requesting support uh for uh doing some

of the work. So I think if in light of

uh uh the support that would be uh

available in the coming year if you

could uh briefly uh present what we're

uh going to do from our side uh in terms

of supporting countries. Yep. Sure.

So, thank you Daniel. I think for

supporting countries um usually the