Measuring individual poverty with OLS and off-the-shelf household - - PowerPoint PPT Presentation

Measuring individual poverty with OLS and off-the-shelf household budget surveys

Valérie Lechene UCL and IFS, Krishna Pendakur SFU, Alex Wolf IFS

KCP Project: Intra-Household Allocation of and Gender Differences in Consumption Poverty, with Theophiline Bose-Ducker, Isis Gaddis and Talip Kilic - Dissemination Workshop

Nov 5. 2019, World Bank HQ, Washington DC

What we do and why:

We develop a method to measure the share of household consumption that go to each individual in a household. Individual resource shares are not directly observable The method is simple, theory consistent, easy to implement. Poverty is experienced at the individual level, as there might exist inequality within households. Any interest in poverty requires its measurement. Effective and cost effective policy designed to alleviate poverty should target individuals.

Dream data

Table 1a: Dream Data Man Woman Child Total Food 400 300 200 900 Clothing 50 75 25 150 Shelter 200 200 200 600 Transport 125 63 63 250 Other 45 37 37 120 Total 820 675 525 2020 As a share of the household total expenditure, the amounts in green are resource shares. Poverty line: $1.90 per day, ie $694 per year per person. The woman and the child are poor but the man is not. Dream data: Cherchye, Demuynck, De Rock and Vermeulen (2017); Brown, Ravallion and van de Walle (2019); Bargain, Lacroix and Tiberti (2019)

Dream data

Table 1a: Dream Data Man Woman Child Total Food 400 300 200 900 Clothing 50 75 25 150 Shelter 200 200 200 600 Transport 125 63 63 250 Other 45 37 37 120 Total 820 675 525 2020 As a share of the household total expenditure, the amounts in green are resource shares. Poverty line: $1.90 per day, ie $694 per year per person. The woman and the child are poor but the man is not. Dream data: Cherchye, Demuynck, De Rock and Vermeulen (2017); Brown, Ravallion and van de Walle (2019); Bargain, Lacroix and Tiberti (2019)

Dream data

Table 1a: Dream Data Man Woman Child Total Food 400 300 200 900 Clothing 50 75 25 150 Shelter 200 200 200 600 Transport 125 63 63 250 Other 45 37 37 120 Total 820 675 525 2020 As a share of the household total expenditure, the amounts in green are resource shares. Poverty line: $1.90 per day, ie $694 per year per person. The woman and the child are poor but the man is not. Dream data: Cherchye, Demuynck, De Rock and Vermeulen (2017); Brown, Ravallion and van de Walle (2019); Bargain, Lacroix and Tiberti (2019)

Dream data

Table 1a: Dream Data Man Woman Child Total Food 400 300 200 900 Clothing 50 75 25 150 Shelter 200 200 200 600 Transport 125 63 63 250 Other 45 37 37 120 Total 820 675 525 2020 As a share of the household total expenditure, the amounts in green are resource shares. Poverty line: $1.90 per day, ie $694 per year per person. The woman and the child are poor but the man is not. Dream data: Cherchye, Demuynck, De Rock and Vermeulen (2017); Brown, Ravallion and van de Walle (2019); Bargain, Lacroix and Tiberti (2019)

Real Data

Real Data Man Woman Child Total Food 900 Clothing 50 75 25 150 Shelter 600 Transport 250 Other 120 Total 2020 Poverty line: still $1.90 per person per day, or $694 per person per year, or $2080 for this household per year. The household is poor, hence we conclude that all members of the household are poor.

Real Data

Real Data Man Woman Child Total Food 900 Clothing 50 75 25 150 Shelter 600 Transport 250 Other 120 Total 2020 Poverty line: still $1.90 per person per day, or $694 per person per year, or $2080 for this household per year. The household is poor, hence we conclude that all members of the household are poor.

Real Data

Real Data Man Woman Child Total Food 900 Clothing 50 75 25 150 Shelter 600 Transport 250 Other 120 Total 2020 Poverty line: still $1.90 per person per day, or $694 per person per year, or $2080 for this household per year. The household is poor, hence we conclude that all members of the household are poor.

Dream data vs real data

Real Data Man Woman Child Total Food 400 300 200 900 Clothing 50 75 25 150 Shelter 200 200 200 600 Transport 125 63 63 250 Other 45 37 37 120 Total 820 675 525 2020 With the same poverty line: With dream data, we observe that there is intra-household inequality, and we conclude that the woman and the child are poor, while the man is not poor. With real data, we assume that there is no intra-household inequality, and we conclude that the man, the woman and the child are poor. With real data, we need to fill in the blanks with a model.

What has been done before

In practice:

◮ Measure consumption at the household level, divide by the number of

household members and compare against a per-capita poverty line (1.90$ per person per day)

In theory:

◮ Chiappori (1988, 1991, ...) Collective model ◮ Browning, Chiappori and Lewbel (BCL, 2013) ◮ Dunbar, Lewbel and Pendakur (DLP, 2013)

Practice very far from theoretical results:

◮ BCL is demanding in terms of data and hard to implement. ◮ DLP tried to make it easier in both dimensions, but still not trivial to

estimate.

What we do

We develop a linear method, L-DLP, to estimate individual resource shares as in DLP:

◮ Theory consistent ◮ Allows for multiple men and women ◮ Recover structural parameters from OLS estimation

We develop a pre-test to check whether the methods will work. We implement our method on the LSMS data. We find that there is gender inequality inside households in some places.

Notations and definitions (BCL)

h = 1,...,H households, containing individuals of types t. A household consists of Nt

h individuals of each type t; let Nh = ∑t Nt h

yh denotes the observed household budget. Market price vector p, shadow price vector Ap where diagonal A ≤ 1. A characterises scale economies. We allow for them, but don’t estimate them. Efficient collective households: collections of people (with utility functions) living together, who reach the Pareto frontier. Decentralisation: The household problem reduces to a set of individual problems.

◮ It is as if each type gets a shadow budget; denoted

yt

h, spent at prices

Ap.

◮ The household purchases enough to satisfy everyone.

Resource share, denoted ηt

h, is the fraction of the household budget

going to type t.

◮ ∑t ηt

h = 1.

◮ a type’s shadow budget is

yt

h = ηt hyh; a person’s shadow budget is

ηt

hyh/Nt h.

Notations and definitions (BCL)

h = 1,...,H households, containing individuals of types t. A household consists of Nt

h individuals of each type t; let Nh = ∑t Nt h

yh denotes the observed household budget. Market price vector p, shadow price vector Ap where diagonal A ≤ 1. A characterises scale economies. We allow for them, but don’t estimate them. Efficient collective households: collections of people (with utility functions) living together, who reach the Pareto frontier. Decentralisation: The household problem reduces to a set of individual problems.

◮ It is as if each type gets a shadow budget; denoted

yt

h, spent at prices

Ap.

◮ The household purchases enough to satisfy everyone.

Resource share, denoted ηt

h, is the fraction of the household budget

going to type t.

◮ ∑t ηt

h = 1.

◮ a type’s shadow budget is

yt

h = ηt hyh; a person’s shadow budget is

ηt

hyh/Nt h.

Notations and definitions (BCL)

h = 1,...,H households, containing individuals of types t. A household consists of Nt

h individuals of each type t; let Nh = ∑t Nt h

yh denotes the observed household budget. Market price vector p, shadow price vector Ap where diagonal A ≤ 1. A characterises scale economies. We allow for them, but don’t estimate them. Efficient collective households: collections of people (with utility functions) living together, who reach the Pareto frontier. Decentralisation: The household problem reduces to a set of individual problems.

◮ It is as if each type gets a shadow budget; denoted

yt

h, spent at prices

Ap.

◮ The household purchases enough to satisfy everyone.

Resource share, denoted ηt

h, is the fraction of the household budget

going to type t.

◮ ∑t ηt

h = 1.

◮ a type’s shadow budget is

yt

h = ηt hyh; a person’s shadow budget is

ηt

hyh/Nt h.

Assignable goods

Assignable goods: Private goods consumed only by one type of individual in the household. E.g., clothing. W t(y) denotes the observed household Engel curve function for type t’s assignable good: the fraction of household expenditure commanded by that good at prices p and household budget y. wt( y t) denotes the unobserved individual Engel curve function for type t: the fraction of expenditure commanded by the assignable good for type t at shadow prices Ap and shadow budget y t. Key to identification in BCL and DLP and here: The household-level Engel curves at market price are equal to the product of the resource share of type t by the Engel curve at shadow price and shadow budget: W t(yh) = ηt (yh)wt y t

h

Assignable goods

Assignable goods: Private goods consumed only by one type of individual in the household. E.g., clothing. W t(y) denotes the observed household Engel curve function for type t’s assignable good: the fraction of household expenditure commanded by that good at prices p and household budget y. wt( y t) denotes the unobserved individual Engel curve function for type t: the fraction of expenditure commanded by the assignable good for type t at shadow prices Ap and shadow budget y t. Key to identification in BCL and DLP and here: The household-level Engel curves at market price are equal to the product of the resource share of type t by the Engel curve at shadow price and shadow budget: W t(yh) = ηt (yh)wt y t

h

Dunbar, Lewbel and Pendakur

DLP impose sufficient conditions to identify resource shares with just Engel curves for assignable goods within households.

◮ resource shares do not depend on the household budget so that

ηt(y) = ηt

◮ individual Engel curve functions are given by the Almost Ideal demand

system of Deaton and Muellbauer (1980): w t(y) = αt + β t lny

◮ preferences are similar—but not identical—across people, such that

β t = β

Implies W t(y) = ηt αt +β

• lny +lnηt −lnNt

troublesome ηtβ lnηt.

Dunbar, Lewbel and Pendakur

DLP impose sufficient conditions to identify resource shares with just Engel curves for assignable goods within households.

◮ resource shares do not depend on the household budget so that

ηt(y) = ηt

◮ individual Engel curve functions are given by the Almost Ideal demand

system of Deaton and Muellbauer (1980): w t(y) = αt + β t lny

◮ preferences are similar—but not identical—across people, such that

β t = β

Implies W t(y) = ηt αt +β

• lny +lnηt −lnNt

troublesome ηtβ lnηt.

Dunbar, Lewbel and Pendakur

DLP impose sufficient conditions to identify resource shares with just Engel curves for assignable goods within households.

◮ resource shares do not depend on the household budget so that

ηt(y) = ηt

◮ individual Engel curve functions are given by the Almost Ideal demand

system of Deaton and Muellbauer (1980): w t(y) = αt + β t lny

◮ preferences are similar—but not identical—across people, such that

β t = β

Implies W t(y) = ηt αt +β

• lny +lnηt −lnNt

troublesome ηtβ lnηt.

Dunbar, Lewbel and Pendakur

DLP impose sufficient conditions to identify resource shares with just Engel curves for assignable goods within households.

◮ resource shares do not depend on the household budget so that

ηt(y) = ηt

◮ individual Engel curve functions are given by the Almost Ideal demand

system of Deaton and Muellbauer (1980): w t(y) = αt + β t lny

◮ preferences are similar—but not identical—across people, such that

β t = β

Implies W t(y) = ηt αt +β

• lny +lnηt −lnNt

troublesome ηtβ lnηt.

Linear DLP

Let W t

h = at h +bt lnyh +εt h

where at

h = ηtαt +ηtβ lnηt −ηtβ lnNt h,

and bt = ηtβ Let Wh = ∑t W t

h be the household budget-share for all assignables:

Wh = ah +blnyh +εh where b = ∑bt = β regress W t

h and Wh on lnyh, collect

bt and b and compute

• ηt =

bt / b If the denominator ˆ b = ∑ bt

h is close to zero, you get garbage.

Linear DLP

Let W t

h = at h +bt lnyh +εt h

where at

h = ηtαt +ηtβ lnηt −ηtβ lnNt h,

and bt = ηtβ Let Wh = ∑t W t

h be the household budget-share for all assignables:

Wh = ah +blnyh +εh where b = ∑bt = β regress W t

h and Wh on lnyh, collect

bt and b and compute

• ηt =

bt / b If the denominator ˆ b = ∑ bt

h is close to zero, you get garbage.

Linear DLP

Let W t

h = at h +bt lnyh +εt h

where at

h = ηtαt +ηtβ lnηt −ηtβ lnNt h,

and bt = ηtβ Let Wh = ∑t W t

h be the household budget-share for all assignables:

Wh = ah +blnyh +εh where b = ∑bt = β regress W t

h and Wh on lnyh, collect

bt and b and compute

• ηt =

bt / b If the denominator ˆ b = ∑ bt

h is close to zero, you get garbage.

Linear DLP - Extensions

Pretest

◮ ˆ

b is the estimated coefficient on lny in the Engel curve for, e.g., all clothing at the household level.

◮ If it is insignificantly different from zero, these methods won’t work.

You can allow the resource shares, the intercepts and slopes of Engel curves to depend on demographics z, including N. You can test the per capita model within L-DLP.

Linear DLP - Extensions

Pretest

◮ ˆ

b is the estimated coefficient on lny in the Engel curve for, e.g., all clothing at the household level.

◮ If it is insignificantly different from zero, these methods won’t work.

You can allow the resource shares, the intercepts and slopes of Engel curves to depend on demographics z, including N. You can test the per capita model within L-DLP.

Linear DLP - Extensions

Pretest

◮ ˆ

b is the estimated coefficient on lny in the Engel curve for, e.g., all clothing at the household level.

◮ If it is insignificantly different from zero, these methods won’t work.

You can allow the resource shares, the intercepts and slopes of Engel curves to depend on demographics z, including N. You can test the per capita model within L-DLP.

Men’s Food Engel Curve

All Food Engel Curve

Men’s Food Engel Curve, vs ln Expenditure

All Food Engel Curve, vs ln Expenditure

Men’s Food Engel Curve, vs ln Expenditure

All Food Engel Curve, vs ln Expenditure

The Man’s Resource Share

The estimated slope of the man’s food share is bm = −0.065 The estimated slope of the overall household food share is b = −0.145 The man’s resource share is ηm = −0.065

−0.145 = 45%

Suppose the total household spending y is US$2000 over the year. Then, the man’s budget is y m = ηmy = 0.45∗US$2000 = US$900 If we compare this to the poverty line of US$694 per year, we would say this man is not poor. But, we know that someone else in the household is poor, because there’s not enough remaining ($1100) to give each of them $694.

The Man’s Resource Share

The estimated slope of the man’s food share is bm = −0.065 The estimated slope of the overall household food share is b = −0.145 The man’s resource share is ηm = −0.065

−0.145 = 45%

Suppose the total household spending y is US$2000 over the year. Then, the man’s budget is y m = ηmy = 0.45∗US$2000 = US$900 If we compare this to the poverty line of US$694 per year, we would say this man is not poor. But, we know that someone else in the household is poor, because there’s not enough remaining ($1100) to give each of them $694.

The Man’s Resource Share

The estimated slope of the man’s food share is bm = −0.065 The estimated slope of the overall household food share is b = −0.145 The man’s resource share is ηm = −0.065

−0.145 = 45%

Suppose the total household spending y is US$2000 over the year. Then, the man’s budget is y m = ηmy = 0.45∗US$2000 = US$900 If we compare this to the poverty line of US$694 per year, we would say this man is not poor. But, we know that someone else in the household is poor, because there’s not enough remaining ($1100) to give each of them $694.

The Man’s Resource Share

The estimated slope of the man’s food share is bm = −0.065 The estimated slope of the overall household food share is b = −0.145 The man’s resource share is ηm = −0.065

−0.145 = 45%

Suppose the total household spending y is US$2000 over the year. Then, the man’s budget is y m = ηmy = 0.45∗US$2000 = US$900 If we compare this to the poverty line of US$694 per year, we would say this man is not poor. But, we know that someone else in the household is poor, because there’s not enough remaining ($1100) to give each of them $694.

The Man’s Resource Share

The estimated slope of the man’s food share is bm = −0.065 The estimated slope of the overall household food share is b = −0.145 The man’s resource share is ηm = −0.065

−0.145 = 45%

Suppose the total household spending y is US$2000 over the year. Then, the man’s budget is y m = ηmy = 0.45∗US$2000 = US$900 If we compare this to the poverty line of US$694 per year, we would say this man is not poor. But, we know that someone else in the household is poor, because there’s not enough remaining ($1100) to give each of them $694.

Data

Living Standards Measurement Study (LSMS): Household surveys collected by national statistical offices since 1980s, with World Bank assistance in the design and implementation. These data are standardised to some extent, and are the best tool available for cross country comparisons of poverty in low- and middle-income countries. About 40 countries, and often several waves. In total 87 country-waves. We analyse the most recent waves from 12 countries for which LSMS data include clothing expenditure by type of individual (men, women and children), a measure of total expenditure for the household, and a minimal set of demographic variables (age, sex and education level of household members). We also use the Bangladesh Integrated Household Survey, which contains information on individual food consumption, so that we can consider using food as the assignable good.

Data - Descriptive statistics

Country total N single compositions Our N Nuclear budget std dev Albania 3599 239 mw, mwc 3279 612 11084 6477 Bangladesh 6434 143 mw, wc, mwc 4288 1472 6460 6879 Bulgaria 3018 801 mw, mwc 2099 412 13117 7954 Ethiopia 4717 503 mw, wc, mwc 3845 1481 3092 3645 Ghana 8687 1922 mw, mc, wc, mwc 6313 2195 5096 4835 Iraq 17513 288 mw, wc, mwc 14297 5487 26188 14287 Malawi 12271 1030 mw, wc, mwc 10873 5488 3189 3758 Nigeria 4600 349 mw, wc, mwc 3556 1013 6656 20322 Tajikistan 1503 54 mw, mwc 1275 192 10483 6250 Tanzania 3352 320 mw, wc, mwc 2677 1133 7219 5164 Timor Leste 4477 229 mw, wc, mwc 3788 1577 4954 4116 Uganda 3117 257 mw, wc, mwc 2468 1014 2462 2262 Bangladesh 6434 143 mw, wc, mwc 3929 1330 6511 6969 (food)

Pre-test

Country sample N budget std dev slope at t-test of % of sample share mean slope significant Albania 3279 0.0411 0.0416 0.0139 4.6 84 Bangladesh 4288 0.0407 0.0207

• 0.0157
• 21

99 Bulgaria 2099 0.0355 0.0398 0.0144 5.1 90 Ethiopia 3845 0.0716 0.0636

• 0.0109
• 3.5

65 Ghana 6313 0.0476 0.04

• 0.0021
• 1

62 Iraq 14297 0.07 0.0465 0.0209 14.8 99 Malawi 10873 0.0246 0.0361 0.0092 10 98 Nigeria 3556 0.0171 0.0235

• 0.0017
• 2

50 Tajikistan 1275 0.0578 0.0502 0.0075 1.8 5 Tanzania 2677 0.0436 0.0578

• 0.0022
• 1

12 Timor Leste 3788 0.0223 0.0205

• 0.0025
• 1.8

48 Uganda 2468 0.0545 0.0521

• 0.0039
• 1.2

5 Bangladesh–Food 3929 0.5621 0.1507

• 0.1181
• 14.6

99

Pre-test

Country sample N budget std dev slope at t-test of % of sample share mean slope significant Albania 3279 0.0411 0.0416 0.0139 4.6 84 Bangladesh 4288 0.0407 0.0207

• 0.0157
• 21

99 Bulgaria 2099 0.0355 0.0398 0.0144 5.1 90 Ethiopia 3845 0.0716 0.0636

• 0.0109
• 3.5

65 Ghana 6313 0.0476 0.04

• 0.0021
• 1

62 Iraq 14297 0.07 0.0465 0.0209 14.8 99 Malawi 10873 0.0246 0.0361 0.0092 10 98 Nigeria 3556 0.0171 0.0235

• 0.0017
• 2

50 Tajikistan 1275 0.0578 0.0502 0.0075 1.8 5 Tanzania 2677 0.0436 0.0578

• 0.0022
• 1

12 Timor Leste 3788 0.0223 0.0205

• 0.0025
• 1.8

48 Uganda 2468 0.0545 0.0521

• 0.0039
• 1.2

5 Bangladesh–Food 3929 0.5621 0.1507

• 0.1181
• 14.6

99

Predicted Resource Shares, Selected countries

Evaluated at mean mh Evaluated at all mh % with Country sample N men women ch. men women ch. an eta ou est est est mean mean mean [0,1] std err std err std err std dev std dev std dev Albania 3279 0.28 0.24 0.14 0.29 0.24 0.13 0.0561 0.03 0.03 0.03 0.44 0.36 0.16 Bangladesh 4288 0.31 0.27 0.13 0.31 0.27 0.13 0.0005 0.01 0.02 0.01 0.11 0.12 0.06 Bulgaria 2099 0.30 0.37 0.19 0.30 0.38 0.18 0.0734 0.04 0.04 0.06 0.14 0.22 0.21 Iraq 14297 0.27 0.24 0.04 0.27 0.24 0.04 0.0124 0.01 0.01 0.01 0.13 0.13 0.07 Malawi 10873 0.31 0.27 0.12 0.31 0.27 0.13 0.0143 0.03 0.03 0.01 0.22 0.17 0.09 Bangladesh 3929 0.29 0.24 0.17 0.30 0.23 0.17 0.0501 Food 0.02 0.02 0.01 0.11 0.12 0.08

Gender gaps

Estimated Resource Shares and Gender Gaps, Selected Countries Households with Both Men and Women Present Evaluated at all z Evaluated at mean z Gender Gap sample N men women men women mean mean est est est sig std dev std dev std err std err std err Albania 3279 0.3413 0.2808 0.2823 0.2448 0.0376 0.3501 0.253 0.0311 0.032 0.0571 Bangladesh 3850 0.3419 0.2848 0.3096 0.2525 0.0571 *** 0.1136 0.0951 0.0132 0.0127 0.0237 Bulgaria 2099 0.3171 0.4435 0.3048 0.3702

• 0.0654

0.1501 0.2116 0.0371 0.0408 0.0683 Iraq 14040 0.3346 0.2871 0.2685 0.2327 0.0358 *** 0.1387 0.1217 0.0092 0.0086 0.0166 Malawi 9490 0.3633 0.2797 0.3116 0.2532 0.0584 0.225 0.1404 0.0281 0.0291 0.0542 Bangladesh 3522 0.3281 0.2263 0.2922 0.2217 0.0704 *** Food 0.1167 0.1145 0.0153 0.0144 0.0273

Individual poverty

Estimated Poverty Rates, Selected Countries country per-capita men women children all people est est est est est std err std err std err std err std err Albania 0.003 0.055 0.037 0.075 0.053 0.001 0.043 0.041 0.074 0.031 Bangladesh 0.111 0.022 0.086 0.224 0.117 0.004 0.009 0.019 0.043 0.016 Bulgaria 0.003 0.024 0.027 0.223 0.054 0.001 0.039 0.032 0.120 0.031 Iraq 0.000 0.000 0.002 0.146 0.063 0.000 0.001 0.001 0.070 0.030 Malawi 0.629 0.469 0.585 0.727 0.626 0.004 0.025 0.035 0.030 0.009 Bangladesh 0.104 0.036 0.173 0.093 0.104 Food 0.005 0.018 0.031 0.025 0.011

Discussion (1/4)

Evidence of substantial within-household consumption inequality. Current standard practice for poverty measurement in developing countries— based on asking whether or not per-capita household income consumption falls below a threshold— can provide a misleading picture:

◮ as it ignores within-household inequality, and so mischaracterises

poverty rates.

◮ For example, if a household has income slightly above the poverty line,

then by the per-capita method would call it non-poor, but even a small amount of within-household inequality will result in some members being poor.

Within-household inequality may be biased against certain groups. For the 5 countries for which we estimate resource shares, we see statistically significant gender gaps in resource shares that favour men

• ver women in two countries and we find no statistically significant

evidence of gender gaps that favour women. These gender gaps in resource shares result in gender gaps in poverty rates.

Discussion (2/4)

If within-household inequality is real, and affects the incidence of poverty among men, women and children, then its accurate measurement is of paramount importance. In terms of gathering data on assignable goods, there are two strategies available:

1

Collect data on assignable person-level consumption flows for all categories of goods and services (aka: dream data in Table 1).

⋆ With these data, we would not need a structural model such as ours to

estimate resource shares—we could measure them directly.

2

We could direct resources to gathering assignable consumption flows for 1 or 2 categories of goods and services that can be measured well and which represent a large fraction of total household expenditure.

⋆ With these data, we could estimate resource shares using our structural

model (or any household model that bases identification on assignable goods).

Discussion (3/4)

For Bangladesh, we use two different assignable goods. Clothing, which is roughly 4 per cent of the household budget, and food, which is roughly 56 per cent of the household budget.

◮ Clothing has a venerable history as an assignable good used in this

literature (e.g., survey of Donni and Molina 2018; Calvi 2019; etc). However, the use of clothing is due to its availability in public-use datasets, not to its superiority in other ways.

We find in our work that using food data as an assignable good to identify resource shares delivers estimates that are very similar to those generated from clothing data in Bangladesh. This lends credibility to the identifying assumptions of the model, when clothing is well behaved.

Discussion (4/4)

Food data have five advantages over clothing data.

◮ More plausibly assignable as clothing can be handed down from

member to member. The same food cannot be eaten by two members.

◮ Food consumption is typically measured in quantities (eg grams of

legumes), clothing is measured in dollars of expenditure. There may be more unobserved quality heterogeneity in clothing than in food.

◮ Food budget shares are known to be downward sloping (e.g., Engel

1857, 1895), and therefore satisfy the identifying restriction of our model.

◮ Clothing is much more durable than food. Consequently, observed

clothing expenditure may not equal clothing consumption, due to infrequency of purchase. If total expenditures are polluted with infrequent purchases, the estimates may be polluted with endogeneity.

◮ Food shares are typically much larger than clothing shares. This is not

a gain in terms of the model in any formal sense, but it does seem like a worthwhile auxilliary feature.

All together, this suggests that the collection of person-level food consumption presents signigicant advantages that have to be weighed against its costs and other difficulties.

Conclusions (1/4): Summary of the contribution of this project to the agenda of individual poverty measurement

We show how to estimate the resource share of each person in a collective household via simple linear regressions of assignable goods Engel curves. This may be implemented with off-the-shelf consumer expenditure micro-data, such as the LSMS. We show how to quickly determine, for any data set, whether the method can be applied. We apply the model to data from 12 countries, and investigate resource shares, gender gaps and individual poverty. We find that equal sharing--the implicit assumption underlying household-level poverty calculations---is always rejected. We also find evidence of large gender gaps in resource shares, and consequently in poverty rates, in a few countries.

Conclusions (2/4): So what, in practice, as a poverty economist?

Why use L-DLP? How to use L-DLP? What to do with measures of individual resources? Why use L-DLP?

1

What do the identifying assumptions mean in terms of behavior?

⋆ Resource shares do not depend on the household budget. ⋆ AIDS of Deaton and Muellbauer for individual Engel curves. ⋆ Preferences are similar, but not identical, across individuals. ⋆ These assumptions do not seem unreasonable as structural analysis of

behavior goes.

2

What is the gain over per capita measures?

⋆ Allows consideration of intra-household inequality.

Conclusions (3/4): So what, in practice, as a poverty economist?

How to use L-DLP?

◮ Resource: 2 papers which will be circulated to all participants of the

workshop:

⋆ “Measuring individual poverty with OLS and off-the-shelf household

budget surveys”, Valérie Lechene, Krishna Pendakur and Alex Wolf (2019)

⋆ Title TBC, Theophiline Bose-Ducker, Isis Gaddis, Talip Kilic and

Valérie Lechene

◮ These resources will detail the steps to take to implement the method

(including Stata code), carefully explain the identifying assumptions of the model, as well as all the steps required to produce all measures presented today.

Conclusions (4/4): So what, in practice, as a poverty economist?

What to do with measures of individual resources?

◮ Compare with per capita results and with knowledge of the

environment.

◮ Use the results to analyse any individual outcome (health, schooling,

happiness...)