11. | ..1. .6110“? .111 1 1|- 3.... .1...- 1 . . 1 1.....11. .1. 1 _

hug. 3M... «$111K»... huk§i 3%. .mmamwfiz . :11: . 31”.. ... .4. x“... mgflwfifibai; V... - .
1. T. . W3)“. Costu‘; .l3 .n. , 5.1.1.1 1.01.1.“v 0. 2 . Wmmflfmm‘ . 3n", Sivan»: fl film"!

11 . 21% .1? . V. . He : 1. V... «m... at: 4mm“)

. A 9 .. v o Iv k . 1d,?

. V .15.... 11.1.. 1.. ..

. 1.4.... 1 U... a
.

«flung... .. k1“...qr1
1.- . .4. WI.” II § I. V
. . .. "3...... w... r . .
. ..\ . .
..

1 {1.1.1.1
«1.1%...1. -fiw... 111m. , .. _ : , . 1...
, .m .. 11 - 1.1.1.1.... .. 1. .. .... . . . .. .11.
V Mira .Vndnoviu): . .. . _ .. .1. . ..
. my”! 1 . fig . . .V. . . . 1.. _.. ha... imwwmutd. 181...... .
1.1.1111. ..1V . . . .. .. .4 . . . .. 3mm. Jana”... . .

. L ..
. 1.1V. ....... _ 3.. .. -1. .. . .
n . .. 1 9| 11 Q v
1... . . v1 k1_v.*nnw.$wahfiflhfimmnww . ..
. .2... . . 1 .. .1.V.w.oxyu.1.....w. “filmnmflmflomnv. .. .. . . 1
. . 1 .1 . 1. v}. I - l~ - I c ,..Vo.' I.‘
. . . . . . .1. mmunWBEV1Lrhur-wwwfia 1.1.1.... 1 .. ..... 1.. n
U. 1 al A I .1- .0» . .1
O. u 11-1 1 :01 .Mh1. d11.1.1I»>.v 11.~1b| 1 .1 .11! 1. .-
.. .. 1... .. . . .. . .
1 1... 1 11.5
- ...11.'

c ‘1
, hangar... .
.. 1 . .:..

I 0 II A
o. I if... v! . .
.W‘vfiluqovo 1 .n. . ..
‘l . . . _. ... .
c U! I ‘.1\ I. u D . . .0 1! I 9
1. , . 1 . 1 p I 1 1 .I
- \ A . . I 1‘ .O!
. u .
1 . . 1
.1 . 1.. . 1. ... A'r. . 1 .. a.
1 1 p. ..1 1 . 1 1 .1 11' . .1111 011.111..
. . o o . 1. . ...1 . ...V1:...U1Uodrvvltbufluulr....1.!1.h11 l 1
.. , 1 111.1 . .1..111l..1.l.ll.l.oa 11
. ‘

l

. 1 . .1.
. . 1,
V , V $115k]. .mfibI31110111
. irritqg
1....1uflfl. 11.1.1.3. 111.0:- I. 1
.1. 111......11111 15.111.19-
191...!)

1.1.1 11.1.1...

1 n 11.3.13!
.. I u. " 'ROIOI‘W‘...‘
3" .HIJCO

Hammmflmr

.. 11.1..
.. . . .. .....111.11...1unP.1.r11 1.1.1.
.1112..- 1. 1.: nldtrflluw‘urllfluw‘llni
. .1.111..u1\.l..1.1....1:14«1.\.1.111111.1n&l111fl. \111
“fling “moll‘firlvo. .9111}! 1.MU11.1 ...U..1
.V.. . .. Vv...11.l1u.1...1.”1h11:um.1.u1111u1$ -
II"... 1111. 1.11
. (1“...
1.1”31." 1.191.“...
g1

9; ll IIIIIIV‘a. -
2".11 a; 11 . .. 1
1 .0.11.1.I.§111u.1u1 1 .

It. . I I I

‘1‘ (.1

Pilununufififl. . . 1..
nah? \(4 '1 D lelle. ‘1“1 I
..n1111..r1.1l.n1. ... 1.|1.....1,.111..11,1I.I...1t11.1.u1..11.n1|

. .1 .1 .. .11
.1 . 11.1. .1..1.1. 111x111. . .1 .6... .11....1r.11:
P9111. .. 1 . : , ..1
7.15219. 111.13.611.11»: .. - .
1.1-5.3911
.'.’OI|Y‘O.1I
1.-. 11.11.

h l‘tlw!‘ I‘
. 1 11%tol
miti..-
. .1 1:9}.1... II
1 l. 1. . .1Ur vvuu
..
.11

,, 1 .1
1,. 1‘ v
1 , 1'
. 1 I, 1
1 I o
. . .
'3 ' <

.’ u
'1
U o ‘ '

1

1‘ ’ 1
"3g”
.1! 1. 1

. , .‘1‘3'9'

‘ Milt. ;
31

.1?

1., .1 v. .
v 000‘.
1.. 11. . .111-.. . o... . v . ...
1. .v. 11.6111. 1. o. . ..1. u . 1 .
1 it; < 1 . 115.. 0.1.111
O 1 iv 1 1 11.... I .111 it...1 .
I- 1 411 1.1....118'.‘u£10‘«1|‘011.
11 ll... 1.10.1.1II. .\1
. . 1. 1'l‘. .0
.1 .1
L .1 4‘ ~ 1
11.0 11

 

llllllllllIlllll'flllllllllEljlflfifllllllll

31293 018

This is to certify that the

dissertation entitled

The Impact of Adult Mortality in Agricultural
Households: Evidence from Rural Tanzania

presented by

Kathleen G. Beegle

has been accepted towards fulfillment
of the requirements for .

Ph . D . degree in Economics ’

 

 

 

mm!» 95, M7

M5 U i: an Affirmative Action/Equal Opportunity Institution

 

0-12771

 

LIBRARY

Michigan State
University

 

 

 

PLACE IN RETURN BOX to remove this checkout from your record.
TO AVOID FINES return on or before date due.
MAY BE RECALLED with earlier due date if requested.

 

 

 

DATE DUE DATE DUE DATE DUE
G; 51 002 to 5 a <

AU %‘ z AUG (it 2604

MAY :0 11 man 5

 

 

 

 

 

 

 

 

 

 

 

 

 

moo Wm.“

THE IMPACT OF ADULT MORTALITY IN AGRICULTURAL HOUSEHOLDS:
EVIDENCE FROM RURAL TANZANIA

By

Kathleen G. Beegle

A DISSERTATION
Submitted to
Michigan State University

in partial fulfillment of the requirements
for the degree of

DOCTOR OF PHILOSOPHY

Department of Economics

1997

ABSTRACT
THE IMPACT OF ADULT MORTALITY IN AGRICULTURAL HOUSEHOLDS:
EVIDENCE FROM RURAL TANZANIA
By

Kathleen G. Beegle

Sub-Saharan populations are challenged with increasing adult mortality rates due
to the HIV /AIDS epidemic, although there is a lack of research on the impact of this
epidemic at a household level. This dissertation explores issues surrounding the impact
of prime-age (15-50) mortality on individual and household behavior in one rural
region of Tanzania. Drawing on panel data from the Kagera region, the main chapter
.of this thesis focuses on time allocation of survivors and household activities. Hours in
farming and home production as well as nonfarm labor participation rates are
examined. Surviving individuals are divided into six demographic groups (by age and
sex) to study the potential differential impact of deaths across groups. Various
activities of households with respect to farming as well as household composition are
investigated.

Generally, adjustments in hours in the two activities examined were not
significantly associated with prime-age deaths of household members. Participation
rates in wage employment were lower only in households with an adult death in the
future 6 months. Households with past adult deaths had a significantly higher
probability of new adults joining the household. Farming activities showed adjustments

in participation of individuals across various crops and reductions, but no response in

acres cultivated per adult resident. Among household with a male death 7-12 months
past, participation in subsistence food crops (maize, cassava and beans) declined while
in tree crops (coffee and banana) participation was not reduced.

Utilizing the same data set, the second chapter explores the consistency in the
retrospective reporting of deaths in households. Comparing the first full household
interview with the baseline (enumeration) survey revealed significant underreporting of
deaths in the former survey. Notwithstanding the differences in the survey
instruments, these differences are hypothesized to be linked to inconsistency in
classification of the deceased as a household member and errors in dating of deaths and
age of deceased. Additionally, change in headship was significantly associated with
probability of underreporting. In addition to informing the dialogue on the quality of
retrospective data, the results of this analysis were applied to the use of the

retrospective data in the first chapter.

ACKNOWLEDGMENTS

I would like to express my gratitude and sincere appreciation to my advisor,
Professor John Strauss, for guidance, advice, counsel and encouragement. I would
also like to thank the other members of my committee, Professors David Neumark and
Robert LaLonde, who provided insight and valuable input.

I am indebted to Martha Ainsworth and Mead Over for bringing me onto their
project, allowing me access to its data and their numerous helpful comments. The
Population Council provided funding in 1995-1996 for which I am grateful.

I have had the good fortune of making many friends at Michigan State. I
especially want to thank Jen Tracey, Kelly Hallman, Heather Bednarek and Jess Reaser
for their invaluable support in this last year. They always make me laugh and try their
best to keep me focused and a little less crazy.

I give special thanks to my family. They never ask too many questions and
always listen to my complaints. Finally, though she will never read this, I thank my

mother whose memory is with me always.

iv

TABLE OF CONTENTS

LIST OF TABLES ................................................................................ vii
LIST OF FIGURES ................................................................................ ix
CHAPTER 1

THE IMPACT OF ADULT MORTALITY IN AGRICULTURAL HOUSEHOLDS:
EVIDENCE FROM RURAL TANZANIA

I. Introduction .................................................................................. 1
11. Literature Review .......................................................................... 6
Economics of Time Use ............................................................... 6
AIDS and the Macroeconomy ........................................................ 9
Impact of AIDS at the Household Level .......................................... 14
III. Model ..................................................................................... 18
A Basic Model of Time Allocation ................................................ 18
Dynamic Impact of Death ........................................................... 22
Time Use and Prime-Age Adult Death ............................................ 22
IV. Data ....................................................................................... 27
Data Setting ............................................................................ 27
Kagera Health and Development Survey .......................................... 29
Individual Time Allocation Patterns ............................................... 31
Household Activities ................................................................. 34
Reporting of Prime-Age Deaths .................................................... 35
V. Empirical Specification ................................................................. 38
Time Allocation ....................................................................... 38
Household-Level Activities ......................................................... 45
Estimation Techniques ............................................................... 48
VI. Results .................................................................................... 50
Farm Hours ............................................................................ 51
Farming Activities of Households .................................................. 54
House Hours ........................................................................... 60
Farm and house Hours by Per Capita Expenditure ............................. 61
Nonfarm Self-employment and Wage Employment ............................. 62
VII. Conclusions ............................................................................. 63

APPENDIX 1 ...................................................................................... 84

Sampling Strategy and Attrition ........................................................... 84
Data Editing .................................................................................. 86
APPENDIX 2 ...................................................................................... 91
APPENDIX 3 ..................................................................................... 103
APPENDIX 4 ..................................................................................... 107
APPENDIX 5 ..................................................................................... 1 1 1
APPENDIX 6 ..................................................................................... 117
LIST OF REFERENCES ....................................................................... 123
CHAPTER 2
EXAMINING REPORTING ERRORS IN RETROSPECTIVE MORTALITY DATA
1. Introduction ............................................................................... 129
H. Recall Error .............................................................................. 131
III. Data ...................................................................................... 134
IV. Incidence of Death in Kagera Households ......................................... 136
V. Consistency in Reporting of Deaths ................................................. 139
Comparison of the Two Surveys .................................................. 140
Probability of Underreporting in Wave 1 ........................................ 144
VI. Conclusions ............................................................................. 145
LIST OF REFERENCES ....................................................................... 151

vi

LIST OF TABLES

CHAPTER 1
Table 1: KHDS Sample of Individuals ................................................... 70
Table 2: Sample Statistics .................................................................. 71
Table 3: Impact of Adult Household Member Deaths (15-50) on Hours .......... 72
Table 4: Impact of Adult Household Member Deaths (15-50) on Hours .......... 75
Table 5: Impact of Adult Household Member Deaths (15-50) on Probability
of Adult Joining Household in Last 6 Months ............................... 78
Table 6: Farm Hours of Adult Joiners and Other Resident Adults, Sample of
Households with Any Past Death .............................................. 79
Table 7: Impact of Adult Household Member Deaths (15—50) on Hired Labor
in Last 6 Months .................................................................. 80
Table 8: Impact of Adult Household Member Deaths (15-50) on Land
Cultivated in Last 6 Months, Household Fixed Effects .................... 81
Table 9: Impact of Adult Household Member Deaths (15-50) on Crops
Farmed by Any Household Member in Last 7 Days, Probit. ............. 82
Table 10: Impact of Adult Household Member Deaths (15-50) on N onfarm
Self-employment and Wage-employment of Men (20—5) .................. 83
APPENDIX 1
Table A1 . 1: Sample of Households in the KHDS by Wave .......................... 89
Table A1.3: Sample Attrition in the KHDS ............................................. 89
Table A1.3: Observations Dropped from KHDS Data ................................ 90
APPENDIX 2
Table A2.1: Sample Statistics ............................................................. 91
Table A2.2: Hours Equations .............................................................. 92
Table A2.3: Hours Equations ....... ; ...................................................... 95
Table A2.4: Probability of Adult Joining Household in Last 6 Months ............ 98
Table A2.5: Hired Labor in Last 6 Months ............................................. 99

Table A2.6: Land Cultivated in Last 6 Months, Household Fixed Effects ....... 100
Table A2.7: Crops Farmed by Any Household Member in Last 7 Days,

Probit .......................................................................... 101
Table A2.8: Wage Employment and Nonfarm Self-employment of
Men (20-50) .................................................................. 102
APPENDIX 3
Table A3. 1: Impact of Adult Household Member Deaths (15-50) on Farm
Hours, Random Effects Estimates ....................................... 103
Table A3.2: Farm Hours Equations, Random Effects Estimates ................... 104

vii

Table A3.3: Impact of Adult Household Member Deaths (15-50) on Farm

Hours, Random Effects Estimates ....................................... 105
Table A3 .4: Farm Hours Equations, Random Effects Estimates ................... 106
APPENDIX 4
Table A4. 1: Impact of Adult Household Member Deaths (15-50) on House
Hours, Random Effects Estimates ....................................... 107
Table A4.2: House Hours Equations, Random Effects Estimates .................. 108
Table A4.3: Impact of Adult Household Member Deaths (15-50) on House
Hours, Random Effects Estimates ....................................... 109
Table A4.4: House Hours Equations, Random Effects Estimates .................. 110
APPENDIX 5

Table A5.1: Impact of Adult Household Member Deaths (15-50) on Hours,
Sample of Households Below Median Expenditure per Capita ..... 111
Table A5.2: Hours Equations, Sample of Households Below Median
Expenditure per Capita ..................................................... 113
APPENDD( 6
Table A6.1: Impact of Adult Household Member Deaths (15-50) on Hours,
Sample of Households Above Median Expenditure per Capita ..... 117
Table A6.2: Hours Equations, Sample of Households Above Median
Expenditure per Capita ..................................................... 120

CHAPTER 2
Table 1: Reporting of Deaths During Enumeration and at Wave 1 Interview 149
Table 2: Probability of Reporting No Adult Death in Past 24 Months at
Wave 1 ............................................................................ 150

viii

LIST OF FIGURES

CHAPTER 1
Figure 1: Tanzania and Kagera Region ............................................................ 68
Figure 2: Participation Rates for Last 7 Days .................................................... 69
CHAPTER 2
Figure 1: Households Reporting a Prime-Age Death, Unweighted .......................... 148
Figure 2: Households Reporting a Prime-Age Death, Weighted ............................. 148

ix

Chapter 1
THE IMPACT OF ADULT MORTALITY IN AGRICULTURAL HOUSEHOLDS:
EVIDENCE FROM RURAL TANZANIA

I. Introduction

An estimated 1.9 million adults have died of Acquired Immune Deficiency
Syndrome (AIDS) in Africa through 1994 (World Health Organization, 1995). Despite
the projected decline in the number of new Human Immunodeficiency Virus (HIV)
infections by the year 2000, deaths from AIDS in Africa will continue to rise well into the
next century. While there are other more prevalent diseases in Africa, the characteristics
of HIV/AIDS suggest that its economic and demographic impact will be profound.‘ In
the absence of the AIDS epidemic, prime-age deaths are relatively rare in Africa, about
six deaths per 1,000 persons ages 15-50 per year. In an area with high levels of HIV
infection, prime-age mortality rates can double, triple or even quadruple (World Bank,
1993). Moreover, HIV/AIDS is not restricted to poorer populations. Unlike other major
diseases in Africa, HIV /AIDS is prevalent among the better educated and higher income
Africans in urban areas, with some evidence indicating higher rates among these groups.
Although HIV/AIDS is typically considered to be primarily an urban epidemic, there has
been increasing concern about the implications of HIV in rural areas where rates appear

to be increasing (National Research Council, 1996).

 

‘ See Ainsworth and Over (1994a) for an overview of the economic implications of the AIDS epidemic in
the context of African development.

1

2
Barnett et a1. (1995) assert that in rural areas the main burden of the impact of

HIV /AIDS falls on the labor economy of the household, but little empirical evidence
exists of this effect. They hypothesize that increased adult mortality and morbidity
associated with HIV /AIDS will result in, among other outcomes, a reduction in range of
crops grown, in particular cash crops, due to loss of labor available to the household, an
inability to purchase farm inputs, and diminished capacity to manage household
economies and market agricultural produce. As households try to sustain farm production
and shift labor towards farming, other important activities, such as fetching clean water,
will be compromised.

Drawing on a panel data set from the Kagera region of Tanzania, the Kagera
Health and Development Survey (KHDS), this dissertation estimates the impact of prime-
age deaths on activities and time allocation in rural households among surviving
household members. The potential reallocation of time by individuals and changes in
household activities are of interest for several reasons. First, adjustments in labor supply
may give insight into the mechanisms that households use to smooth consumption in the
event of the death of an adult member. Much research has shown that households in rural
areas are able to smooth consumption in the face of weather and other income-related
shocks, though the means by which they smooth are not well researched (Townsend,
1994). From the data used in this study, there is evidence of consumption smoothing
among households experiencing an adult death (Over and Dorsainvil, 1996).

Second, if time allocation is adjusted to buffer households from the shock of an

adult death, it may be the case that individuals within the household have differential

3
outcomes, where some groups bear a larger share of the adjustments. In other words,

household level outcomes may mask differential outcomes across household members as
households smooth consumption or otherwise cope with an economic shock. To date,
most research examining household responses to various types of income shocks have
emphasized that certain types of households (e. g. landed versus landless) may be better
able to smooth consumption. Much less analysis has examined the potential differential
effects of income shocks on individuals which may be an important factor in the design
and implementation of policies (Quisumbing, 1996).

Third, changes in labor activities may be associated with changes in diversification
over income sources. In a rural region like Kagera, diversification can provide security
against the volatile income stream associated with agriculture.2 The economic viability of
households may be compromised if, for example, households must shift participation of
household members away from off-farm activities or cash crops to subsistence farming.

Understanding the interaction between individual labor participation rates of
survivors and prime-age adult mortality will be informative in the application of
macroeconomic simulations used to estimate the impact of the HIV /AIDS epidemic on
economic growth. Cuddington (1993) proposes that this type of analysis can be used to

evaluate the economic outcomes associated with different health sector interventions.

A

 

2 See the study by Collier et al. (1990) for a description of the degree of diversification over economic
activities by households in rural Tanzania. Collier et al. note that while rural households have a narrow
range of employment activities these households engage in an array of distinct economic activities within
farming. Reardon (1997) finds that earnings from nonfarm wage employment (not necessarily in the farm
sector), in addition to nonfarm self-employment, is of dominant importance to rural farm households in
Africa. The effectiveness of income diversification at reducing income risk depends on the covariance
between income sources. The lower the covariance between sources, the greater the potential reduction in
total income risk.

4
Analysis of individual and household behavior can, in turn, inform those undertaking

these types of simulations. Finally, if high rates of adult mortality result in farm labor
shortages that cannot be easily made up for by hiring in labor, then we may observe a
decrease in crop production, as speculated in World Bank (1992). This has implications
for food security. Brown et al. (1994) contend that diminished labor quantity and/or
quality will increase the incidence and depth of household poverty and, thereby, food
insecurity. They conclude that the implications of the AIDS epidemic for food security
across Africa are “potentially huge” where labor power is crucial to agricultural
production. Furthermore, this effect may be heightened by arguably equal or higher HIV
infection rates among women who, in some parts, account for the majority of the
agriculmral labor force. Moreover, agricultural production in Sub-Saharan Africa shapes
the availability and price of food for urban markets and is a major export good.

We anticipate that labor supply (both participation rates and hours across sectors)
will serve as a response mechanism to deaths or other idiosyncratic shocks in households
in part because there are very limited options for risk coping and management in this
region.3 Rural regions like Kagera lack formal credit and insurance markets, and have
limited options for savings and asset accumulation, i.e. a lack of financial institutions and

a thin land market. While informal institutions and arrangements among households may

 

3 Alderman and Paxson (1992) broadly classify strategies that mitigate risk into two groups: risk-
management and risk-coping strategies. The former attempts to reduce income variability and the latter
attempts to smooth consumption within and across households. Risk-coping arrangements that strive to
smooth consumption across households are classified as risk-sharing strategies. See Dercon (1996) for
further discussion regarding the limited options for risk coping and management in Tanzania. Dercon
examines whether risk management is different depending on households’ level of liquid asset holdings.
The analysis in this study, on the other hand, focuses on potential risk-coping mechanisms within
households, specifically those tied to time allocation.

5
partially accommodate households, labor serves as the pivotal resource to all households.

Although Kagera lacks a large local urban labor market, off-farm labor opportunities
do exist, in addition to nonfarm self-employment.

With respect to farm and household chore hours, this analysis finds that in general
neither male or female prime-age deaths are associated with changes in farm hours of
surviving household members. Other analysis of these data finds that new household
members are joining households affected by a prime-age death, thereby mitigating the
extent of labor loss which would restrict the med of others to increase hours in
households with a past death. These results suggest that one adult death is not necessarily
a one—person reduction in the size of the labor pool of the household. Hiring of labor, an
alternative to acquiring a new member, was not associated with an adult death. Short-run
participation in coffee farming, the main cash crop, and banana farming decreases after a
male death, but this decrease does not persist. Allocation of labor to other food crops
(maize, cassava and beans) declined in households with a distant male death. While acres
of land cultivated decreased in households experiencing a past male death, acres cultivated
per resident adult was not lower in households reporting a death. Thus, households
undertake strategies ranging from bringing in new members and changing some aspects of
farming at least temporarily after an adult death. On the other hand, while the lack of a
sustained increase in farm hours among survivors in households experiencing a prime-age
death is suggestive, the question of the impact of adult mortality on total and per capita

farm output is unresolved.

6
N onfarm wage employment participation among adult men declines in the months

preceding a death, but is not significantly lower in the months following an adult death.
Wage employment is a important source of cash income and income diversification for
rural households. Contrary to speculation by Barnett and Blaikie (1992) and others,
surviving members in these data do not reduce participation rates in the wage sector

This chapter is organized as follows. Section 11 reviews the existing literature
with respect to time allocation studies and studies specifically examining the impact of
HIV/ AIDS. A basic model of time allocation is presented in Section III. Section IV
describes the Kagera Health and Development Survey (KHDS), the panel data used. The
empirical specifications estimated are presented in Section V and results are described in
Section VI. Section VII concludes and draws tentative policy implications from these

results.
II. Literature Review

Economics of lime Use

There is a large literature on the theory and determinants of labor supply, as well
as some empirical evidence bearing on the economics of time use in developing
economies (see Becker, 1965; also see luster and Stafford, 1991 , for a review of this
literature). Some of this research focuses on total or off-farm labor supply. Rosenzweig
(1980) analyzes market labor supply responses to wage rates in India, finding evidence

consistent with the neoclassical-competitive model of labor markets. J acoby (1993)

7
shows that Peruvian peasant households respond in total time worked to the productivity

(or the shadow-wage) of its members. Other work has examined time allocation across
activities in agricultural households (such as Evenson et al. , 1980, Khandker, 1988, and
Mueller, 1984). For example, Skoufias (1993) demonstrates that the intrafamily
opportunity cost of time (market wages in this case) is an important determinant of time
allocated to the market, domestic production, schooling and leisure for adults and children
in rural Indian households.

A few researchers have also examined the responsiveness of time allocation in
rural households to income shocks, such as weather shocks. Rose (1995) finds that Indian
farm households in areas with greater dispersion of rainfall have higher labor market
income, indicating a greater allocation of labor to off-farm labor activities ex ante, an
alternative to own-farm production. At the same time, households compensate ex post for
the productivity effects of short-run weather shocks by increasing their work in the labor
market afterwards. Likewise, Fafchamps (1993) explores the relationship between the
risk associated with farming facing West African farmers and their farm labor allocation
across seasons only. Kochar (1996) examines labor supply changes as a method to
smooth consumption and finds that rural Indian households adjust labor market hours in
response to idiosyncratic crop income shocks. Kochar emphasizes that in developing
countries idiosyncratic shocks which are independently distributed across households
are more important than aggregate shocks at the village level. One difficulty in
analyzing the former is developing a measure of this idiosyncratic shock. Kochar

constructs a measure of profit shocks from observed profit to analyze labor supply

8
responses. However, the shock measure, profits which are not net of family labor, is a

function of endogenous labor supply decisions. J acoby and Skoufias (1995) observe
adjustments in child school attendance in response to seasonal income and wage
movements in rural India. They find negative impacts on schooling of negative shocks
which can help explain low rates of human capital accumulation in the region.

Adjustments in time allocation in response to illness shocks and health status have
also been examined. Morbidity will lower the patient's productivity and can imply time
spent caring for the ill individual, usually by other household members who may also be
otherwise be engaged in productive work. Pitt and Rosenzweig (1990) provide evidence
of the reallocation of time from market activities and school to household "care" activities
of adult women and adolescent children associated with infant illness in Indonesia. In
response to infant morbidity, teenage daughters significantly increased participation in
household activities and decreased school and market activities, as compared to response
of teenage sons. Kochar (1995) found that market labor supply responds to the
prevalence of illness among other household members using data from rural India. She
concludes that farm households may in fact be more vulnerable to demographic shocks
than to crop income shocks.

Own—illness, either chronic or acute, will also impact an individual’s own time
allocation decisions. In work by Pitt and Rosenzweig (1986), using data from Indonesia,
short-term illness of an individual or their spouse is shown to decrease labor supplied by

farmers. Pitt et al. (1990) examine the relationship between one’s own health endowment

9
and labor market activities, concluding that individuals with greater health endowment are

more likely to participate in activities in which productivity is sensitive to health status.
While the occurrence of prime-age adult mortality in households is similar to
economic shocks examined in the studies cited above, certain characteristics of this type
of shock distinguish it. First, while morbidity diminishes labor capacity over the duration
of illness, a death results in the permanent loss of an individual. This implies that
households will need to adjust to the long-term loss of this individual's labor and related
income, in addition to the loss of management skills and acquired human capital
investments. Second, in the case of higher mortality due to AIDS, the deaths are
preceded by severe debilitating illness resulting in extreme limitation in daily living
activities and typically necessitating care by others. Moreover, this shock is targeting
prime-age workers who are typically the most economically productive household
members. Finally, the shock of adult mortality due to illness may be anticipated to some
extent by the household. Thus, households may adopt ex ante strategies to cope with an
impending adult death. These characteristics of prime-age mortality suggest that we
cannot equate it with other sorts of economic shocks facing agricultural households, such

as variation in weather.

AIDS and the Macroeconomy
Most of the empirical analysis of the economic impact of high prime-age adult
mortality due to HIV/AIDS and other fatal illnesses has been at the macro level. There is

a significant body of research estimating population projections under the AIDS

10
epidemic (see, for example, Bongaarts, 1996). A handful of studies have addressed the

macroeconomic impacts of the epidemic, simulating economic growth rates and GDP
per capita given HIV/AIDS incidence projections. If the epidemic only reduced
population size and growth rates, we would actually expect per capita income to rise as
a result.4

Cuddington (1993a) and Cuddington and Hancock (1994) modify the standard
neoclassical (Solow) growth model to simulate the impact of the AIDS epidemic on key
macroeconomic aggregates for Tanzania and Malawi, respectively. In these studies,
AIDS is hypothesized to impact growth through its negative effect on the size of the
labor force and reductions in domestic (private and public) savings resulting from
increased medical expenditure. In addition, workers afflicted with AIDS are modeled
as having lower productivity due to morbidity associated with their affliction. Non-
afflicted workers can also suffer lower productivity (or increased labor days lost) due to
the affliction of others although their participation rates in the labor market do not
change, nor do we see in-migrants with respect to the labor pool. Cuddington and
Hancock (1994) find that the epidemic depresses the annual growth rate of real GDP in
Malawi, but they find little effect on per capita GDP growth rates. Cuddington (1993a)
finds that under the “medium” case scenario GDP is 15-25 percent smaller in Tanzania
by 2010 than it would be without an AIDS epidemic while per capita GDP is reduced

by 0-10 percent. Both these studies use the simplifying assumption of full employment.

 

4 As noted by Over (1992) and citations therein, the large reduction in population size due to bubonic
plague of 14th century Europe resulted in increases in wages and incomes for survivors. However, the
characteristics of the HIV/AIDS epidemic are very different from the bubonic plague for a number of
reasons including rate of transmission, length of infection and mode of transmission.

1 1
Cuddington (1993b) adapts his model of Tanzania by incorporating labor surplus

(unemployment or idle capacity) and dual labor markets. Surprisingly, his simulations
produced roughly similar estimates of the impact of AIDS to those produced in his
single sector model. All three of these studies note that the macroeconomic effect
depends critically on the labor productivity and saving loss parameters assumed. These
parameters capture the fraction of worker productivity and savings lost due to an AIDS-
related illness. Because little empirical work has addressed this topic at the household
level, these studies typically rely on assumptions about “reliable” ranges for responses
among households and individuals, such as identifying “medium” case scenarios for
reduced productivity or savings.

Kambou et. al (1992) construct an 11 sector computable general equilibrium
model of the impact of AIDS in the Cameroon economy. They examine the
consequences of the epidemic where the impact is modeled as a one-time 0.8 percent
decline in overall size of the labor force, with disproportionate reductions among urban
(generally more skilled) labor. They conclude that the effect on the economy is
“devastating” with average annual GDP reduced by 1.7 percentage points from 4.3
percent. Per capita estimates are not reported. Over (1992) models the effect of the
epidemic over the years 1990-2025 drawing on aggregate data from 30 sub-Saharan
African countries. In his analysis, as well as reducing savings, AIDS can
disproportionately impact labor supply across different skill categories. Over calls this
the “socio-economic” gradient of the epidemic. He finds that under the scenario where

half of medical costs are financed from savings and AIDS cases are twice as frequent

12
among more skilled labor, GDP growth rate is reduced by nearly one percentage point

and GDP per capita growth rate falls by .15 percentage points per year. These
reductions in GDP growth rate are relatively large given the low or stagnant levels of
growth observed in most African economies in the late 19803.

These studies may overstate the negative effect of AIDS on economic growth for
several reasons. First, these simulations rely on AIDS case projections that may
overstate the number of future cases as individuals may be adapting or changing their
behaviors in response to superior knowledge of HIV IAIDS. Reliable estimates of the
future number of HIV / AIDS cases is further confounded by the serious under-reporting
of present cases in many developing regions. As Bongaarts ( 1996) emphasizes, the data
on levels of infection are considered to be least reliable for Africa.

Second, Bloom and Mahal (1997) note that extended family networks and
community based organizations may provide medical services which mitigate the
explicit costs associated with the illness. In addition, given the paucity of data,
reasonable assumptions of savings or consumption responses to increases adult
morbidity and mortality are unclear. They even suggest that households with an HIV
infected member may increase precautionary savings. Although given that the amount
of HIV testing is very low, especially in rural areas, and, therefore, for most
individuals HIV status is unknown, it may be that precautionary savings will respond to
the probability of an HIV infected household member. Bloom and Mahal comment that
another reason these studies may overstate the macroeconomic impact is that AIDS may

be more prevalent among the poor who have lower productivity, earnings and

13
utilization of medical services. This characteristic of the epidemic appears to be

generally true for Asian countries, but is not apparent for those in Africa.5 As
Cuddington ( 1993b) addresses, Bloom and Mahal reiterate the possibility that many
developing countries have surplus labor that could mitigate potential output losses.
These previously cited studies all assume that, while productivity effects for persons
afflicted with AIDS may vary, an AIDS death results in the reduction of one laborer in
the economy. They do not allow for the possibility that some surviving labor may
work more in response to one lost laborer, even in economies with no labor surplus.
Rather than simulate the impact of the epidemic, Bloom and Mahal estimate the
effect of AIDS prevalence on the rate of growth of GDP per capita using cross-country
data from 51 developed and industrial countries over the period 1980-1992. They
conclude that the effect is insignificant, while controlling for other covariates and the
possible simultaneity in AIDS prevalence and economic growth. Furthermore, they
stress the need to explore this result “in terms of labor supply and relative shifts in
labor supply. . .caused by AIDS morbidity and mortality.” The lack of a significant
impact of the epidemic is not surprising given the small sample size, the combination of
developed and developing countries, and the possibility that the characteristics of the

epidemic with respect to socioeconomic factors, as note above, may vary across

 

5 See references in Bloom and Mahal of evidence of the link between HIV and poverty, or lower
socioeconomic background, among Asian populations. See Over (1992) and Ainsworth and Over (1994b)
for citations noting the positive link between socioeconomic status and HIV prevalence in Africa. It is
possible that this relationship in Africa may reverse itself as those with higher income and education are
also more knowledgeable about the disease and may be more likely to take precautions against
contracting HIV. In Africa at least, as yet this has not been observed (National Research Council, 1996).

14

countries, particularly in the sample of countries most severely affected by the epidemic

(those in Africa and Southeast Asia).

Impact of AIDS at the Household Level

A handful of studies have examined outcomes of the epidemic at the household
level. Gillespie (1989a, b) models the hypothesized impact of AIDS on farming systems
in Rwanda using data on household labor supply by gender to crops across seasons. He
does not have data on deaths experienced in households, but rather evaluates households’
ability to cope with loss of labor. For several crops and within each of 5 geographical
zones and 4 months, Gillespie calculates the “labour balance” as the difference between
“labour availability” and “labour requirement”. labour availability is computed as the
time remaining after subtracting the time spent in essential nonfarm tasks (water and fuel
collection, food processing, housework, private needs and sleep). labour requirement is
the time spent on the crop within zone and month. He assesses regional vulnerability to
adult mortality by focusing on the labour balance available to households for each crop.
Gillespie’s conclusions characterized the Kagera piedmont, the western region of Rwanda
bordering the Kagera region of Tanzania, as a relatively secure system with respect to
sensitivity to loss of labor via AIDS morbidity and mortality. His analysis does not allow
for changes in labor allocated by gender or the potential for hired labor (or even other
farm inputs) to accommodate labor shortages. Cleave (1974), for example, surveys a

large number of studies of African farmers and finds that traditional allocations of work

15
between sexes are modified in response to labor shortages as well as to accommodate

cash-crop production.

17/“ Barnett and Blaikie (1992), using a detailed field study, present mixed anecdotal
evidence about the farming systems and activities of 129 households in the Rakai district
of Uganda in response to an AIDS death. They find no significant change in agricultural
production in the district. Barnett et al. ( 1995) and Barnett (1994) analyze data from
rapid assessment surveys of households in Uganda and communities in Tanzania and
Zambia. They find some discernible evidence of impact in only one community in
Uganda where they note a shift over the period of 1989-1993 to subsistence crops most
pronounced for poorer households. In the Tanzanian communities, these studies found
little evidence of an impact of morbidity and mortality due to the epidemic. In Zambia,
those farming systems most vulnerable to labor loss are not the severely exposed regions
with respect to HIV prevalence and number of AIDS cases.

Tibaijuka (1997) draws on data collected in 1989 from households in one village
in the Kagera region of Tanzania. Of the 220 households surveyed in this village, her
study focuses on the 10 households that reported an AIDS death since 1985 which was

confirmed in a hospital.6 She evaluates the impact of the death on labor both before and

 

6 The restriction of this analysis to these 10 households based on hospital confirmation of cause of deaths
may produce a rather atypical sample of households experiencing an adult death. For example, average
per patient hospital costs for the 12 adult deaths in these households was about 30,000 Tanzanian
shillings. Consequently, most households reported selling assets (including productive ones such as land
and cattle) or borrowing funds to pay for these expenses while a few were able to pay with family
savings. In comparison with the KHDS data (the data used in this study) on deaths of adults due to AIDS
as diagnosed (not necessarily confirmed) by any health practitioner (about 60 individuals) and unadjusted
for price level changes, Tibaijuka’s hospital costs are more than double the amount reported for all
medical expenditures (hospitalization, drugs and consultations) in the KHDS. Furthermore, in
Tibaijuka’s sample, 9 of the 10 households reported the death of the household head (in all but one case,

16
after the event. For the impact preceding the death, she estimates labor supply response

based on the reported length of illness and composition of the household.7 For adult
deaths, she assumes that the labor days diverted by illness of the victim and nursing care
by a surviving adult is double the length of illness, on average 6 months. This presumes
that every day of illness preceding the death results in the reallocation of one laborer to
nursing among survivors as well as zero labor activity of the sick individual. Given this
approach, it is not surprising that the percent of potential labor days diverted (computed
for the year preceding the death) is quite large, about 30 percent on average but as high as
57 percent in one household. The assumption that each sick day pulls one survivor
completely out of labor activities into nursing care seems rather extreme especially in
light of the main economic activity of self-employed farming that may be more flexible in
terms of caring for sick household members. Tibaijuka does not offer any evidence to
validate this assumption.8

Tibaijuka further explores the impact of death on labor stock by comparing levels
of labor before and after the death(s) where labor is calculated as the man-equivalent
amount of labor by weighting individuals across age categories and summing. In all
cases, household labor stock declined, but dependency burden (labor stock divided by

consumption units) was worse in less than half (4) of the households after the death(s).

 

the head was also a male). In the KHDS, only a quarter of households with any prime-age death
reported the death of the household head.

7 It is not clear if length of illness is the duration of illness when the patient was bedridden or when the
victim had to be nursed, as opposed to the observed onset of illness.

8 Analysis of the KHDS data suggests that this assumption vastly overstates time devoted to nursing a
fatally ill household member. The average weekly hours reported in caring for a sick household member
among prime-age women (the main care givers) in households with an adult death sometime over the next
6 months was under 5 hours; for other survivors, average weekly hours were well under 5 hours.

17
Tibaijuka’s study does not compare any of these outcomes to events occurring over the

same period in households experiencing no death. So, it is unclear that the changes
observed from the period before to after the death are necessarily unique to households “in
which someone dies of AIDS (or, more generally, any illness resulting in death) rather
than outcomes linked to events occurring across all households .‘

The lack of studies of the impact of an adult death at the household level or
across individuals stems mainly from the difficulty in collecting data with the
appropriate information to analyze the issue of adult mortality within households. By
examining adjustments in time allocation in households before and after the event of an
adult death using (a detailed panel data set, this study extends the literature in several
ways. It expands our understanding of how rural households smooth consumption in
the event of economic shocks. Evaluating changes in time allocation of individuals will
be insightful for evaluating the burden of increasing prime-age mortality across
different demographic groups. In particular, are children (females) more hours
absolutely and relative to adults (males). This is also a test of Sen’s (1966) model of
labor surplus in peasant agriculture in which remaining members work harder after a
person leaves either through out-migration or in this case mortality.

The results also will be useful in applications to macro projections on the impact
of the AIDS epidemic which incorporate labor supply responses across different
sectors. For example, in rural labor markets, if individuals are, in the short-run at
least, working more in response to one adult death ex post, then modeling labor supply

reductions as equal to the number of prime-age deaths will overestimate potential output

18
losses. Analysis of adjustments in time allocation will also inform the reasonable range

of estimates of productivity reductions (as assessed by hours) due to the illness
preceding death for both non-survivors and surviving individuals. For instance, do we
observe reductions in hours among survivors in the months preceding a death? This
work may also lend evidence towards the issue of food security and cash crops

production in households with a prime-age adult mortality.

III. Model

A Basic Model of Time Allocation

In the basic model of time allocation, households maximize utility which is a
function of a vector of home-produced agricultural commodities consumed by the
households (C,), a vector of market-produced and purchased goods (Cm), a set of home
produced, non-marketable goods (Ch) and a vector of leisure over all household members
(To, conditional on Z, which is a set of taste/preference shifters9:

(1) U = U(C,, Cm, Ch, T.; Z)

 

9 This approach in which households make decisions to maximize household utility (a unitary model of
household decision making) greatly simplifies the process of intrahousehold decision making. The reduced
forms estimated are generally consistent with collective models of household decision making as well as a
unitary model, although collective models suggest that who owns assets and nonlabor income is important.

Furthermore, the following model is static. It ignores the time dimension to utility maximization where
households maximize the discounted sum of utility over its lifetime. Therefore, it does not capture the
implications of an intertemporal planning problem such as insurance, savings, and even household
composition adjustments over time. Glomm and Palumbo (1993). for example, model optimal intertemporal
consumption decisions in which households face endogenous survival probabilities. Their results emphasize
that understanding the pattern of incprne is necessary to interpret consumption decisions. In this study, the
nature of the variable we are examining, time, to some extent minimizes the intertemporal issue. In
particular, your stock of time can not be transferred across periods.

19
Time is divided between the following non—overlapping activities: farm labor (f),

wage employment (w), domestic work (h) and leisure (I). Where T is a vector of time
over all members,
(2) T=Tf+Tw+Th+Tl

Each individual’s health status is produced as follows:
(3) H=H(C,,Cm,Ch,T,,ui,uhh)
where health depends on consumption of home-produced and market-purchased
commodities, leisure, and unobservable individual (i) and household (hh) specific health
endowment (u). There exists some latent threshold level of health, H“, below which the
individual dies.10

Home produced, non-marketable goods, Ch, are produced as a function of time
and health, implicitly effective time in this activity:

(4) Ch = Chan , H)

Farm output (Q) is produced in accordance with a farm production function which
includes effective hired labor (Uh), effective family labor (L°f) and a fixed stock of farm
assets (A). Effective units of hired and family labor are assumed to be perfect substitutes
in farm production. Effective labor in some activity j is a function of time allocated to the
activity and the health status of the individual:

 

'0 In a dynamic model, current health would also depend on lagged health. Current health status could
alternatively be modeled as the lagged stock of health plus gross investment minus depreciation, where
households choose and produce gross health investment in each period (Grossman, 1972). Moreover, in
addition to its impact on effective time in different activities, health could be included in the utility
function.

20
In this construct, poor health can reduce the productivity of each time unit of

labor. In the extreme, at some point, poor health can drive effective labor to zero (even
before reaching H*, death).

The household’s budget constraint is given by:

(6) PmCm = Pa(Q[Leh+Ler . A] ‘ Ca) ' WLh + WTw + Y

where the p's are vectors of prices for market and home-grown marketable commodities,
w is a vector of wage rates, and Y is non-labor income. Wages are not the same for all
individuals, but can vary across household members and hired labor types. The budget
constraint can be re-evaluated such that the market value of consumed goods (including
leisure) must equal full income where full income is farm profit, the value of all time
available and non-labor income.

With respect to time allocation, the solution to this problem yields the
conventional result that our time demand function for individual i in activity j will be a
fiinction of prices, wages and full income. Full income consists of farm profit (1:), the
value of the total stock of time to all household members and non-labor income.
Additionally, individual and household-level unobservables enter this equation:

(7) T"i = Tj,(p,,,, p,, w, T, n, Y, o, , om; Z)

Unobservable specific variables were introduced through the health production
function but could feasibly be incorporated elsewhere. For example, household-level
unobservables could enter in farm production and individual specific endowment may
affect effective labor, conditional on health. Market wage offers may also respond to

unobservables.

21
In this framework, household farming decisions are separable from consumption

decisions. In other words, under the assumption that households can find market
substitutes for labor used in farming, household decision making is recursive. One
substantial prediction of this recursive model is that while changes in demographic
characteristics of the household may affect consumption and time allocation decisions,
they will not affect total farm labor demand, output or profits. That is, under
separability, farm decisions made with respect to some level of maximized profit (it) are
not affected by an adult death or health status of household members. 11 This is not to say
that the household’s farm labor supply is unaffected by the morbidity and mortality of
household members. Farm labor supply, as well as time allocated to other activities, will
respond to taste and preferences, and full income. In addition to it, full income includes
the value of the total stock of time of all household members and non-labor income which
will be affected by an adult death and the preceding illness.

If separability does not hold, then, in addition to the impact on labor supply of
household members , deaths may impact outcomes at the household level with respect to
farming practices such as total farm labor demand and output. In other words,
decisions made with respect to maximization of farm profit will be a function of the

demographic composition of the household.

 

” See Singh, Squire and Strauss (1986) for a full discussion of the topic of separability or recursivity in
agricultural household models. Pitt and Rosenzweig (1986) develop a similar model focused specifically on
farmer’s responses to health with regard to farm labor and profit; they test for and do not reject their
assumption of separability of consumption and production decisions using data on households in Indonesia.
The restrictive assumptions underlying separability make these results difficult to generalize to other settings.
Newman and Gertler (1994) develop a model considering multiple activities in a multiple-person household
under the assumption that consumption and production are joint decisions.

22

Dynamic Impact of Death

One shortfall of the above model of time allocation is the lack of dynamics with
respect to an adult death and outcomes for surviving household members. As an
economic shock, we expect that there will be a time dimension in the various possible
ways that households respond. The static model considers only permanent adjustments to
the economic shock of a prime-age death. Clearly, household responses to a death are not
constrained to a once-and-for—all adjustment. First, if prime-age deaths are caused by
illness, as in the case of HIV /AIDS, households may anticipate the death and adopt
coping strategies preceding the event. These ex ante adjustments can extend over the
duration of the illness leading to death.

Ex post, household responses may also vary over time for a number of reasons. 12
For example, if households cope with a death by initially selling nonproductive assets,
then time allocation may be unaffected by a death only up to the point where households
exhaust this wealth. If other households temporarily aid household experiencing an adult
death, then, again we will be more muted responses closer to the event. On the other
hand, time allocation responses might be larger in the period closest to the death if, for

instance, replacement labor is not readily available.

Time Use, Household Activities and Prime-Age Adult Death

 

‘2 Barnett and Blaikie (1992), for example, hypothesize that the impact of AIDS deaths occurs over a
period of at least five years particularly in settings where multiple members die of HIV/AIDS.

23
The preceding illness and subsequent death of a household member will affect the

allocation of time of remaining members through full income. 13 There are several aspects
to this full-income effect. First, an adult death changes the total stock of household time
available for allocation in two ways. First, the total time of the deceased is gone. Among
individual survivors, the death also implies constraints in the short-run on time use. For
example, we normally associate deaths with customs such as funerals and mourning,
which includes social customs mandating periods of mourning where mourners cease
participation in economic activities. A death due to illness such as AIDS may imply
demands on non-afflicted members’ time in caring for the ill and assistance in seeking
medical care prior to death. While these are not necessarily exogenous constraints on
time, they are, nonetheless, activities that now compete with other activities to which
individuals allocate their time.

We expect that the illness that precedes death will affect the effective time of the

dying household member through health status. 14 As well, the quality of labor inputs of

 

'3 This treats the event of death as an exogenous shock to the household. It should be noted that even in
cases of a fatal disease, such as AIDS, life can be prolonged‘depending on health inputs. For example,
higher levels of health care and more advanced medications is one reason cited for significantly longer
durations of AIDS in HIV /AIDS patients in the US than in African countries (see, for example, Ryder
and Mugerwa, 1994). To this end, households can potentially influence the timing of death. In the
region studied in this analysis, these high—cost health care and advanced medication options are not
available to households so it is presumed that the degree to which households can prolong life for
household members afflicted with “full-blown” AIDS is minimal. Furthermore, given the low level of
testing for HIV in Africa, especially in rural areas, health care intended to delay the onset of AIDS is not
considered. For diseases that are not always fatal, but nonetheless are the cause of death for some
individuals, it is more difficult to assert that the death is exogenous because health inputs, endogenous to
households, most likely influenced, the outcome of death.

‘4 Obviously the extent to which individuals suffer physical disability or limited activities will depend
upon the specific illness with which they are afflicted. O’Dell et al. (1996) study physical function
among a sample of American adults with AIDS and find mild to moderate disability in instrumental
activities of daily living (such as grip and reach) although severe disability is rare. Their results may
understate morbidity among Africans with AIDS where general levels of health are quite lower. At the
same time, this implies that the differential in health status between adults with a fatal illness and others

24
other members may be diminished if they contract illnesses themselves from the presence

of infections and diseases. 15 Health status of afflicted and non-afflicted adults can impact
effective labor across activities. If potential employers can make inference regarding
individual effective labor, wage offers in the labor market will respond to health status.
Of course, piece-rate wages would automatically adjust for changes in effective labor.
Time-rate wages may also adjust, albeit imperfectly, for changes employers perceive in
effective labor, via health status.

An adult death usually mandates expenditures on terminal medical care and
funerals which can be a significant portion of household cash income. 16 Households
experiencing a death may need to raise cash to accommodate these expenses while at the
same time maintain other basic needs expenditures. Estimates from the Rakai district of
Uganda found that medical and funeral costs were approximately one to three months of
median household cash income (Konde-Lule et al. , 1995). Two-thirds of households sold
property to supplement cash income to pay for these expenses. The majority of these
expenses can be funeral costs rather than medical costs. The data from Kagera, Tanzania,
used in this study, indicate that funeral expenses were at least double that of medical
expenses for deaths of prime-age adult household members. Cost of burial and mourning

ceremonies can be quite high when social norms mandate that households pay for food,

 

may not be larger in African countries relative to developed regions, as non-afflicted adults in the former
have lower levels of health.
'5 The illness that results in adult mortality is also an important consideration with respect to the potential
morbidity among survivors as a consequence of this illness. AIDS deaths are associated with numerous
apportunistic infections, unlike other fatal illnesses such as cancer.

See references cited in National Research Council (1996) for evidence of the economic burden of medical
care and funerals for AIDS-related deaths in Sub—Saharan Africa.

25
lodging and, perhaps, transportation of mourners. These expenditures result in a

reduction of full income, and, thereby, may affect time allocation.

In addition to the expenditure associated with death, assets such as land rights and
other property held by the individuals may be inherited by individuals outside the
household after a death. Women may be especially vulnerable to the loss of a spouse
and, consequently, ownership of productive or otherwise valuable assets. 17

Finally, the event of deaths of non-member relatives may also influence full-
income through non-labor income. These deaths may be associated with financial and
labor shocks to households with no household member deaths if there exist Why. _.
arrangements across households. This may result if, for example, some non-members are
actually former household members who migrated for employment purposes and no
longer met the criterion for membership. Consequently, remittances received by a
household may be vulnerable to deaths of non-household members. At the same time,
households may contribute to medical and funeral expenses for deaths in their extended
family. Therefore, not only are potential remittances into the household decreased with
non-household member deaths, but expenditures associated with deaths may be incurred
in households in which no death had occurred.

Ties across households in the Kagera Health and Development Survey are
apparent in the flow of resources over the course of the panel. Over each of the 6 month

intervals between the four household interviews of the panel, we observe at least one new

 

n Women’s access to and legal title over land, in particular, may be susceptible to customs that restrict
the rights of women to own or inherit land. Tibaijuka (1997) comments that a widow without children is
especially vulnerable to being forced to leave her dead husband’s estate and she is usually not welcomed
back to her natal clan.

26
adult (15-50) member in approximawa 15 percent of households. One or more adults are

observed leaving about 20 percent of households between each interview of the panel. In
addition to the movement of individuals, households are observed lending or receiving
money or goods as loans or gifts. Moreover, there is evidence in these data that a portion
of funeral expenses in households experiencing a death were financed by remittances sent
into the households for that specific purpose. In addition, households reported sending
remittance to other households to assist in funeral expenses. Thus, the event of non-
household member deaths as well as household member deaths may affect changes in
current-period full-income.

We can extend the full-income effect of an adult death to include an impact on
farm profits if we drop the assumption that effective hired labor and family labor are not
perfect substitutes. If farm production and consumption decisions are not separable, then
farm profit will also depend on the health status of household members.

The death of an adult member also implies the loss of some of the managerial
capacity or stock of experience available to the households in farming. If no market
exists for these farm inputs, then separability also fails and, consequently, farm
productivity and profits can be negatively affected. Rosenzweig and Wolpin (1985)
present empirical evidence from India which suggests that farm-specific, experientially
obtained knowledge in rural farm households contributes to agricultural profits,

While analysis of individual time allocation can be informative about differential
outcomes across groups in households, it cannot be used to infer changes in activities at

the household level. As noted earlier, changes in the household’s portfolio of farming

27
will not be affected by demographic changes if farming decisions are separable from

consumption and leisure choices. Even if hours of individuals are not fully adjusting for
lost labor supply incurred by a death, cash income or farm output may be maintained if
households pull in new household members or, in the case of farming, hire more labor,
on-farm. Therefore, issues such as the impact of deaths on agricultural output need to be
addressed by examining variables at the household level. Changes in farming patterns at
the household level associated with a prime-age death would suggest that farming and

consumption decisions are not recursive.

IV.Data'

Data Setting

The data for this study are drawn from a research project conducted by the World
Bank and the University of Dar es Salaam in the region of Kagera in Northwest Tanzania.
The explicit objectives of the project were to measure the economic impact of fatal illness
in the region and to propose cost-effective strategies to help survivors. This study was
originally undertaken as one part of this multifaceted project. Other areas researched
by this project include analyzing the impact of adult mortality on various outcomes such
as household composition, child nutrition, and consumption. Figure 1 presents a map
of Tanzania and the Kagera region.

Tanzania and a host of other countries generally clustered around lake Victoria

(for example, Uganda, Kenya, Rwanda, Malawi and Burundi) are some of the most

28
seriously HIV/AIDS affected countries in the world. 18 Circa 1990, HIV seroprevalence

among 15-49 year olds in Tanzanian urban and rural areas was 8.9 and 5.4 percent,
respectively (Over and Piot, 1993). Kagera is estimated to be one of the regions in
Tanzania most affected by the HIV/AIDS epidemic (World Bank, 1992). Across the six
districts in Kagera, there is substantial variation in HIV prevalence and adult mortality.
Population based estimates from 1987 for 15-54 year olds identify the most affected
district as Bukoba town, the regional capital of and only urban area in Kagera, with HIV
sero-prevalence levels at 24.2 percent (Killewo et al. 1990).19 Bukoba rural and Muleba
have lower, although quite high, sero-prevalence levels, both at 10 percent. Karagwe
observed a sero—prevalence rate of 4 percent. The least affected districts, and most rural,
were Biharamulo and N gara (both under 1 percent).

Killewo et al. (1994) describe the socio—geographical patterns of HIV
transmission in Kagera region mainly in terms of distance of wards (small geographical
areas within districts) by road from Bukoba town. They conclude that the spread of
infection stems from urban areas through which two main roads to Uganda pass. On
the other hand, the truck route that passes through Biharamulo and Ngara from Zaire

via Rwanda and the frequent crossing of people between Ngara and neighboring

 

'8 It should be noted that in Africa HIV/AIDS is primarily contracted through heterosexual contact.

'9 Killewo et al. (1990) employed a multistage cluster sampling technique to generate regionally
representative seroprevalence estimates for the population of Kagera. As noted in National Research
Council (1996), information on national or regional representative seroprevalence levels in Africa is
rarely available. Seroprevalence studies are more widely available for subgroups with the highest risk of
HIV infection (i.e. commercial sex workers or patients receiving treatment for sexually transmitted
diseases). Other studies focus on HIV incidence among low-risk populations such as pregnant women
seeking prenatal care and blood donors, both of which are presumably not representative of the overall
population. Bongaarts (1996), in fact, reports HIV prevalence for only urban populations in Africa (circa
1993) for lack of comparable data for ma] areas.

29
Burundi have not spread HIV in these districts. Despite having much less contact with

the neighboring country to the west (Rwanda) than the other border district of Ngara,
the Karagwe district has a notably higher rate of infection.

One could argue that to best estimate the impact of adult mortality you would
want to examine households in an area with a low incidence of HIV/AIDS so that the
event is unanticipated for all households. Despite the fact that HIV levels are very high
in Kagera, the probability of a prime-age adult death is still quite low. According to
the 1988 Census, the annual number of prime-age deaths per 1000 was 15 in the most
affected districts of Bukoba rural and Bukoba urban. If households do anticipate the
event of a prime-age adult death even before the onset of illness we may see more
muted responses in time allocation than we would otherwise observe if the event (death
and the preceding illness) was unanticipated. This raises an alternative issue which this
study does not focus on: household responses to the perception of risk of an adult

death.20

Kagera Health and Development Survey
The Kagera Health and Development Survey (KHDS) surveyed over 800

households in the Kagera region four times from 1991-1994 with an average interval

 

2° Households in the Kagera region do appear to participate in at least one type of risk-sharing activity:
indigenous/ informal savings and emergency associations, called “bujuni”. These organizations,
organized and run predominantly by women, appear to serve as insurance against a variety of potential
crises. They also fund income generating activities for members. Further evidence of risk-sharing can
be inferred from remittances across households, such as those tied to funeral expenditures (and other
expenditures tied to illnesses).

30

21,22 . . . “
These four mtervrews Wlll be referred to as waves”

. between surveys of 6-7 months.
1—4. The sample is a random stratified set of households where households thought to
have a high risk of an adult death were over-sampled. Households are drawn from 51
communities, mostly villages (also referred to as clusters), in the six districts of Kagera.
Appendix 1 details the sampling strategy and household attrition during the panel; this
analysis does not exclude households that do not complete the panel. The KHDS has
information on time use and other individual and household characteristics, including
detailed data on past deaths of household members and non-member relatives. In
addition, data on physical infrastructure, assistance programs, schools, health facilities
and commodity prices in the community were collected in each wave.

The panel nature of the data has advantages over cross-sectional data. It allows
for the test and potential control for time-invariant, unobservable variables. In addition,
for households with deaths during the panel, we have information before and after death
occurs. This will enable us to examine effects of a future death, without relying on
retrospective data about behaviors individuals or households engaged in during the time
preceding a death.

In the KHDS questionnaire, data on participation in and time allocation over the

last 7 days to several activities were collected for household members seven years and

 

2' For further description of the project and data see Ainsworth et al. (1992) and World Bank (1993).

22 A key component of this analysis is response of individuals labeled as household members. There are
numerous ways to define household membership. In this study, the criterion for household membership
was living in the same dwelling as the head and eating meals with the head for at least three of the 12
months preceding wave 1. Recent joiners, individuals who were not residing in the households in the
previous interview, were considered household members if they intended to stay in the household until
the next interview.

31
older. Individuals were also asked about retrospective participation in activities covering

the last 12 or 6 months.23 Participation describes the supply of labor in the last 7 days
and in the last 12/6 months. These activities are defined as follows:

- Wage employment: primary and secondary employment hours working for a
firm, the government or an individual other than a fellow household member. This would
include wage employment in agriculture.

- N onfarm self-emploment: self-employment hours working for yourself or
someone in the household, such as trader/merchant and including fishermen and
excluding self-employment in farming or livestock activities.

° Farm self-emplomnt: hours supplied to a farm owned by someone in the
household. The hours spent in the last 7 days were collected over several agricultural
activities: work on the household's shamba(s)/garden(s); work on collective land for the
community; processing crop output for sale; tending livestock; and processing animal
products for sale. Allocation of time to specific tasks in farming (such a weeding,
planting or harvesting) is not available.

- Home production: hours in various other activities aggregated into one measure
of home production hours: fetching water, collecting firewood, preparing meals, cleaning

house, and caring for ill household members.

Individual lime Allocation Patterns

 

23 Henceforth, 12/6 will be used to denote the reference periods in waves 1-4. The reference period in
wave 1 is the 12 months preceding the interview date. For the following waves, the reference period is the
time between interviews which is generally between 6-7 months.

32
Participation rates across three main economic activities by age and gender for the

last 7 days are presented in Figure 2. These are locally weighted scatterplot smoothing
(“lowess”) curves for participation rates (Cleveland, 1979). The sample used is based
on the first wave of data collected in late 1991 and early 1992.

Overall, 70 percent of the sample reported working last week in one of the three
areas of economic activity: wage-employment, nonfarm self-employment and farming.
Prime-age adults (15-50 years) had the highest levels of participation in economic
activities in the last week and over the past year. About 80 percent of prime—age men and
women reported working in one of these areas last week. Over the last year, 18 percent
of all individuals reported no work; unemployment over the last year is 6 and 5 percent
for prime-age men and women (15-50) respectively. The vast majority of individuals
who were engaged in at least one economic activity (wage employment, nonfarm self-
employed, or farm self—employed) in the last year reported farming as the area of their
main job in the past year (93 percent of prime-age women and 77 percent of prime-age
men). Only a small fraction of adults, predominantly men, did not participate in farming
in the past year.

At least two-thirds of the entire sample in any wave of the survey had farmed in
the last 7 days. As an indicator of participation in agricultural activities, this figure
provides a lower bound, given that the category "wage employment" includes farming or
working in other agricultural activities for others.“ Among wage earners, however, only

a quarter reported working in farming.

 

2" Kagera is predominantly made up of small-holder coffee-banana farms and observes two cropping
seasons, although tree crops and cassava have fairly continuous cultivation over the year. Although bananas

33
Agriculture in sub-Saharan Africa, as in other developing regions, typically

displays a gender division of labor in agricultural tasks and across crops. In the KHDS
sample we find that a significantly higher percent of prime-age women reported farming
the main food crops (beans, maize, cassava, potatoes, yams, groundnuts and sorghum) in
the last 7 days than prime-age men. Although, farming of the main tree crops (bananas
and coffee) and even sugar cane in the last week was not reported more often by men.
Tibaijuka’s (1984) earlier 1982-1983 survey of Kagera finds a similar pattern of the
gender division of labor, where women are involved in almost all activities. Including
cash crop farming, with the exception of livestock rearing, her study finds that women
generally exhibit higher hours in various farm activities than men.

Average weekly hours in farming among the KHDS sample are low compared to
Western standards. Individual farm hours per day worked in the last week averaged just
over 4 hours among prime-age men and women. Of course, daily average hours
computed for all persons (including those with zero hours) for the last week are lower,

well under 3 hours for prime-age adults. While this may suggest that farmers are

 

are typically considered a food crop, they are becoming increasingly commercialized. Almost all the
households in the sample from the KHDS had some land holdings (93 percent). Average landholdings of
households was 4.2 acres while median acreage was 2.8. In addition, on average households used .5 acres
of land owned by persons outside the household (for which no rent payments in cash or kind were made to
the owner(s) in the vast majority of cases). Among specific farm tasks, less than 4 percent of persons
reported farming community plots or processing either crop output or animal products in the last 7 days.
While Collier's et al. (1990) study of a national sample of rural Tanzanian households from 1980 does report
that most of the labor force was primarily involved in farming, the study also found that the time on the
communal shamba represented approximately 15 to 21 percent of total work time in rural villages. This is a
drastically higher share than found in this sample of households from Kagera. Furthermore, a negligible
number of households in the KHDS reported owning any land jointly with other households. This difference
may be in part explained in a historical context by Ujamaa villages and the process of collectivization during
the period 1967-1980. Other evidence suggests that communal farming even ten years before the KHDS
was rare in Kagera. Tibaijuka’s (1984) 1982-1983 survey of agricultural households in Kagera identified all
tree-crop land as under clan or individual ownership. In almost all villages, communal tenure was found
only on marginal “rweya” land, the surrounding grassland.

34
underemployed, low average daily hours alone are not necessarily evidence of

underemployment. Cleave (1974) surveys a large numbers of smdies of agriculture in
Sub-Saharan Africa and notes that seasoml peak times may be associated with much
higher time spent of agriculture and competing demands for time in nonagricultural
activities also need to be considered. In Kagera, seasonal labor demands are less
pronounced for farms with permanent banana and coffee holdings.

A large number (20%) of prime-age adults reported no participation in the last
week across the set of economic activities (wage, nonfarm and on-farm employment).
Nearly a third (31%) of prime-age adults reported illness as the primary reason for not
participating. However, for the entire sample of the prime-age adults, those reporting a
chronic or acute condition (43 % of all adults) are not less likely to have worked in the last

week.

Household Activities

The households sampled by the KHDS clearly have strong ties to agriculture. In
wave 1, three-quarters of households report having at least one adult member engaged in
farming in the last 7 days; over the past year, 98 percent reported having at least one
adult engaged in farming on household lands. At the same the, a large share of
households had members engaged in other income-generating activities. Forty-three
percent had a member employed for wages, while 30 percent had a member in a nonfarm
self-employment activity in the last year. Sixty percent of households had a member

participating in either wage employment and/or nonfarm self-employment. Only 9

35
percent reported only farming as an income source among wage employment, farm self-

employment, nonfarm self—employment and remittances from individuals.

Households also hired labor for farming activities. In every wave of the survey,
roughly one-quarter to one-third of households reported hiring labor in the last 12/ 6
months.25 With the exception of manure (which was used but typically not purchased),
few farms used other non-labor farm inputs such as pesticide or fertilizer in farm

production.

Reporting of Prime-Age Deaths

Households report deaths for the 24—month period prior to wave 1 and in the
intervals between waves for two groups of people: household members and non-
coresident relatives. The definition of household membership of the deceased was left
to the respondent for those deaths reported in wave 1. During the panel, the deceased
was considered a household member if he/ she had qualified as a member in the
previous wave or had been residing in the household at the time of death. So, some
people who moved into the household just prior to death were considered to be
household members.

Almost all adult deaths (at least 95 percent) were attributed to illness, rather
than accident or injury, childbirth, suicide or homicide. The average length of illness

was 12 months. Cause of death was reported by households and, if diagnosed,

 

25 The KHDS does not contain detailed information on use of hired labor on farms. Tibaijuka (1984)
found that hired labor in Kagera is used for weeding and harvesting coffee sometimes, but only on larger
farms. Otherwise, she found that hired labor is mainly employed in the strenuous task of clearing land
for establishment of perennial crops.

36
confirmed by a medical practitioner. About half of deaths were reported to be caused

by AIDS; another third were cited as unspecified or unknown illness. Other illnesses
cited as cause of death include: asthma, cancer, diarrhea, dysentery, gonorrhea,
parasite, malaria, malnutrition, measles, meningitis, poison, syphilis, tetanus, and
tuberculosis.

Without reliable diagnosis by medical practitioners (or, more specifically,
laboratory-confirmed diagnosis), it is difficult, if not impossible, to infer the true
proportion of deaths due to HIV / AIDS. Households may incorrectly identify the cause
of death because of lack of information, misconceptions and stigma surrounding
HIV/AIDS. In any case, this study is concerned with the event of an adult death, and
not isolating deaths due to AIDS from other illnesses that afflict prime-age adults.”5

The KHDS relies on households to accurately report information about deaths
experienced therein and was not designed to assess the consistency in the reporting of
deaths by households. While the panel does not have overlapping recall periods, the
enumeration data collected in the baseline, which was fielded about 8 months prior to
wave 1, can be used for comparison with the wave 1 data. The enumeration (baseline)
survey consisted of interviewing every household in the 51 villages selected for
participation and collecting data on prime-age deaths in the last 12 months. The sample
of 838 households was then drawn from the entire sample of enumerated households

(see Appendix 1). About 7-8 months later, in wave 1, those households selected report

 

26 As mentioned in the previous section, it is important that we believe that deaths were, in most cases,
not caused by potentially curable illnesses. Otherwise, after contraction of a disease which itself may be
an endogenous occurrence, the consequence of death would be endogenous to household and individual
health care decisions.

37
prime-age deaths in the last 24 months. Therefore, we would expect all deaths reported

at enumeration to be re-reported in wave 1 if we believe that the survey instruments
which are not identical are adequately similar. Comparison of these data revealed
marked underreporting of deaths in wave 1. These recall errors are hypothesized to be
primarily linked to membership classification and dating errors in reporting of deaths
by households. In the former case, households may be re-classifying deaths between
household membership and non-membership. 27 This is of concern because the analysis
takes into account both the classification of deaths (member versus non-member
relative) and the date of death. In a more general context, this highlights concerns
about collecting retrospective data even when the event is something as serious as a
death in the household.28

About a third of adult deaths during the panel (after wave 1) are individuals who
joined the household between interviews. Thus, they resided in the household for less
than 7 months before death which may indicate that these individuals entered the
household specifically because they were seriously ill. Some of the reasons these

individuals would go to these households include seeking terminal care and saving the

 

27 One would expect that the problem of memory effects plagues reporting on minor events, such as
expenditure for minor items (see Deaton, 1997), as opposed to seemingly major events such as deaths.
Thus it is presumed that deaths reported at enumeration are not forgotten by wave 1, but rather,
inaccurately dated or reported as non-household members.

28 This conclusion to this finding suggests not separating deaths of household members from non-
coresidents, but rather modeling responses to either type of reported death. The number of non-
coresident deaths reported is large and includes distant and closer relations. Furthermore, it is possible
that respondents selectively report non-resident relatives who died from a larger set of deceased non-
coresident relatives.

38

29 This raises the concern that the event of an adult

family the cost of shipping a body.
death in a household is not exogenous if fatally ill individuals are selecting households in
which they will die.30 It is the case that all of these persons were related to someone in
the household and most were close relatives, primarily with a parent in the household and

also siblings or a spouse.

V. Empirical Specification

Time Allocation

For estimation of individual time allocation equations, the sample has been divided
into 6 groups: children (7-14), young women (15-19), young men (15-19), adult women
(20-50), adult men (20-50) and older adults (51+ years). Young adults are separated
from other prime-age adults (20-50) mainly because school participation is still a decision
for these individuals. In fact, 66 percent had been enrolled in the past 12 months in wave
1. While this would seemingly reflect secondary-level enrollment, due to repetition and
late entry, 3 number are in fact still at the primary level (grade 7 or less) despite their age.

Two equations are estimated for time allocated to activity j by individual i in

period t as derived from equation 7:

 

29 In a 1990 study in Abidjan, Cote d’Ivoire, research suggested that AIDS-related deaths were probably
understated due to the desire of seriously ill persons to die in their home areas (National Research
Council, 1996).

, 3° This point alludes to the notion that contracting HIV/AIDS is also not exogenous but the result of
distinct patterns of behavior, particularly with respect to sexual activity. This is of concern because
potentially the observed “impact” of a death in the household may be identifying the impact of other
,characteristics of the individual or household, not controlled for in other right-hand side variables. This
is addressed below.

39

(8) m=l30+l31vht+ BZXit+ BBth+B4Dht+Di+Uhh+8it
and:
(9) (Tjit ' Tjit.l )= 90 + 91 Vht + 92 Dht + 8it

V is a vector of village by wave dummy variables to capture all village-level price
and infrastructure effects. V also includes dummy variables for the month and the year of
interview for seasonal effects and overall changes in regional prices across years.

X is a vector of individual characteristics including age, education and headship to
capture the individual’s marginal productivity on farm and proxy for individual wage
rate.31 For children and young adults (15-19 years), education and headship are
excluded, as these individuals may still be in school and very few are household heads.
For these samples, some parental information is included describing orphan status,
presence of parents in the household, education of parents, and status as child of the
household head.

Household production characteristics (components of full-income), which include
characteristics of the household head (age, sex, and education) and a measure of assets
(inherited land), are in the vector Z. Land purchased with either cash or credit, which is

likely to be endogenous with respect to past labor supply decisions, is excluded from the

 

3 ' Village-by-wave interactions are presumed to pick up market wage effects. Wages as an explanatory
variable were examined, but are not included in the estimation for several reasons. First, the wages
reported for various agricultural tasks in the community questionnaire were missing numerous observations
and showed little variation across adult males and females. Given the low numbers of women with wage
employment and probable gender and wage differentiation by agricultural task, it seems dubious to apply one
community wage rate to all adult men and women. Because of the low numbers of individuals working for
wage employment in the sample, calculation of cluster averages of reported wages or predicted wages from
individuals in the sample of households would have to be based on very low numbers of observations.
Measures of marginal productivity or shadow wages, which are probably a better estimate of productivity or
the value of time given the high fraction of adults who are farmers, cannot be adequately estimated from the
data.

40

land measure. Household composition variables, which are also believed to be
endogenous, will be excluded from the right-hand-side. The age and sex composition
among surviving individuals within the households is posited to be influenced by the

32 Variables describing health status of

occurrence of adult deaths and other covariates.
individuals and other household members are also excluded because of problems with the
self-reported nature of the data (with the exception of weight and height measures) and the
simultaneity of current health status and time allocation.

D is a set of variables describing deaths, explained in detail below. The terms U,
and um, are individual and household time-invariant unobservables, respectively. The
“typical” error component, which is assumed to be independently and identically
distributed with mean 0 and variance 02,, is an.

In equation 8, the set of variables describing deaths, D, will measure the effect on
levels of hours in activity j for individuals in households with a particular death shock
(specified as, for example, a past/ future or male/female death). In equation 9, the set of
variables describing deaths, D, assess the differential in the change in hours across
periods associated with deaths. A potential concern in regards to equation 9 is timing of
deaths where deaths occurring before t-1 will affect Tim as well as Tj,,. Even if deaths

occur between t and t-1, both are affected if future deaths influence hours or participation.

In these cases, the vector D describes changes observed in households with versus those

 

32 Other analysis of household composition in the KHDS sample showed a high level of turnover of
household members of all ages and revealed a significant impact before and after the death of a member on
the probability of another adult joining or leaving the household (Ainsworth et al. , 1995). This issue is
addressed in more detail below. See Rosenzweig (1988) for a discussion of the supposition that household
composition responds to the economic environment facing households.

41
without deaths over time although it cannot necessarily be used to infer differences in

levels of hours or participation for individuals in households experiencing a death to those
with no deaths.33

D is a vector of variables describing the future and past deaths of adults 15-50
years who are household members. Variables describing future deaths capture two
simultaneous occurrences. Future death variables should identify households with very ill
household members, which may influence time allocation - regardless of the household’s
expectation or knowledge of the impending death. At the same time, households may be
adjusting activities, not only to account for a potentially ill household member, but in
preparation for an anticipated death. The future death variables, therefore, identify an ex
' ante death coping strategy an_d_ an ex post illness response. If it is the case that households
cannot (or do not) accurately anticipate deaths, then the future deaths variables will
capture only the latter effect. Past death variables describe ex post responses to factors
such as income shock associated with funeral expenses and shock to potential stock of
labor in the household.

While this model motivates the inclusion of death as an economic shock, among
other covariates, it does not lead us towards any particular specification of the effect of
death. The shock of an adult death has several dimensions that need to be captured: the

discrete event itself, the time to and from the event and characteristics of the deceased.

 

33 For example, suppose individuals in households increase hours in the months preceding a death while
following the death hours retum to the level which would be observed had no death occurred (some
“normal” amount of hours). Households with a death will exhibit a decrease in hours compared to those
with no death, although levels in hours after the death will not be different, ceteris paribus.

42
Several alternative depictions of this impact were explored and the results of one of these

are presented. 34

The death variables identify households in which at least one person died while
residing in the household.35 This includes individuals who joined the household and died
shortly thereafter. These new arrivals are, in some cases, individuals who joined the
household and died between interviews - residing in the household less than 6 months.
As noted above, about a third of all deaths during the panel are such cases; we cannot
identify these cases for deaths preceding wave 1. Ideally, we would like to separate
deaths of household members into those that were “long-term” members and the new
arrivals, but small cell sizes in this larger, more descriptive set of death variables restrict
us to grouping them together. Whereas some individuals move into the household shortly
before death, it is also possible that some of the non-member relative deaths reported are

individuals who recently exited the household, although these cases cannot be identified.

 

3" The various alternative specifications examined included, for example, variations in an Almon
polynomial in the months since/to an adult death (see Greene, 1990) and dummy variables for less
aggregated monthly groupings (such as single month intervals and 3 month intervals). Characteristics
such as education of the deceased could also be introduced. The small number of deaths, especially once
prime-age deaths are segregated by sex, makes obtaining results from a more descriptive set of variables
increasingly difficult.

35 Household members need not be related to other household members, although in most cases the
deceased is closely related to the head or otherwise has a close relative in the household (such as a child,
spouse, parent or sibling). Originally, this analysis included deaths of relatives who were reported as
living elsewhere divided into two categories: close relatives and distant relatives. Close relatives
included any individuals identified as having a child, sibling, or parent in the household or being related
to the head as a spouse, grandchild, grandparent, niece/nephew, or in-law (child, parent, or sibling).
Distant relatives included any other relative of the head or spouse of the head not aforementioned.
Generally the deaths of adult relatives residing outside the household were not significantly associated
with changes in participation or hours. These measures of non-household member deaths are problematic
in that we can not differentiate those deaths that are more “economically” significant from the full sample
of deceased relatives. Furthermore, the full set of variables describing death may be multicollinear. For
example, households recently experiencing a death may have better information received during funeral
and mourning ceremonies about deaths of non-member relatives. This does not appear to be the case as
the results presented below are strikingly similar to those estimated which included the corresponding
variables describing deaths of relatives who are non-household members.

43
A set of variables are included to identify any male or female prime-age (15-50)

death in the household in the future and past between certain time intervals from the
interview date. The set of time intervals includes deaths occurring within 12 months from
the interview date.36 This interval is subsequently divided into two periods to infer
differences of the impact of death depending on distance of death from the interview date:
0-6 months and 7-12 months from interview date.37 Each dummy variable is 1 if such a
death occurred and 0 otherwise. The data on deaths used from the panel is collected from I
7 months preceding wave 1 through the deaths reported in wave 4. As noted earlier in
the text, the deaths reported over the 24 month recall period in wave 1 are suspected to
be reported with measurement error in both timing and classification of deaths as
members versus non-members of the household (See Chapter 2 for a more detailed
analysis of these potential recall errors). To circumvent this problem, the wave 1 data
on deaths has b-eken ”truncated to include. those deaths reported in under 7 months of
wave 1. Although there is no way to infer at what length in the recall period the data
become inconsistent, a 7 month retrospective for wave 1 approximately matches the
retrospective period for waves 2-4 -- the time between households surveys, roughly 7

months on average. Nevertheless, in one sense the data collected in waves 24 may be

more reliable. The surveys in waves 24 elicit information about the status of all

 

36 This categorization will not distinguish between households never experiencing a death and those with a
death outside the 12 months to/ from the interview data. Thus, the control group is not households that do
not experience a death, but rather those that had no death within a year.

37 Few deaths occurred within 2 weeks of the interview date, where we would expect to observe periods
of mourning and other flmeral activities. Thus it is presumed that this approach of categorizing deaths in
periods covering 0-6, 7-12 and 0-12 months in the past will not be capturing these types of short-run
activities associated with deaths.

44

individuals previously recorded as household members in the last interview, some of
whom may have died. In wave 1, there is no roster to infer membership in the
household 7 months in the past.

Only the most recent death for each category is identified and so the dummy
variables do not take into account multiple deaths within a household over the specified
time interval. This is rarely observed given the classification of deaths by sex, limitation
to deaths of household members and truncation of deaths prior to 7 months preceding
wave 1.38

The panel has information on deaths over a period of up to 28 months for those
households that completed all four waves of the panel. To capture the censoring of
observed future and past deaths, variables were included to identify the length the panel
ahead and in the past for each household. For a household ' observed in wave 4, the
survey has no information on future deaths, despite appearing to signify the household as
having no future deaths. Therefore, dummy variables to identify the last interview and .
second-to—last interview, where future death variables are truncated, are included. This is
not the equivalent to identifying the first, second or third interview because the sample
will include individuals in households that do not complete the panel. Likewise, a
variable to identify the first interview, where no deaths are observed 7-12 months past, is

included.

 

3‘ The 157 deaths of prime-age adults occurred in 140 households. Five households had two deaths of
different sexes and seven had two deaths of the same sex. One household reported three deaths and one
reported four not of the same sex. The majority of these deaths occurred at least two waves apart, so
most of the future and past death variables are accurately capturing the number of deaths (one) within
category.

45
The activities that are examined are farm heurs, home hours, wage-employment

participation, and self-employment nonfarm participation. The participation equations are
estimated for prime-age men, given the very low levels of participation for other
demographic groups and the high degree of censoring in hours fer wage employment and
nonfarm self-employment.

Table 1 presents a description of the original samples of individuals and the sub-
sample used for estimation. The original sample of individuals initially includes persons
that subsequently die over the course of the panel (labeled “N on—survivors”). This is not
the full sample of deaths, but rather those deaths where the deceased was surveyed at least
once before dying. This means that the category “N on-survivors” excludes all household
member deaths from wave 1 and those individuals who arrived into and died in the
household between waves of the survey (so-called short-term non-survivors). Non-
survivors reported in Table 1, persons who were observed in one wave and subsequently
die during the panel, have been excluded from the samples used as this analysis does not
want to confound the impact of an individual's future death or current illness on their own

current time use with the impact of a past or future adult death on other '3 time use.

Household-Level Activities
level farm activities are also examined; these include indicators of labor hiring, area of
land cultivated, area cultivated per adult 15 + years resident, and the farming of coffee,

banana and other food crops. In addition to these changes in farming, the probability

46
that households had an adult joining the household is also explored.”'40 Joiners do not

include non-survivors (i.e. individuals who die during the course of the panel).

Analysis of hired labor is limited due to lack of detailed information about farm
inputs collected in the KHDS. The data pertaining to hired labor include only total
man-days hired and total amount paid to laborers for farm activities over the last 12/6
months, both suspected of being very noisy measures of hired labor usage. Analysis of
these measures as well as an indicator of any hired labor was restricted to waves 2-4
because in wave 1 the past death variables (only past 0-6 months in wave 1) do not
precede the recall period of hired labor (12 months). Like hired labor, area of land
cultivated refers to the preceding 12/6 month period, and therefore excludes wave 1.

Area of land cultivated by the household is reported by the household respondent and

 

39 The majority of individuals who migrate into households during the panel come from within Kagera
region. Among all joiners, about half are children and half are prime-age adults; there were few joiners
over the age of 50. Women make up about two-thirds of all prime-age adults joining households.
Almost all individuals who identified marriage or divorce as the reason for immigration were prime-age
women. “Other family problems” and “Other reasons” account for the largest share of reason for in-
migration. See Ainsworth et al. (1995) for a more detailed description of in and out migration over the
KHDS panel.
40 As noted above, Ainsworth et al. (1995) also examine changes in composition in response to deaths
experienced by households. The analysis here differs in several significant ways, aside from various
differences in the set of explanatory variables used. First, Ainsworth et al. do not separate male and
female deaths. This may be an important consideration as the results for farm hours show that responses
vary by sex of the deceased. Second, they include a set of variables that includes a dummy variable if a
death occurred and then a variable for the months since/to the death (as well as months squared and
cubed), where households with no death have these variables set to zero. Their past death variables
include those that occurred in the past 6 months. Thus, with respect to past deaths, the event of a leaver
and joiner can not be confirmed as necessarily occurring after the death since both events occurred during
the span of some 6 month period. (While date of death is documented, date that an individual joined the
household is not.) Ainsworth et al. point this out, but they conclude that their results simulating
probabilities past these 6-7 months are not problematic. The coefficients used to derive these simulated
probabilities were estimated from the data that include the aforementioned data covering the 6 month
preceding the interview. Thirdly, unlike Ainsworth et al. , I truncate the retrospective data from wave 1.
As a result, I have significantly fewer past deaths in my sample. Finally, their analysis did not include
controls for the censoring of future deaths as well as the variability of the length of past and future death
observations across waves. Although they examine probability of adults joining and leaving the
households, the analysis in this study is restricted to adults joining the household. Probability of an adult
leaving was examined but not found to be significantly associated with adult male or female deaths.

47

not measured by the survey team. Area of land cultivated is a measure of the
extensiveness- of farming while area of land cultivate per resident adult (15 +) describes
the intensity of farming. These short-run measures of cultivation may not be very
responsive to shocks given the predominance of tree-crops in Kagera and analysis of
changes in number of trees under cultivation (which is available) may need to be
examined under a longer panel.

Theflprobabilities of any household member reporting farming coffee, banana
and other food crops. in the last seven days, potentially more responsive variables than
land cultivated, are also examined.41 Other food crops includes any member farming
maize, cassava and beans. Although these probabilities cannot reveal information about
output or yields, they may indicate the extent to which crops, tree crops in particular,
are neglected as a consequence of an adult death.

The estimations of these household activities include the same set of household
characteristics as the individual time allocation regressions for equation 8:
characteristics of the household head (age, sex and education), acres of inherited land,
and month and year of interview. Variables to identify number of past interviews as
well as cluster-by-wave interactions are also included. The latter set of interactions
should capture spatial and temporal variation in, for example, weather and prices that

may influence farming decisions.

Estimation Techniques

 

4' Area of land cultivated is reported for all crops. The data do not have measures of area cultivated for
specific crops, which would be difficult to assess given the high degree of intercropping of plots.

48
The analysis of time allocation is confounded by two econometric issues:

censoring of the dependent variable and potential omitted variable bias. Censoring is of
greater concern in the analysis of wage employment and nonfarm self-employment where
a large share of the sample do not participate and, thus, a large fraction of the sample
have hours censored at zero. Under censoring, Ordinary Least Squares (OLS) estimates
will be biased; this problem is typically addressed using a Tobit model.42 For analysis of
participation decisions, probit estimation can be used with a binary dependent variable
which is zero for non-participants and 1 for participants.

The omitted time-invariant variables u, and um, may result in bias of OLS
estimated coefficients. These unobservables may be a “random effect” in which case the
heterogeneity is distributed randomly across cross-sections and is not correlated with
other explanatory variables. Random effects can be estimated efficiently using a
Generalized Least Squares (GLS) approach. On the other hand, if the time-invariant
heterogeneity is correlated with other explanatory variables (either at the household level
or among individuals), then the random effects estimator will produce biased and
inefficient estimates. The fixed effect approach will control for this form of time-
in'variant heterogeneity.

Household unobservables in particular may be correlated with the death variables

as follows. Death occurs if health, which is a function of household unobservables and

 

42 The Tobit model does not necessarily address self-selection by individuals into activities. Further,
Tobit estimation restricts the parameters on participation in the activity to be equal to those determining
hours allocated to the activity (see Greene, 1990). These problems could be addressed using Heckrnan’s
two-step estimation. Ideally one would want identifying variables to appear in the participation equation
but not in the hours equation. Such instruments in this context are difficult to identify.

49

the non-survivors’ individual unobservable component (see equation 4), falls below some
threshold level. If all deaths were caused by fatal illnesses randomly distributed in a
population, then we would not suspect the set of death variables to be correlated with
households unobservables or the unobservables of survivors. The presumption is that a
significant portion of deaths are caused by AIDS, a disease typically contracted through
distinct patterns of behavior.43 Conditional on household unobservables, the individual
specific unobservables of survivors would not seemingly be correlated with the variables
describing deaths of others which are a function of household and the non-survivor’s
individual effects. Controlling for individual heterogeneity will address households
unobservables in addition to survivors’ individual time invariant characteristics that may
bias OLS estimates such as ability and motivation.

The fixed effects approach also addresses the stratified choice based sampling
employed to construct the KHDS sample where households that had already
experienced as death or had sick adults were much more likely to be surveyed (see
Appendix 1). In this analysis, the stratification criterion, those with and without any
sickness or past death, is not included in the estimation equations although the concern
is that it is correlated with the right-hand-side variables indicating death. That is, the
stratification presumably emphasized households that were more likely to experience a
death after the enumeration survey. Where the selection variables are right-hand-side

variables (and there is random sampling within strata), Wooldridge (1997) shows

 

43 Philipson and Posner (1995) consider the rational choice individuals make to engage in risky sexual
activities (such as unprotected sex and multiple sex partners) that influence the contraction and spread of
HIV/AIDS and other sexually transmitted diseases in sub-Saharan Africa.

50
unweighted estimators are better than weighting in specific cases including fixed effects

models.

Presence of individual specific unobservables will be tested using the Breusch-
Pagan Lagrange multiplier test in the farm and home hours equations, which have high
participation rates (less censoring) across the 6 demographic groups. The Breusch-Pagan
test compares the random effects estimates with those produced by OLS. Hausman tests
are estimated for all variables across the random and fixed effects estimations, as well as
for the subset of variables describing deaths in the household. For the participation
equations, the probit model is estimated with Huber-adjusted standard errors for

clustering at the household level.44

VI. Results

The means and standard deviations of the samples are presented in Table 2.45
Results for response of hours in farming as well as house hours to past and future deaths
are presented in Tables 3a-c for any death in the past/ future 0—12 months. Tables 4a-c
present results for deaths in the past/future 0—6 and 7-12 months. In all cases, the
Breusch-Pagan test is rejected suggesting that heterogeneity is present. In most instances,
the Hausman test of the individual fixed versus random effects specification did not reject

the null hypotheses that the individual unobservables are uncorrelated with the set of

 

4" Chamberlain’s (1980) fixed effects logit model (or conditional logit) was estimated as well as. Results
from these estimates exclude men who were never observed entering or leaving the activity at least once
and, thus, samples were significantly smaller.

45 Only . the results for the set of variables describing deaths are included in the tables of the main text;
results for other explanatory variables are presented in Appendix 2 in addition to summary statistics for
variables omitted in Table 2.

51
explanatory variables. The Hausman test on the subset of variables describing the deaths

of household members is rejected in most cases, but not necessarily in every instance
where the overall Hausman test is insignificant.46 The individual fixed effects estimates
are presented in the set of Tables 3 and 4; for comparison, the random effects estimates

are presented in Appendices 3 and 4.47’48

Farm Hours

There no evidence that surviving members work more hours per week on the farm
as a consequence of an adult death in the past or future year. That is, almost all estimates
for level of farming hours in Tables 3a-c are small (under 3 hours) and statistically
insignificant. The only statistically significant estimate is for women 15-19 years, where
female deaths within the future year are associated nearly 6 hours less per week in
farming. With respect to the change in weekly farm hours associated with deaths over
the 0-12 month period, likewise, most estimates are small and statistically insignificant.

Women 20-50 years and children had decreasing hours associated with male deaths in the

 

46 In comparison with the results from this study, Skoufias (1993) examined time allocation in rural India
also using a panel data set and adopted random effects estimates after failing to reject the null hypothesis
of exogeneity of the heterogeneity component. He notes that his results imply that “ . . .OLS estimates
from individual time allocation data in a cross section are likely to be free from omitted variable bias
(p.292-293).” Re-evaluated at a 10 percent level of significance rather than 1 percent half of his 12
estimations reject the Hausman test statistic.

‘7 Fixed effects estimates are the mean-differenced estimates. First-differencing produced similar results.
48 This is an unbalanced panel since some households drop out and individuals enter and exit households
during the panel. In addition, observations for some individuals were not used due to missing values.
Consequently, some individuals are observed only once during the panel; these observations are not
observed in the fixed effects estimations. Random effects and Hausman tests were estimated on both the
full sample and the subsample of individuals with multiple observations. The results of these tests were
generally very similar and the coefficients in the two random effects estimates were very similar.
Furthermore, household fixed effects models were also estimated (within each of the 6 demographic
groups) and produced estimates very similar to the individual fixed effects estimates on the set of death
variables.

52
past year, while women 15-19 had increasing farm hours for past female deaths. In

levels, however, both groups of women did not have significantly different hours in
farming from other households with no such death. Farming hours increased across
waves for men 15-19 and children in households experiencing a future male death with
the former group increasing hours by slightly more than 5 per week. Although the
estimates of levels of hours for men 15-19 is of similar magnitude, it is not statistically
significant.

When the 0-12 month period is divided into two segments, we observe some
significant differences in weekly hours across individuals in households with or without
an adult death (Tables 4a-c). Women 20-50 in households with a male death occurring
sometime in the future 6 months work 6 hours more in farming activities over the last
week compared to women in households with no death. Future male deaths are also
significantly associated with large increases in farming hours across waves for these
women. Consistent with higher hours for women with future male deaths, women with a
recent past male death had significantly declining hours in farming from the previous
interview, about 9 hours, despite having insignificant, albeit lower, hours in farming for a
recent past male death when assessed in levels. This shows that difficulty in using change
in hours across interviews to assess the impact of deaths when these events may have ex
ante and ex post impacts. For adult men male and female adult deaths were not
statistically associated with differentials in levels or changes in hours in farming.

Among younger adults, 15-19, men had 7 more hours in farming in households

with a male death within the upcoming 0-6 months. Female deaths within 0-6 months are

53
associated with 7 lower hours among young women. Children experiencing a male death

had lower hours in farming for deaths 0—6 and 7 -12 months past, consistent with the lower
hours observed for past males death in the overall 0-12 month period.

Among older adults, 51-65, the impact of deaths was very imprecise. Despite
large coefficients, most estimates were not statistically significant. Female deaths in the
future 7-12 months were associated with an almost 10 hour reduction in farming
compared to households with no such death.

Other covariates included in the estimations are very briefly discussed. Most of
the set of individual and household variables are not individually significant, although
jointly, they are in some cases (see Appendix 2, Tables 2a—c and 3a-c). This may because
a large factor in time allocation in farming and home activities in this area is one’s
demographic group which has already been accounted for in the segmentation of the
sample. Young men (15-19) whose father is deceased are working many more hours in
farming per week (about 12). Young women (15-19) that are children of the household
head are working fewer hours (about 10) per week than other young women, perhaps
because they are more likely to advance to secondary levels of schooling. Children in
households with more inherited land report significantly higher weekly hours in farming.

Increased education of the household head is also positively and significantly associated
with farming hours for children. In almost every case, the set of month and year dummy

variables are jointly significant as are the village-by-wave interactions.

Farming Activities of Households

54
Despite a handful of significant differences in farm hours among households with

and without death, the results find little association between prime-age deaths and
adjustments in farming hours. This may imply that overall farm hours and output have
fallen, or that labor on Kagera farms is underemployed in which case output can be
maintained despite the decrease in total hours. The possibility that output has declined
has serious implications for food security for Kagera households where food
consumption is the dominate category of household consumption expenditure and the
majority of food consumption expenditure is on home-produced items. Alternatively,
households might increase hired labor to maintain output. Households might even use
unhired labor. In the short-run, it may be the case that individuals residing in other
households are supplying labor to households with a prime-age death. Over time,
households may even acquire new members to increase the stock of labor. These
possibilities suggests that linkages across households and within villages not only in terms
of cash or in-kind remittances but also with respect to the movement of individuals (in the
short or long term) are potentially important methods for coping with an economic shock
such as an adult death. These issues also point out the complexities of assessing the labor
endowment available to households. If labor resources of households transcend the
composition of the household unit then policy makers may not be able to effectively
assess relative vulnerability of households to the AIDS epidemic in terms of labor
endowment (see Barnett et al., 1995).

While the labor supplied by neighboring households is not measured, we can

assess the movement of new individuals into households in response to an adult death.

55
Table 5 presents results of probit estimates of prime-age adults entering households over

the 6-7 month period from the previous interview. Because these individuals join
sometime over this 6-7 month period, we cannot know if deaths in the same period (0-6
months past) preceded or followed the event of a joiner. Two sets of past death intervals
were estimated: 0-12 and 7-.12 months. In the latter case, we know that deaths preceded
the addition of a new adult resident. Future deaths 0—12 months are included in both
estimations.

Female deaths 7-12 months past are not associated with new members (male or
female), female deaths in the past year are associated with an increased probability of an
adult joining by 10 percentage points. This is a sizable increase from the mean
probability of an adult joining the household between any two waves (around 15 percent).

Male deaths are significantly associated 7-12 months and 0-12 months in the past. The
probability of a prime-age adult joining more than doubles for households with a death in
past 0-12 months. Future deaths are not associated with significant changes in the
probability of adults entering.

Separate regressions for male and female joiners revealed that the effect of a past
male death is significantly associated with an adult female joining the household.49 On the
other hand, adult males were more likely to enter households that had experienced an

adult female death as well as those experiencing a male death in the past 0-12 months.

 

‘9 Because of the infrequency of future deaths and probability of any adult joining the households, future
deaths had to be combined into 0-12 months rather than divided into 0-6 and 7 -12 month periods.
Likewise, because of the low probability of both males joining households and any future male/female
death, estimates on the impact of future deaths could not be produced for the probit analysis of males
joining households in response to a death.

56
Female deaths rather than male deaths had a stronger association with males joining.

Thus, male and female deaths draw new members of the opposite sex into the household
after the event. Members who die are apparently not replaced with members of the same
sex. The focus on the probability of adults joining stems from the presumption that
new prime-age adults are actually participating in labor activities. In the first interview
since joining the household, prime-age joiners had average hours in farm and domestic
activities slightly lower than other individuals. In households with a past prime-age
death, there is no significant difference in hours of adults who joined the household in
the last 6 months and other resident adults already residing in the household (Table 6).
Given the mean predicted probability that a new adult will join after a death (about .3)
and the average hours of new joiners (9), the average household with a past death will
have 3 hours of labor supplied by a new member.

The probability of entrance of new adults is significantly higher in households
with older or better educated household heads for both male and female joiners. Acres of
land, however, was significantly associated with the probability of a prime-age man
joining and not a prime-age woman in spite of the higher levels of farm hours among
adult women.

In addition to replacing the labor stock with new members, households could
offset a death by hiring in more labor to work on the farm. Three indicators of hiring
practices are analyzed and results are presented in Table 7. As in the case of probability
of joiners, our measures of hired labor cover the span of the preceding 6-7 months. Two

sets of past death variables are examined (7 -12 months and 0-12 months). Future death

57
variables have been omitted from these analyses. The probability that households hired

labor in the last 6 months was not associated with a death more than 7 months in the past
or deaths 0—12 months past. Acres of land inherited and years of schooling of the
household head were both significantly associated with a higher probability of hiring of
labor.

Any hired labor may be too crude a measure to capture affects on hired labor.
Fixed effects estimates of mandays of hired labor also showed no association with past
deaths. On the other hand, total expenditure (in nominal Tanzanian shillings) on hired
labor (cash and in-kind) was significantly lower in households experiencing a past female
death.50

The lack of an impact of deaths on labor hiring practices along with the increased
probability of new residents suggest that households prefer to replace labor with new
members rather than hiring in labor. Households suffering a death lose not only farm
labor but also labor in domestic activities (for which a labor market is presumably non-
existent).51

If households are not completely replacing the lost labor of the deceased via
individuals joining the household or hiring labor, we may expect that the extensiveness of
farming will decrease. Indeed, acres of land cultivated in the past 6 months significantly

declined for households with a male death 7 -12 months past by nearly one acre (see Table

 

5° With respect to the household fixed effects estimates of mandays of labor hired and total labor
expenditure, it should be noted that there is a high amount of censoring of the dependent variable at zero
as only a third of households reported hiring any labor.

5 ‘ This hints of the notion that households have some optimal size and composition in terms of permanent
members.

58
8). Intensity of farming, as measured by land cultivated per resident adult (15+), was

not significantly associated with the death of a female or male adult in the household.
Given the preponderance of tree-crops, we would expect that farms do not drastically
alter cultivation on lands to which it has access.52 Acres of land inherited by members of
the household was significantly associated with a greater extensiveness and intensity of
farming (acres cultivated and acres cultivated per adult resident, respectively). As well,
cultivation is positively associated with age of household head, but not the schooling of
the head. Female headed households, about a quarter of all households, cultivated
significantly less acres controlling for acres of land inherited. Intensity of cultivation
was not significantly different for female headed households from that of male headed
households.

If it is the case that total labor hours devoted to the farm do fall after a prime-age
death, then we expect to see output and maybe yields from farms fall, unless the marginal
productivity of the deceased member was very low (i.e. there is under-employment on-
farm or underutilization of labor along the lines of Arthur Lewis' labor surplus model).
This would be contrary to the conclusions of Collier et al. (1990) of a labor shortage in
Tanzanian rural labor markets surmised from the small numbers of individuals looking
for work. Ignoring possible seasonal shortages, labor shortage in the Kagera sample of
households may be unlikely given the low number of hours adults are reporting in
economic activities. Unconditional on numerous factors such as poor health, average

hours do not appear to be constrained in the KHDS. It is interesting to note that lower

 

52 Land areas cultivating only food crops (not including banana) are presumed to take up a small share of
cultivated lands as opposed to lands intercropped with tree-crops and food crops.

59
farm output (not per capita output) is not necessarily a negative outcome given that these

households also have one less member to feed and shelter. However, even if this were
true, decreased production of certain crops (such as export cash crops) may be of concern
to the national economy.

Table 9 presents results for the probability of any household member farming of
various crops in the last 7 days. Households with a male death in the past 0—6 months
showed decreased participation in coffee farming. This suggests that, at least in the
months immediately following a male death, cash cropping may be curtailed in favor of
food crops. Participation in coffee and banana crops is not lower for households
experiencing a death more that 6 months in the past. This may reflect the previous
evidence that male deaths are associated with new adults joining the household over a 6
month period, thereby re-establishing farming of trees within 7 months of the death.
Households with a more distant (7-12 months) death did not show lower participation in
coffee or banana farming but did have significantly lower activity in farming of maize,
cassava and beans. This decline is consistent with the reduction in the number of persons
for which these subsistence, non-commercialized foods are being grown, although they
cannot tell us about outcomes on a per capita level. It is contrary to the perception'that an
adult death will result in a labor crisis and in increased production of subsistence crops
relative to export crops (see, for example, National Research Council, 1996). To what
extent these result holds in the long term is not certain. Barnett et al. (1995) found
evidence from Uganda that households shifted farming away from cash crops over a four

year period.

60

House Hours

Results for hours in various household activities are also included in Tables 3a-c
and Tables 4a-c. The category of house hours includes specific household chore activities
as well as caring for ill household members. These activities make up a large portion of
weekly non-leisure hours, particularly for women, although caring-for-ill hours are a
relatively small portion (for example, 6 percent of hours for women 20-50). In addition
to the loss of potential labor for household chore activities (similar to the loss for
farming), a fatally ill resident requiring care by others places a new demand on other
household members. The extent to which these demands on labor constrain individuals
can explain the lack of increase in farming hours. We might not see increases in farm
hours if survivors have to spend more time in household chore activities as well as caring
for an ill member in the months preceding death.

In levels of hours, men 20-50 are the only group that showed significantly
different hours in house activities for a death within 12 months, with nearly 6 more hours
per week associated with future male deaths. For women 15-19 and 20-50, hours were
decreasing by 5-7 hours per week in response to future male deaths. As households
approached a future male death, women were working less in house activities than
households with no future death. On the other hand, past deaths of both males and
females were associated with increasing hours (just under 4 hours) for older adults (51-65

years) .

61
When months are divided into two periods, in a handful of cases deaths within 7-

12 months are associated with changes in hours across waves and in levels. Given the
inclusion of caring for the ill in the measure of house hours, we would have expected
closer future deaths (within 0-6 months) to have a larger impact than deaths farther from
the interview date. Furthermore, in several cases these coefficients are negative. For
example, women 20—50 in households with a male death 7-12 months in the future worked
8.5 less hours per week.

As in the case for farming hours, most of the set of individual and household
variables are not individually significant, although jointly, they are in some cases (see
Appendix 2, Tables 2a-c and 3a-c). Men 2050 who are the head of household work
significantly more hours in household activities, about 6 hours per week. The village-by-
wave interaction variables and the month and year dummy variables are significant for

both the OLS estimates and the individual fixed effects estimates.

Farm and House Hours by Per Capita Expenditure

A household’s ability to cope with an adult death may be dependent on the
household’s level of wealth. Wealthier households will have more assets and, perhaps,
greater access to credit, both of which will ' affect the extent of time allocation
adjustments. For instance, some households will be better able to finance medical and
funeral costs through sale of assets. Likewise, such households will be better able to hire
replacement labor for an ill or deceased household member. To address the concern that

the poorer households are most vulnerable to an adult death, the sample of households

62
was separated into two groups based on median expenditure per capita. Households are

categorized on the basis of their wave 1 expenditure per capita where number of residents
is adjusted for length of residence and adult equivalency.

Results for farm and house hours for these separate samples are presented in
Appendices 5 and 6. Generally, individual farm and hours were not significantly

associated with adult deaths for persons from either set of households.

Nonfarm Self-employment and Wage Employment

Changes in wage employment and nonfarm self-employment are important
because these activities are the main sources of cash income to rural households. As
mentioned above, notwithstanding the possible loss of income after an adult death, adult
deaths are associated with large, immediate expenditures for both medical expenses and
funeral activities. Due to low levels of participation among most samples, only adult
males 20-50 were examined. Participation in wage employment is nearly three-times that
of nonfarm self-employment (30 percent to about 10 percent). Low levels of participation
in self-employment required aggregating future and past death variables into one indicator
of any male/ female death in the past/future. These results are presented in Table 10.

Men in households with a female or male death were significantly less likely to
report participation in wage employment within 6 months of a future death. For past
deaths, there was no difference in participation for men in households with or without a
death. The magnitude of decreased participation was large for both male and female

deaths, at least 50 percent of the mean level of participation. In fact, the magnitude of

63
deceased participation probability was larger for male than female deaths (20 to 15 .6

percentage points). Nonfarm self-employment participation showed no significant
changes associated with male or female deaths.

Activities such as wage employment, that pull individuals out of the households,
may be more difficult to sustain in times when the household has a fatally sick resident.
Individuals in households with an ill female may incur additional household

responsibilities, including caring for the ill individual as well as children, even if hours in

farming or house activities do not increase in response to future deaths. Considering the _ .

emphasis on the importance of diversification of income sources in rural settings where
the main activity, farming, has a high level of risk , we tentatively find that one crude
measure of diversification (participation in wage employment) is not compromised by

prime-age adult deaths.

VII. Conclusion

Despite much concern about the impact of the AIDS epidemic on households in
developing countries, little empirical research has addressed this topic at the household
level. This study examines one area of the potential impact of a prime-age adult death:
adjustments in time allocation and household activities. Using a panel data set from rural
Tanzania, participation in activities across different demographic groups and at the
household level was examined. Understanding the impact of prime-age mortality on these
outcomes serves multiple purposes. This section links the results outlined in the previous

section with goals established in the introduction.

64
Analysis of farm and house hours across six demographic groups generally found

small and insignificant changes in labor supply of individuals in households experiencing
any prime-age adult death. The lack of any increase in hours for past deaths is notable
for children in particular for whom deaths are presumed to result in higher farm hours
assuming there will be an acute shortage of farm labor (Barnett and Blaikie, 1992). This
ties back to one of the points raised earlier that some of the areas in the region around
Lake Victoria which are most affected by the HIV/AIDS epidemic are those least
vulnerable to labor shortages. While we do not observe increases in farm hours, this
cannot be used to infer changes in school participation, a very important activity for
children omitted from this analysis. School participation could decline due to inability to
pay for school fees, uniforms and other costs.

Both female and male deaths are associated with an increased probability of new
prime-age adults entering the household. At least in the short-run (the 2-year length of
this panel), new members may help households smooth consumption by compensating for
lost labor of the deceased. Although given that the mean probability of a new adult
joining is around .30, new joiners are not observed in most households that experience a
prime-age death. For those that do, this partially explains the lack of increases in hired
labor or hours in farming of surviving household members after a death. It also suggests
that households most vulnerable to prime-age deaths may be those without access to new
household members. At the same time, it challenges policymakers to consider how to
infer relatively vulnerability to deaths in terms of access to labor that may not be easily

observable to outsiders.

65
To infer the impact of deaths on farming activities, further research examined

another farm input (hired farm labor), crops grown, and land cultivated. These results
help identify outcomes of increasing adult mortality rates for household level food
security as well as outcomes for the agricultural sector. While deaths were not associated
with changes in hiring of labor, coffee and banana farming in the last 7 days by any
household member did decrease in the 6 months following a male death in the household.
This is suggestive that some farm activities are temporary scaled back as a result of male
deaths. This analysis did not find that households experiencing a death shift cultivation
towards subsistence food farming. In fact, participation of individuals in farming of other
subsistence food crops (maize, cassava and beans) in households with a male death 7-12
months past declined. Without measures of aggregate output, we cannot draw
conclusions about per capita changes in output. Nevertheless, the reduction in
participation of individuals in these subsistence food crops but not tree crops among
households with a male death draws attention to the point that prime-age adult deaths are
a loss of labor but also a reduction in the number of individuals consuming food and other
goods.

Results on changes in acres cultivated showed declines in acreage associated with
more distant male deaths. Intensity of farming (area cultivated per adult resident) was
unaffected by adult death indicating that yields will be maintained. To further explore the
affect of deaths on agricultural outcomes, more detailed data on farming practices would
be need. For example, one extension of this analysis would be to examine outputs across

different crops to see if households produce proportionately more cash crops to food

66
crops. Ideally, we would carefully follow households across cropping seasons, rather

than use aggregated indicators of activities in the last 6 months or past 7 days.

Households experiencing a death do not appear to have reduced their
diversification over income sources. While we see significant decreases in wage
employment of adult men in response to a future female or male adult death, past deaths
are not associated with changes in either wage employment or nonfarm self-employment.

Likewise, as mentioned, coffee farming is not lower in households with a death more
than 6 months past. This may have important implications for households’ ability to cope
with the often large costs associated with deaths if a household’s cash income stream is
interrupted in the months preceding a death. Further examination into net changes in
assets or even net sales of livestock may provide evidence on how households finance the
high costs of funerals, not to mention medical care, associated with prime-age deaths.

These results can be applied to macro estimations of the impact of AIDS in a few
dimensions. They suggest that an adult death of males participating in the rural labor
market may not be offset by male survivors in the household, at least as measured by
participation. Cuddington’s (1993) comments in his analysis of the macroeconomic
impact of AIDS in rural Tanzania that estimates of the associated labor supply loss for an
AIDS-stricken worker and survivors before death and the reduction in the labor pool afier
the death are not available. He assumes (as do other macro studies cited) that one AIDS-
related death reduced the labor pool by one. This is not true for household labor
available to households for farming, as deaths are associated with a higher probability of

new adults joining. Application of this result to regional measures of labor participation

67
requires understanding where new members came from and their previous labor supply

including their sector of employment. Areas with increasing mortality rates may draw
down labor in other areas. These results also shed light on the affect of a future AIDS
death on labor supply of non-afflicted adults. Lack of ex ante adjustments in time
allocation among survivors narrows the range of reliable estimates of the fraction of labor
productivity (hours) lost per AIDS case to equal to or under one.

This analysis has focused on time allocation and activities of agricultural
households using a large and detailed panel data set in comparison to the few studies in
this area which have, to date, been small and illustrative. To complete the picture of the
impact of HIV/AIDS, though, these results ultimately nwd to be incorporated into results
on an array of adjustments and outcomes that households in this region may undertake
when confronted with an adult death. These would include issues such as liquidation of
assets, nutritional status of survivors and schooling of children. Further, these results
need comparison to other types of farming systems, such as areas with lower population

density and without tree crops.

68

 

 

         
 

  

I Dodoma .5.
Dar u ‘i .

TANZANIA em... -=-

 

ZAMBIA

L '5: ‘0.-

 

Source: Mukoyogo and Williams (1991)

Figure 1: Tanzania and Kagera Region

memfimfimmg

RmmnEmMmmm

Pammemmg

 

 

 

Figure 2.A: Farming Last 7 Days

 

Mflu
‘ ‘ Hmflm

    

. :‘.:“!4
7172737475767"
Ag
Figure 2.8: Wage-employment
Last 7 Days

 

«L-

 

Maw
““"Rme

 

Figure 2.C: Nonfarm Self-employment last 7
Days

70

40:3 05 Me 8.58 on. 2.56 0:. 9.? =5 .85 $8. 3 watch?
€523“! Pa 802.5282 .839 MERE. 6.5. RES .9 one SEES Eat vows—88 mcoumtomne he coir—3.2. :8 a he _ 530%? com “€82

 

 

 

 

new mom: 2 com NV ONE mHoN no.3 333‘
5mm vow~ am So E E8 memo» OméN mama—2
can 83 3 can mo 85m mum?» 8-0m magnum
amm 3.2 m vov m 002 Emu» 3-2 mow“:
mom vmg N 05 N $2 Mao» 3-9 meadow
S: Nam N. 53 Na ca; «53% E4. GBEEU
6v 6 as 6 S 8 4%
Eoztaéoz 30335 $953.82 £0323
£3365 3 $3365 550388 “so—com

 

80:33on N $2
a 55 32830 $8335

35 2: 5:6 weer—030.5%“

 

 

agree as” $9 a a:

71

.8935? as S 58: £8332

use 2. .5 flaws» 2583 25 5 _.9< use. 9 5589.? 8» £32“ 95 a 58.. 835:9 9.982%» 25 95 use a £558 23 £99 .3352: 2288.. a 5.. 555
Ba .83: mamas—5 5.88 mar—«Pan .9853 32809 .15389 95 598:8 .8 £88 E3: 083.. 3855955 3 Ba 9833—. fig 55 5828 Eu 5 85 882 .5082

 

 

 

 

82 32 83 32 3.9 «an a.” 2558
98.9 985 98.9 «85 989 855 98.9 985 98.9 855 955.9 985 388 979 23E
8— _ .9 ~85 98.9 585 989.9 585 :2 .9 255 E _ .9 :55 c 5 _ .9 255 2.85 55 8:5...
82.9 985 9859 «85 92 .9 :55 as .9 255 62.9 :55 92 .9 985 €88 2- 9 .39
a: .9 985 22.9 985 92.9 585 62.9 855 92.9 585 92.9 585 2.85 95 .85

, 53% use»: 333:3 35:.
955.9 585 98.9 585 98.9 985 99.9 985 :2 .9 585 98.9 855 3.85 9-9 23:".
92 .9 :55 92 .9 £55 6 _ .9 955 92 .9 «85 92 .9 985 62 .9 985 3.85 55 23:9
93 .9 985 99 .9 £55 99 .9 255 99.9 585 82 .9 585 99.9 585 £85 2- 9 up.
92 .9 855 92 .9 .955 82 .9 585 an: .9 985 an: .9 985 c: .9 585 385 Po .25
.538 k25.9.2 33533 39.59
$55.53 $33
98.9 925 959.9 9:5 93.9 55 959.9 5955 959.9 255 2:9 :55 a a.” .89 a s
5.35 9 an. =528_o_§5
Evie—33.8 895.252
959.9 585 99.9 89.5 959.9 855 659.9 9555 92 .9 585 855.9 585 8 a.» .89 a c as 9 22
8:596:98 5:09:92qu own?
955.9: 559.: 6.5.9 :55 69.8 9859 92.9 § .2 955.5 9 58.: 55.9 98.9 as 9 .3 92. 25585:
699.8 552: 995.8 52 .2 995.9: 52.2 93.9: 599.: 993 c 52.: 955.9 5955 an. 9 .3 38.. 55
flag: as» 539 2855959 as» 2-2 as» 2-2 289. 39 33> $3.39
933. 8:2 8E 8:2 8E 8.555

 

 

 

 

 

 

 

monmusm nan—«m ”N 2an

72

.30:

8:82 .8 :5 83.05 388.? 35323 .8 «.2 2.5 .N 5208“ 8m .5 a t... as ”an a : 5:: a 35:3... .825“
8.8m? . ._o>o_ 3:232 05 a $588? .8» 389:8 :8; 2:5 £03838 30 on. 3“— .Ea mace—.28 5 2a macho Bug—Sm £82

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

32 3.2 33 3.2 83 GE comm ~3— ofim Baum
890 306 «8.9 85.0 movd Rod 2:6 Sod 288 d 83 58v 234
mwmd 586 End 5de 02 .o Sod 85d wvmd 833 a. 63 58v via—om
..uu§uREm.a 3.6535 397$
02 .o 536 3N6 :«d :2 .o _N~.o 3nd find W—
98.6 38.6 85$ a8. 0 G59 9 £3 .8 $8.3 @516
3:56 n86 vmwd can; «end. 3:86- awhm Nwwd 2:58 NTo 9—38 58c 032
33.8 23. 3 €36 3.9.5 2506 Goad Ame—NV 68.:
«8.7 59.7 2m; 8N.— wE .N- 356. mood- 1.807 2:58 ad 63 Snow 2“:
33. C S n5. C 355.8 28.8 93.3 83.9 A: 9% Gde
a5~d 08. T 58$ and mad. quad REA- mafia v.55:— NTo 0:58 58v “.3th
Amen. C 8N0. C 83$ 58. 3 Ga" .8 33$ 33. 3 Roe. S
ova.“ 8nd. 524 3.04 mwvd wand #2 .0 2n; 3:88 N70 83 53v viz—om
.md 30 .mm 30 .mm 30 .mm 30 933$» bagnxm
23: .233 V .233 9:8: V 2:3 2:3 V .233 .223 V
8.5: 8.8m 33:
8.8 :02 oméa 5:83

 

 

 

38: 8 63: mean .380: 2288: :22 a. .395 am 038.

73

.3 26:5 .8 88: new .6862

 

 

 

 

 

 

 

 

 

 

 

:62 :3 3m: v3 :02 $56 vnfi Eu 0:5 268mm
~36 ‘26 656.6 nnm6 656.6 36.6 nmn6 666.6 28.6 a. 3.8 586 2:2
$56 65.6 «66.6 35.6 69.6 m5m6 «3 .6 «S .6 282 d .36 586 2:th

.6365362 365265 39¢
3N6 6N~6 56~6 63.6 63.6 2m6 :26 63.6 NM

Amfi .mv Sow-.8 $66.9 A52 .9 A6366 83.9 ~65N.$ $56 .3
v2.— 65m.m Gad «63.6 626 .965- 65.6- 62.— 26.88 ~76 88.6 536 2:2

A6625 $3.8 965.8 fined 86m. 8 @6666 Gun. 9 A666. 3
6.6.7 686 mafia- $566. «on; $50: 6666- 66_ .6 2:88 N76 .86 586 3:2

26:66 9866 365.6 €66.66 866.0 $66.3 $66.9 360$
666.7 6666- n6m.m- 366 25.n- 636- $656. 65.6 3:88 N76 95:6 586 2.25.6

$n 6 .3 36:. C A5966 866. C @5666 $5.8 $66.6 $65. 3
656.7 526 63.6 656.7 6666- $66 36.6- 3686 26:2: ~76 :86 536 225%
.m.m 30 .mm 640 .mm 30 .mm 640 .3332» bugmuw

2:3 a. .33- V 2:3 2:3: V. 2:3 2:3: V. 2:3 2|I:3 V

83: g 8.5: III
66-2 52 2.2 5:53

 

 

 

23: .8 836 “586 .3502 29.88: :22 6o 33:: 5m 23.

74

.66 as“... 8.6 88: 86 .282

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

666— 666— 662 6666 686 6666 656 6666 0&6 26:86
666.6 62.6 666.6 66.6.6 666 666.6 86.6 666.6 883 a. 626 586 22
666.6 666.6 666.6 $66.6 36.6 666.6 666.6 666.6 883 a. .26 586 £25m

”832626.; 2.8.6 26.6 397%

66.6 66_ .6 666.6 666.6 66— .6 E H .6 66— .6 6: .6 N6—
6666 .66 $66.3 A6866 94.666 866.: 266.3 $8.3 866.5

.6— .6. 666.6- 6664 36.6 6664 «66.— 66v._ «6666 26:2: 676 9:53 .686 262
366.66 626.3 866.66 @6666 866.: A667: 666.: 866.66

666.6 166.6 62.7 666.6 63.6. 666.6. 3666.6- «3.636- 65:68 676 ~26 586 032
$66.66 €66.66 666.66 866.3 866.3 866.: A6663 866.:

66_ .6- 6666- 666.6- 666.6 666.6 666.6 666. 7 666.6 A9.5.5.: 676 23:..— 586 0.680%
@6666 $36. 3 666.66 86666 A666. 3 $66.66 88. 3 666.66

666.6- *6666 666.6 :66. 56.6 «66.6.7 666.6. 666.6 65:9: 676 .26 £36 225m

.md 30 .mm 30 .mm 6.5 .m.m 640 .3332» 2632.6
2:3 2:3 V 2:3 2:3 V ”4:46 2:3 V 2:3 2:3 V

8.6: 82m 83: g
66.6 933: $76 62220

 

 

 

25: 8 86-66 W586 352 233.8: :22 3 .895 an 2.5

75

.22. 8:82 .8 .3 3.205 £8832 3.2.63 .8 n? 23. .N 5232 8m .5 a :5 Ba in a z. 32: a Banana»: 3E3...
menofis .. ._o>o_ 3:233 05 a matamao Bu 380:8 anon 9:2 .8§E~8 30 05 Bu 65 885:83 5 8a 388 Egg—Sm ”882

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

v2: 32 33 32 cana 32 coma 32 86 3.38m
Sod and vnmd nnwd amvd mid Sod wood 888 nu .ma 58¢ 032
waod wood and nmnd awmd v36 «86 Sod 238 58¢ 2mg
33. 3H6 coed 3nd mood mood wand 806 33 58w 0.x:
and 8nd 8nd «.36 3:. n2 .o and «86 283 a Ea 58¢ 255m
acad Sad End mnvd 036 8nd and nmnd 238 58a 286m
:46 3nd wand anod Sad 3 ad wand nnvd 3.8 53—. can—om
69:85:»; 38.3% 337%
2:6 2:6 a§.o agad mEd wafio caad a_a.o Na
Gmnd Aaneav anon $3.8 83.3 88.3 83$ 98$
3.52.8 mama nna.a find .vand- :_o_.a7 23... Leon. 3.5:. aTn 9.88 58.. 032
3.2 . my Amnnav 33$ Anmmév $3.0 $3.9 $3. 9 An— v. 0
nam.m ans. 7 3; 6. 2b.? 03.7 awn; iww We. 2.03.2 2:88 ed 883 58c 0.34
annav 2%.: 83.9 Sand $8.3 23.0 89an Sand
not? n3. T ”mad :na nae.a- onto. «2:. maod 228:. aTn 3.8 58c 032
anaav Anatav Goad Sand $36 33.3 Gavav 336
98.7 2a.“- cmad 986 3:- amod. mmna- 3.636. 3:88 oé .33 53v 232
Good 33$ Good 63.0 Aannd 38.8 Angd fined
8nd Sad $56- Sod. anmé. nmoé. 08.7 3nd aficoE aTn 838 58¢ 0.8—om
Amfl .av 38.: 68.9 23.9 Aaooav $3.9 Amvoav Sand
afld wEa- mad. a2: and c3 .n ooh? awfio 25:2: 06 283 58a own—om
83$ Gaga Gamay 9430 Sand Rand 63.8 Anomav
an»; men? 934 wnng *awwé mama 3a; 2o; v.55... aTn .89 53v 295w
Anon. 3 Aamaav Gon .av Anaeav Aa_ mav 83.8 62 .av Annoav
nooa Rad coed woo; 29a 8n.m onwd. 8v.— m5=oE eé 88 58v 225m
.mm So .2 So .mE 30 :2 Bo 33$» $333
.233 .58: V .933 Q33 V .933 .233 V .233 233 V
83: 3 345m .Ehlam
onéa 52 ow-oa 5:53

 

33: 8 83: 38a .3502 2888: .33.. .o .895 ”2. 29a

76

.3 3%... 5 8.8 3 £32

 

 

 

 

 

 

 

 

 

 

 

 

 

 

SN" 5 3mg 3.» XS Eb §— «5 can 0385
3nd 2 d d wfid 94d tad 8nd and onnd 238 nu “ma 536 032
dnod Knd aEd vnvd 35d and dwad nond 238 53d 032
vfid Nbdd med Sad onwd oFd dnmd dnnd and 586 0?:
nnwd S: d nhnd Rod End owed and 3nd 288 a. 88 53v 235"—
vnnd wndd «Ed wand odvd nwdd Edd dbmd 9.83 58v 255m
Nnnd ~nnd Sad 8nd _Sd onod cued nndd “man 636 £38m
..uu§§m.a 3.558% uSuai
aoNd dead 8nd «mud find 2nd nnnd dand hm!
83$ 68$ 38.3 88$ Etc 85.0 dad 598
8: 1.3.6 nnnd. 2nd nndd card. mtg- with- 3.88 NE. 9:83 53—. 0R2
33.9 35.8 88$ 6:9 Agni A2 5.3 93.3 Awnnnv
find. v3 .N- $8.5. nnvfi Ed.“ and. Sad vEn mecca dd 2:3 58d 032
68.6 $2 .9 $99 $2.6 $3.9 $9.3 $8.3 son:
3x: Loan achd. dhnn Sod ~24 n26. 8nd. 3:88 N75. 3.8 58c 032
92.6 83.8 88.8 63.9 3‘an 68.9 Read $3.3
an- mood- nnwd- can? onod one; wood- 62d mace:— od 33 58c 23)—
833 83.3 Gannv 83.3 95.3 82 .D 88. nv anwi
nNn.N- 38nd? need- 30.5. «2.7 nnwd. $.n.n- 086 362: NE. 883 58d via—um
8:.8 $3.8 $3.8 Agni 934.3 93.0 30th 88.3
2nd 2.”; nfid ”flan End. :Ngd- 3:56- «man- «5:08 ed 238 586 £28m
C _ _ .nv Anemd A598 $3.9 Wand 69$ annd Gun. 3
9: d SN; «and 53d onhd. SYN k: d van «.588 24. and 58d own—om
Gwmd Gad finnd 33.6 83.6 8n _ .3 Annvd 83.6
2:- nvad. mud." dnad- *hnné. nth- mood- v.86 2:88 ed “mag 58—. can—om
.mm 30 .md 3O dd 30 .mm 30 3§> bangs
.233 .933 V .233 “.33 V .3 .233 V 2:3 .233 V
3.5: gm 88: :5"—
2-2 :02 2-2 5:83

 

 

 

38: 8 63.: 38a .38»: Banana: =22 .5 3:: 5e 2.5

77

.3 0.2.... 8.. 8...: 8m .6892

 

 

 

 

 

 

 

 

 

 

 

 

 

8m. 8... 8m. 58. 9.... SR 9% .EN .3 295....
89.. $2... .29.. 5.... on... one... .8... m8... 83.... a. .2... 58.. 2.2
3.... E... 9...... $8... $9.... 89.. a8... m8... 2&8 58.. 2.2

89.. $2 ... 2c... 89.. 89.. $8... to... .8... .2... 58.. 2.2
$9.. 9.9.. 3...... 89.. $9.. $9.. 5...... an. .o 2.5... a. .2... 58.. 22...".

2...... E... 89.. $2... «.90 89.. n8... 2...... ea... 58.. 22...".

S... :N... 89.. 2...... o3... 39.. 89.. RN... .2... 58.. 2.5...

”8.39%..sz Emcxuux 39¢

SN... 9:... SN... .39.. 3.... 3.... 8.... 8.... 9.
$9.... $3.9 $.99 .899 .399 .899 $2.9 .399

at.“ :22. 89$- 08.? 89. an... m... .N- .89..- .55.... N.-. 8...... 58.. 28,.
23.9 $.99 $9.9 $8.9 $9.. 9 .399 .899 $3.9

29$- 2.....- vuw..- R... 2:... an... 58$ :39... .585 c... 2...... 58.. 222
28.9 .599 68.9 .899 $8. 9 $99 9 .899 .8... 9

$9. :9. 29.- 8.... can? 8. .o 8.99 «9.... .55.... N.-. .2... 58.. 2...).
$8.9 $2.9 $8.9 .399 .899 $8.9 28.9 $8.9

a... .o E... .8...- 399 mm. .9 $99 :2. .m- 35.9.99 .585 o... .2... 58.. 2...).
$... .9 .299 .89... $.99 .89 9 $.99 $8.9 $8.9

2.99 9.9.. :89..- $9..- $9.. :9... .383. 89$- ...58... 9-. 2.5.. 58.. 22...".
$.99 $8.9 $8.9 $9.... .89 9 $3.9 $9.9 ..$9 9

9......- 8...o. a9...- $N$ 89.. 8.... 89? :99. .58... o... 85... 58.. 22...".
2.99 $.99 $8.9 $.99 .999 .899 $.99 $2.9

29.. 3.9.. 9%... 8...? «$9? $9 .- 3......- S... .58... N.-. .2... 58.. 228...
89.9 .899 .899 $3.9 $2.9 $5. .9 $.99 28.9
.03..- ..R. 8...? 3...... 99.. 3.....- :9.- ..3..- .585 e... .2... 58.. 228..

.2... .30 .2... 9.0 .m... 9.0 .mw. 30 8.8.5: $3529....
953% .932 V .239: .939.« V 9.33 933 V 932‘ .932 V

3.53 Emma 3303 5.8%
8-. m 25...... E-.. 85......

 

 

 

2...... .8 $39 .582 .35.: 29.88: ......< .o .08.... a... 2.5

78

.82. 82...... .8 .3 8.5.8. 988.98 .8892... .5 3 2.3 .~ 225...... 8... .2. 2. 2: 5.... in a :
:99. a 89899.9.“ .8998... 8.8.9... .. .. 3 9 8...... 833.5. 98.9.... .8 099...... 828.9 a 8986.9». 59.... £88.99.
.29.. .1 9.. =8... 29 9. 92.3.90 9.9.93 9.93389 .9 09.5... 9. 9052.8 98.. 83 32 85.09908 .8028... .9
:96. 228...... 29 9. 9.9.88... .8 938...... 88583.. .9 Echo 9.395... .995 .329 9.9.... 92.9.9.8 2.. 39.89 .5262

 

 

 

 

 

 

999. 9.9. afim 3.9 29.5%
- - 999.9 99.9 89.9 93.9 0.3.... a. .39 .98... 09.2
- - 93.9 29.9 59.9 999.9 88.9 .9 .39 .989 09.80"—
..uucuuR39... 3.8.x Bx wSuz-m
999.9 39.9 $8.9 $9.9 999.9 999.9 Na 99989
- - 399.9. "999.9. 3999.. 59.9.9
- - 6.9.99 9.9.99 8.9.99 A. 8.99
- - 95.9 «999 29.9. 999.9. 99.8... 2.9 28.... .98.. 032
2.8.9. - En . .9. - .99. .9. -
£9999 - Sn. .8 - £2 .99 -
.89.. - 2:89.. - .33.... - .55.... N... .2... 58.. 2...).
- ”29.9. - 7.. . .9. - 82 .9.
- €3.99 - ASN.8 - $9999
- 99. .9 - 2&3... - 2.3.9.9 99:2: N:- .m.... .98... 09.2
- - 8999.. 8.9.9.. 3.9.9.. 38.9..
- - .3999 .9399 85.99 @9999
- - .999. 999.9. 98.9- 8. .9. .995... ~79 283 .98... 0.8909
EN. .9. - S 8.8 - $9. .9. -
2.2.9. - $9. .9. - .93.... -
2:89.. - N2 ... - 2:89.. - .58... N... .2... 58.. 22......
- 32.9. - 28.9.. - 399.9.
- 33.99 - 63.99 - .99. .99
- 18999.9 - c9. .9. - mam-9 99:2: 2... .89 .989 28.5”.
o n .. m N . .235. 9.3.3.3.
.552 £22 .ofiom £8.89 8qu £2.98 .9 «9.2

 

 

.582 o .3 .2 22.88.. 22...... 2...... s.

... Essa... 8 $39 2589 .35.... 22.88.. 2...... ... .88.... a 2......

79

Table 6: Farm Hours of Adult Joiners and Other Resident Adults,
Sample of Households with Any Past Death

 

 

 

Women Men
15-50 15-50
Adults who joined household in last 6 months 9.66 9.46
(10.45) (7.78)
[34] [24]
Other resident adults 9.61 9.76
(9.38) (12.96)
[67] 460]
Notes: Standard deviations in parentheses; sample size in brackets.

 

80

.22. 859: .8 .3 83.05 gage
38:63 new flat. 0.3—. .N 538.? 8m 4: a I... v5 ”Rm 3 .3 ”$2 a 8:853.“ 323% 8.8%.: .. 4 9 o 82”.
813:; >883 .8 owns? 888:. a 35% 523 flagp— hofio .3 .8 :88 05 3 Bums—go SEE bananeq 5
omega 9 3528 :03 83 9:2 nausea—coo .8355“. :33 .8 .aoxufln E :26— Eonomsoz 05 a 3:883 .8 3.8:8
=8: 26: 53.5.8 €95 05 he 65 88553; E an echo Emu—5m .NM 238m 05 fi 8855.8 :33 .8 mm .382

 

 

 

 

 

 

 

8mm 93 88 £5 sass..."
83 E .c 83 £3 53 89o .23 53.. 22. a 255m
“anaemia? goiox 32¢
Sod god «2 .c z: .o 33 5 .o N..—
- - - - 38.9 -
€2.83 - $3.5 - 62.8 -
8&3 - 2.2:- - n85 - 3.88 «3 ea 58“. 2a:
- - - - :85.
- 8:83 - $2.5 $3.9
- 2.3mm - $2: - :85 was... 3 .93 as“. an:
- - - 68.3 -
63.25 - Eves - 82 .9 -
5.63.987 - 3.2- - mad. - 3.55 NZ. ea 58.. 22nd
- - - - :83
- 33.88 - 88.5 - S: .e
- $3.88- - 22:- - 82* £85 a: an. 5% ”need
0 . n v m N _ “figs 3353
98.5 BE Bogom 38% 85 22.3.8: .305
eafias as... 3658 do .382 .32 3.2 .22

 

£82 a .3 a .23 8:: 8 82: mean .352 22.035 .33 .3 .895 g 23

81

Table 8: Impact of Adult Household Member Deaths (15-50) on Land Cultivated
in Last 6 Months, Household Fixed Effects

 

 

 

 

 

 

 

Area of land cultivated Area of land cultivated per
adult resident
Explanatory Variables 1 2 3 4
Female death past 7-12 months 0.296 - 0.191 -
(0.381) - (0.177) -
Female death past 012 months - -0. 156 - 0.167
- (0.322) - (0. 150)
Male death past 7-12 months 0850* - -0.282 -
(0.452) - (0.210) -
Male death past 0—12 months - -0.496 - -0.249
- (0.375) - (0. 175)
R2 0.332 0.331 0.227
P-value for joint significance:
Female & male death past 0.132 0.360 0.238 0.210
Sample Size 2323 2643
Notes: Standard errors are in parentheses. ‘ indicates statistical significance at 10%; ” at

5%; and *" at 1%. See Appendix 2, Table A2.6 for additional regressors included but not

reported here.

82

Table 9: Impact of Adult Household Member Deaths (15-50) on Crops Famed
by Any Household Member in Last 7 Days. Probit

 

 

 

 

Coffee Banana Maize, Cassava
or Beans
Explanatory Variables 1 2 3
Female death past 0-6 months 0.127 0.121 0.253
(0.157) (0.171) (0.165)
[0.050] [0.040] [0.066]
Female death past 7-12 months 0.035 -0.017 0.253
(0.216) (0.216) (0.201)
[0.014] {-0.006] [0.080]
Female death future 0-6 months 0.336 0.285 0.251
(0.228) (0.242) (0.237)
[0. 133] [0.088] [0.079]
Female death future 7-12 months 0.285 0.007 -0.029
(0.291) (0.303) (0.312)
[0.113] [0.002] [0010]
Male death past 0-6 months —0.456*"‘ -0.385** 0.161
(0.200) (0.204) (0. 178)
{-0.171] {-0.142} {-0057}
Male death past 7-12 months 0.217 -0.099 -0.648***
(0.278) (0.257) (0.253)
[0.084] {-0.035} {-0.246}
Male death future 06 months -0.009 -0.353 0.289
(0.253) (0.261) (0.241)
{-0.004] {-0.130} [0.090]
Male death future 7-12 months -0.755""" 0.153 -0.346
(0.365) (0.388) (0.366)
{-0.263} [0.050] {-0. 127]
Pseudo R2 0.230 0.254 0.205
P-value for joint significance:
Female death past 0.722 0.775 0.192
Female death future 0.216 0.500 0.565
Female death past & future 0.502 0.738 0.291
Male death past 0.055 0.140 0.022
Male death future 0.110 0.358 0.254
Male death past & future 0.070 0.242 0.033
Sample Size 2643 2889 2893

 

 

Notes: Standard errors are in parentheses and have been corrected for clustering at the
household level; in brackets coefficients have also been converted to change in probability
(dp/dx) evaluated at the mean of all other regressors, which represents a discrete change for
dummy variables from 0 to l. * indicates statistical significance at 10%; ” at 5%; and
"" at 1%. See Appendix 2, Table A27 for additional regressors included but not reported
here.

83

Table 10: Impact of Adult Household Member Deaths (15-50) on
Nonfarm Self-employment and Wage-employment of Men (20-50)

 

 

 

 

 

 

Nonfarm
Wage-employment Self-employment
ExplanarolVariables 1 2 3
Female death past 0-6 months -0.020 - -
(0.221) - —
{—0.007] - -
Female death past 7-12 months 0.127 - -
(0.282) - -
[0.046] - -
Female death past 0-12 months - -0.288 0.143
- (0.216) (0.198)
- [0.012] [0.038]
Female death future 0-6 months —0.527" - -
(0.252) - -
{-0.156] - -
Female death future 7-12 months 0.193 - -
(0.329) - -
[0.070] - -
Female death future 0-12 months - 0.034 0.105
- (0.204) (0.270)
- {-0.092} [0.028]
Male death past 0-6 months 0011 - -
(0.262) - -
{-0.004] - -
Male death past 7-12 months 0.169 - -
(0.311) - -
[0.061] - -
Male death past 0-12 months - 0.067 0.220
- (0.237) (0.257)
- [0.024] [0.061]
Male death future 0-6 months -0.754* - -
(0.436) - -
{-0.201] - -
Male death future 7-12 months -0.490 - -
(0.541) - -
{-0.145] - -
Male death future 0-12 months - -0.647 -0.697
- (0.425) (0.503)
- {-0.181] {-0.121}
Pseudo R2 0.173 0.177 0.140
P-value for joint significance:
Female death past 0.867 - -
Female death future 0.070 - -
Female death past & future 0.218 0.359 0.771
Male death past 0.822 - -
Male death future 0.225 - -
Male death past & future 0.501 0.278 0.169
Sample Size 1793 1622

 

Notes: Probabilities are estimated using probit with standard errors in parentheses corrected
for clustering at the household level; in brackets coefficients have also been converted to
change in probability (dp/dx) evaluated at the mean of all other rcgressors, which represents a
discrete change for dummy variables from 0 to 1. "‘ indicates statistical significance at 10%;
” at 5%; and “"' at 1%. See Appendix 2, Table A28 for additional regressors included but
not reported here.

APPENDICES

APPENDIX 1

APPENDIX 1

Sampling Strategy and Attrition

The KHDS surveyed a stratified choice based sample of households in Kagera,
where households thought to have a high risk of illness were over-sampled. A random
survey even in a high HIV sew-prevalence region such as Kagera will yield few
observed prime-age deaths over a two-year panel. While the region has a high HIV
infection rate, there is much variation across the six districts, with Muleba, Bukoba-
Urban and Bukoba—Rural having the highest rates of infection (Killewo et al., 1990).

A total of 838 households completed the first wave of interviews; of these, 757
complete all four interviews over a period of more than 2 years (see Table A1.l). The
selection of the sample of households for the household survey was accomplished in three
steps.

First, all 549 primary sampling units (PSUs) or clusters in Kagera region (most of
which are villages) were categorized into eight strata based on four ecological zones and
high/low adult mortality rates (AMR) calculated from the 1988 Census. A stratified
random sample of 51 units were selected.1 High mortality sampling units were over—
sampled in each zone. Among the high AMR PSUs, 26 of 61 were selected compared to
25 of the 488 low AMR communities.

Next, all 26,416 households in these 51 communities were surveyed, resulting in

enumeration of about 9% (135,683 individuals) of the total population of Kagera in about

\

l 54 PSUs were originally chosen but 3 were dropped due to insufficient resources.
84

85

9 % of all communities. The enumeration survey was very simple, containing only four
questions: 1. Name of head of household; 2. Number of people in the household 15 years
and older, and 14 or younger; 3. The age and number of weeks any adults (15-50 years)
in the household who where too ill to work (listing up to three adults); and, 4. The cause
of death (illness, accident, childbirth or other) is recorded for up to three adult (15-50
years) deaths occurring in the household during in the preceding 12 months.

Finally, based on the enumeration survey, households were labeled as “sick” or
“well”. “Sick” households were ones in which there was an adult too sick to work or an
adult death in the past 12 months. Only 7.2 percent of households (1913 households)
were identified as “sick”. The sick (well) households were randomly selected within
village from the set of all sick (well) households. Sick households were over-sampled in
each cluster comprising 14 out of every 16 households targeted. A total of 816
households were selected to be interviewed. By the first wave, 7-12 months after
enumeration, 47 of the original 816 households were not interviewed for a variety of
reasons (see Table A1 .2). However, field researchers replaced these households with 69
more households from the enumeration sample. Thus, 838 households were interviewed
in wave 1.

The second wave of interviews began in the Spring of 1992. Among other
smaller differences between the wave 1 questionnaire and the questionnaire in waves 24,
the reference period for recall questions is usually 12 months in the former and the period

between interviews for the latter surveys. The average number of months between each

survey is 6-7.

86

Characteristics of sample attrition across interviews are documented in Table
A1 .2. Attrition is an important issue because households experiencing a death may be
more likely to drop out of the panel, for example, due to dissolution of the household
unit following a death. Thus, the remaining sample of those households that experience
a death and do not drop out of the panel may be least vulnerable to the shock of a
prime-age death. One quarter of the households that drop during the course of the
project (19 out of 81) had moved due to factors related to a death, though the majority
of households that leave the panel do not leave for reasons related to a death. The

analysis in the text does not restrict the sample to households that complete the panel.

Data Editing

The KHDS data used for this study were received completely unedited after
their completion in the field. Each wave of the KHDS was fielded in two rounds about
two weeks apart. This design allowed for some data checks and follow-up in the
second round in the field on inconsistent data collected in the first round which covered
about half of the sections of the household survey. For this study, editing of the KHDS
data consisted of a multitude of verification and consistency checks, which were
facilitated by the panel nature of the data.2 Some observations (either on entire

households or individuals) were ultimately eliminated, and in other cases, corrections to

2 The unedited status of the KHDS extends to even the most basic data checks on variables such as sex,
membership and consistency in identification numbers (for individuals and for households). These basic
consistency checks were undertaken before preceding to other data checks. Overall, there were very few
corrections of these sorts.

87

the data were made based on other information in the panel. More details pertaining to
editing these data follow.

Originally 840 households were in the wave 1 sample. In the second and third
passes of the panel, field researchers added 75 new households as replacements for
households that dropped from the sample after wave 1. Since the criterion for the
selection of these households is not documented, they are not used in this analysis.

Table A1 .3 describes the number of individual observations dropped among the
13,528 cross-section of individuals between the ages of 7 and 65 residing in the 840
households. Two of the 840 households in wave 1 were excluded from the panel when
it was discovered that these two households had the same household head with a
different wife in each. According to supervisor manuals, they should have been
considered one household, and, therefore, household-level sections of the survey should
have been administered only once in each wave. Given the difficulty of combining
certain sections of household level data (such as land and expenditure data) both
households were discarded. All other observations that were excluded had missing
values in either the explanatory or dependent variables.

A small number of corrections were made to particular variables. For example,
for education, where no enrollment is reported in any wave (mainly adults) and highest
grade is not constant, the most frequently reported level was chosen. For cases with four
levels reported and no majority, the largest and smallest values are eliminated and the
average of the remaining observations is calculated. For cases with three different levels

reported, the middle value is used; with two levels reported, the average of the two levels

88

is calculated. The new data for education are also used for education of resident parents
and household heads. Likewise, these checks were made for the education reported for
non-resident parents. For age, where inconsistently reported (e. g. people getting younger
or aging by more than 1 year across back-to-back waves that are 6—7 months apart), all
reported ages were back-dated to the first wave by length of time between waves. An
average “first” age was then calculated from these data and ages for subsequent waves
were recalculated by adding months between waves.

Some variables, such as orphan status, were expected to change in one direction
only and were thoroughly checked for consistency. Corrections were made only when
sufficient information allowed for ascertaining a presumably more accurate response.
Otherwise, the data are considered missing. Data pertaining to land, income, or other
measures that can vary in any direction across waves (as opposed to discrete status

variables) were not imputed when missing.

89

Table A1.1: Sample of Households in the KHDS by Wave

 

 

Number of households
Wave Dates (# dropped between waves)

1 Sept 1991 - May 1992 838

(37)
2 May 1992 - Nov 1992 801

(17)
3 Nov 1992 - May 1993 784

(27)

4 June 1993 - Dec 1993 757
(total dropped = 81)

Notes: Dropped households include some households that completed portions of
the survey, but failed to complete the entire interview or any subsequent
interviews.

Table A1.2: Sample Attrition in the KHDS

 

 

 

 

 

Between enumeration Between waves of the
and wave 1 household survey
Overall rate of attrition (%) 5.8 9.7
Reason for attrition number per cent number percent
Moved (not related to death) 18 38.3 46 56.8
Moved (related to death) 7 14.9 19 23.5
Head away a 9 19.1 2 2.5
Refused 4 8.5 13 16.0
Illness 2 4.3 1 1.2
Not found 1 2.1 O 0.0
Reason unknown 6 12.8 0 0.0
Total 47 100.0 81 100.0

 

 

 

Notes: The 47 households that dropped from the sample between enumeration and wave 1 were
replaced by 69 households selected from a list of “replacement” households compiled from the
enumeration survey. a. The absence of the household head was not a valid reason for dropping a
household from the sample. This occurrence almost entirely ended after wave 1.

688m 5298 can .wEEonom who—ta ER «.1052: mots—E agar—Sofia”. Econ—a we 28 05. E08 0225 28 E mmEEo: ES— :82
8: Eu mEenomzo: 93,—. 458.3.qu :3 8.85: van «was E 2.5: can .33: qun 82.3.: 8? =5 .wEEeé E 252. bEmEEovoE 2 Enos
mEmEE .bowofio 28 Dee 2E wouma Ba EEEEeE 038 £5 E Ease—Ea dim—:88 3.25:8 5:388: 8: En 3&8“. mEun .8 283% .5882

 

 

 

 

#3 o we 9. 2m wmmfl ~88.

.. o 2 w 2 $3 Emu» no-3 SEE

-- H 3 c me E “N Ea?» omém 8E2

-- N o 9 3 mmwm memo» omuom magnum

mm o N 0 CV $2 23> 2-2 .832

R g N N. hm moi 38> 3-2 8380a

NS N 3 2 >3 momv 88» 3-5 5:230
05:88:20 95 Enos go: 253 E3 get 395% 295m

3223 En wEmEE wEmmzz mEmEE mEonomson 03% 230,—. RES
@2506 cemmom

 

 

 

«an menu see @885 823530 a: 2.3.

APPENDIX 2

N In. IQ I 7.xv-Nul— <

8805028 5 3235.. .588 .082

 

 

 

 

 

 

 

 

 

82 3.: 8m 3m. 88 «an 0.8 2908
38.8 .83. .808 83.9 88.8 33.9
38.. 8.... SN... 2N... wNN... NhNd 8 8.0 .30.. 0.080. .. .8 30.. 22.852. 0.080...
83.8 ASN.8 $2.8 85.8 $8.8 $8.8
wood mood. new... 580 Sn... SN... .30: .0 mE328 .0 800%
33.8. .982. .986: 8.82. 836.. 33.0.8
8%.? 03.9 8. .3 Sada Nwm..n 38$ 30.. 32.83.. .e 0m<
.03 .3 .288 So. .8 8mm .8 $8.3 33.3
FM... NewN N. . .m non...” 8v... 3. .m .82.an 29.852. :0 ~95 v8.3.3. .80. .0 0.808 .05. .e 0.0.0...
83.8 .308 88.8 - - -
oNed mum... m3... - - - 8 8.0 .30.. 22.832. .. .8 29.859. .0 .80...
G... .8 ASN.8 .N.N.8 - - -
N8... m8... >3... - - - 8 8.0 .800» +w 30.9.80... .8 0.30.. +w n3.39.8
53. .809 case - - -
.86 8nd Nnvd - - - 8 8.0 8.80.. b v0.0.9.8... u. .8 8.00.. h "9:338
.888 83.8 .808 - - -
82. $3 83 - - - 8 3.0 £8. .. 55 a». .3 0A 329:8 .o 3 e8. v. “05.8.8
- - - €38 $3.8 $9.8
- - - RN... and SN... 8 8.0 6.2.8.5.. E 808.. e: .2 .8 22.83.. E 80.3.. oz
- - - .388 .888 .288
- - - e8... .8... .8... 8 8.0 .8030“. 8:08.. 53 a .8 80000.0 8:03.. Gem
- - - 85.8 $3.8 $3.8
- - - and 3N... 38.. 8 8.0 680000.. ~08... h .8 880000.. 8050".
- - - 89.8 68.8 .508
- - - NnN... .MN... am... 8 8.0 680000.. .0508 u. 8 5800008 .0502
- - - 38.8 88.8 88.8
- - - .080 83... End 3.09.8 .e 800.. 8.8050,...
- - - 85.8 .888 33.8
- - - thd SN.” 5mm... 3.02.8 .o 800» {053).
- - - 63.8 .338 .808
- - - $9... Red 836 8 8.0 .30.. 29.83.. .0 2.50 a .8 .30.. .o EEO
$3.8 - - - - 338
Sn... - - - - v3.6 8 8.0 0.080.. u. 8 0.080".
.388 $2.8 88.8 8% . 8 SN? 8 SN. .8
8.. Sn mnwdm mom...“ 5. .0. wood. 80.... 02
Jadllduflljufiaa 802 éfillqflufiul

 

 

38:8 288 a? 22:.

N 592mm:

91

.8. .0 :0 30 run .0 .8 n8... .0 038.....3.» 3.8.008 8.00.3. .0 .26. 3.0.33. 0... .0 00.00820 .0. 3.00:8 000. 0...... 00.08.80
30 0... .0. .30 080.320.. 0. 0.0 8.8.0 E0306 .n 0.00... 0. 3.008.... 5900 0... .00.... 38.80 00.06.80 0... 0.30.... 0.00. 8...... .0082

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

a... 30. 002 0000 00.0 0.00.00
08.0 08.0 80.0 80.0 80.0 8.80.0.0 2.03 .. 000:5
.000 80.0 .000 080 000.0 00.50.00 .8. 000 .0002
.000 - 00. .0 .00. .0 20.0 0000.00.00.20 0.0.9.8..
800 - 800 800 00.0 8.00.00.00.20 000.300.
"00:85:30 30.3% 0:37..
00.0 00.. 00.. 00.. 00.86.... 0.0.3 0 0M0...>
00.. 00.. 00.. 00.. 00.86.... .00.. 30 5:02
00. 00.. 00.. 00.. 00.0.8.6 303.0... 058.80..
50.0 - 0.00.... 20...... 0.00....
08... - 80.0 000... 000.0 08.. 0.00302. 0.0.00".
0.3.0. - .0000. .0000. 0.8.0.
080 - 8.0 02.0 ...0. 002. .0 M00.02.00 .0 as»
.800. - 0.000. .0000. 0.000.
000.0 000.0. 80.0. 800 0200.00 000.. 0.0.0000. .0 00<
0.000. - .0000. .080. .080.
8.0. - 080 08.0 2.00. .08.. 0.000002. .0 0....
.0000. - 2.00. .0000. .800.
080 - 080 000.0. 0000. 0200.00 000. .0 8.0...
.8. .0. :80. .2.. .0. .00. 0.
0.00. ... . .v 0.00 00.0 000. .0 00.0...
.800. 5.0. .802... 6......
:30 - :3- 3.0.0- 0000 0.00800. .0 08..
6.0.0. .800. :80. 3.00.
«.00.? - 080 0.0.9 2.2.0.? 00.0000 00...
.000... - .80... .000... :8...
an. - an? :00 E... 00...
.0... 30 0.0. .mm 0.". 00.00.50. b00850...
.35.. 05.3 V a... 20.0.. a
800.. 8.0... 800... 8.0".
0000 00:. 800 000.03

 

 

 

0:0..005 0.00... .0N.N< 0.00...

93

3.2 2.3 a. 38 8m ...§2

 

 

 

 

 

 

 

 

 

 

«S «gr 5 a; o.
83. o8.c 8o... 8.55% 2a; _. 055
8o... 83 m8... 352. 3...» Ba 58:
- n8... - “3698.20 223.8:
as... 362850 33:5
.caeuaasu 3.3ku 32¢
no» no» no» 8* 8% no» 8» 8* 3.88.6 9.53 _. own—.5
3» no» no» no» 8» no» no» no» 388% :3» can 552
no.» no» no» no» no» no» no» mo» mug Barnes 8383!—
88.3 - ante - 82.: $63
2:- - and. «2 .m a: .v 32 288:2 £80m
sane - $8.9 $3.9 83.8
5.9 - 38...? Ed. :2 33 .o 38:8 «a 38>
98.9 - 88.8 $8.8 38.9
8...? - 88... m8... n8... 853 32 2388.. we ou<
83.8 - A82: 83.8 39.8
$3 - 89? an? Ste BB 2288: .o ou<
$8.8 - 98.9 88.8 88.8
«86. - _8.c n85. 358.9 8.3.: is .o 3.3.
62.8 - €39 ASN.9 age
:86 - $2. 83 2.3 25° 83.
6% c - 83.: $58 98.8
8m _ - an c 5.“ 8c.» 2388.. 5 52a oz
88 e - s3 8 88.9 as a
new a - «S v 8a.”- 8a m 3.88.. saga 58
as a - G8 3 838 $2. a
an n - .38 2 «an n 5 ¢ 838.. Sea
Ea s - $5 3 88.3 as a
me n - 8N n 82. 8o _ .9888“. sec:
88.3 - 5n 8 83.3 62 a
5. _- - «x v 8. .¢ .82: 32 we 36
62.9 - 32 9 22.8 32.8
:_.§.e. - .3 ¢ ”8.... .3 c 853 o?
23.3 - Gen e 369 an a
£38.: - tan a E _ fin V 03
.mm 80 .md 30 4.: 30 .mm 3o 335. $33
53 59‘ n 52‘ IIISEJ g 33 n a: I SE a
83|||=. 8.3m 8.53 gm
2 S a: 2 S E

 

 

 

 

 

 

among—um £3: 5N.N< 035.

.uN.N< 03a“. .8 88.. 8w .6892

 

 

 

94

 

 

 

 

 

 

an. an. on... «an .5 «an ..E 8.... 0.95m

2.... n8... ..8... ..8... ..8... 8.... 8.... 8.5.5... 9...; . 8...;

2.... 8.... 8.... «8... 8.... 0...... 8.... 8.65.... .8. ...: 582

8.... 82. - 8n... - 8.... - 88.883... 22.88..

...m... 2.... «8... 8.. .. 8.8.8.2.... .3338.

”gnaw... 2.33% 337%
8» 8a 8a 8» ma 8.. 8.». 8.2.8.... 26... .. nun—.5
8» 8a 8» 8.. 8a 8» 8.. 8.88.... .8." ...... 5.82
8.. 8.. 8.. 8.. 8» 8a 8.. 8.8.8... 30.38... 9.88:3
.82.. .83.. .2.... 28. ..

.8... .8... Sn... 8.. . .88.. 2388.. 8.28..
28... ......N... .82.. .82..

. 2...... 8.... 8n... :8“... ...... ... 889.8 ... 28>
88.9 88.9 .89 .«89

.8... .8... .8... «8... 8.3... .8. 2285...... a...
5...... 28.... 2.9.9 .82..

5.... 8.... 8...? 8.9 .82. 29.38.... ......
.8... .08.... 88.... 88.9

8... 08.... 8.... n8 ... .23... ...... ... 8.3..
2.2.9 5...... 28.9 88.9

8. .9 8.... ...... ... :8”... ...... ... 88¢.
.83. 2.8.9 - -

:3. on..- - - 228...... ... .8:

- - 2.8. .. .....n. ..
8.... .3 . 29.38.. ... .88.. ..z
.83. .8. .N.
...n... 3.... 8.88.. .88.... 5....
.5... 2.... ..
.2..- 5.... 8.88.. .28..
.83. .8... ..
.3... .....N . 8.88.. 8.82
.8... .. .3... ..
- - 3.. . :83- .89. ... ......o
83.9 28.9 .899 88.9
88.9 ......9 3.8.... 2.88.? 8.2... a?
......9 28.9 G8 .. .3...

8. ... 8...? - ....3... .383 9?.
...... ...... 30 ...... 3o ...... m... 8.33. 33.3.3
94...: 52. a g I 2...... .V g 3.... n 18.8 II 5.... n

8:... I 8...". 8.5: gm
8.: 2...... v.-. 8......6

 

 

 

883.... 88: 8.9. 2......

95

.... ... .... 8.. .... ... .. u...... ... 8.80.8... .888... 8.8.8. . .25. 32...... 2.. ... 88...... a. 88:8 82. 2.... 8.3.8
30 o... .0.. .65 88608.... .... o... .....Eo 335m .v 2...... ... 68:08.... 58.... 2.. 890 63.8.. 888.86 o... .327. 0...... «E... .882

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

3». 3m. 30. 32 8mm .3. 8mm .3. 0&5 208.6
000.0 000.0 000.0 000.0 000.0 000.0 000.0 000.0 8.88.6 263 * own—.5
n80 000.0 0.0.0 000.0 000.0 «00.0 000.0 000.0 8.88.6 .8... 6.... 5.82
63.0 - 00. .0 - .m. .0 - $0.0 - 8.8.8.0820 Bosnian
03.0 - 3.0.0 - 08.0 - m0. .0 - 8:30.085... 8.6365
..uucuuREw... .50....0... 39¢
we» no» 8% 8» 8m 8» no» 8» 8.88.... 963 .MMEEI
no» no» 8.. no» we» 8.. mo» 8.. 8.85.6 .8» 6.6 6.82
8.» m9. 8.. 8.. mo» no» 8» no» 8.88.6 30.30.... 0.8.0.8....
.33... - 68.... - .03.... - .80.... -
$26 - 36.... - N006 - «and - 63.. 6.2.8.5.. 2.3..."—
.n.....9 - .30.... - .890. - 98.9 -
50.0 - mNNd - 3.0 - an. .0. - 68.. ... 0.58:8 .... 88>
$00.0. - $00.0. - $00.0. - €00.00 -
«00.0 30.0. - 80.0. - 80.0 - 62.3.... 6.8.. 6608...... ... 03.
83.0. - 80nd. - $0000 - 83.00 -
mmmd. - «2.0 - 0mm .0 - 05.0. - 68.. 6.2.82... .... on...
800.0. - A. 8.0. - $00.0. $00.0. -
..8... - ..8... - .8... - 8.9 - 85.... 8... ... 8.2..
Ame—.00 - 28.0. - $2.00 - 82.0. -
Sud. - 0S... - 30.0 - a. ..0 - 6....— 00 88<
83.8 - Sufi... - .23.. 80:0 -
9.35.0 - Sad- - 3.06- - 0n. .0 - 2003...... 0.. 6.83
6.0.00 - 88.0. - 200.0. - 8.0.0. -
«00.0. - .80 - 0.0.0. 2.3.0.0. - 68.3.8 om<
ASN.3 - 8:... - CS... Cum... -
3.. - 8...... - G8... - E... - on...
.m.m 30 .md 30 .m.n_ 30 .mm 30 3.3:...» @835me
.53.. .53.. V .58.. .58.. V .58.. .58.. V .53.. .58.. V
83.5 g 8.5: g
0m-0~ .82 0m.0~ 5:63

 

 

 

82.33 8.6: ”3.9. 03a...

96

.am.u< 035. .3.“ 88: com .882

 

 

 

 

 

 

 

 

 

 

 

 

 

 

v5 38 v5 3m 2%
8o... o8.c o8... 3% 0:3 _. 08:5
956 Red 286 388.6 ..8» 9: £82
- 38¢ - «383082... 29.83:
~36 «auteur—Eu .5281:
32358.... so?» :33
8» 8» 8» 8» 8» 8» 8» 8» 8:556 o>a3§>
8» 8» 8» 8» 8» 8» 8» 8» 8:55... 8» in 5.82
8» 8» 8» 8» 8» 8» 8» 8» «388.6 Burr—BE 8382
33.8 - 83.8 82.8 - 82.8
83.. - 3.? 83 :3. BB 2238.. oaaom
A308 - 23.8 63.8 838
8a.? - :33 8: .¢ 83 32. ... 38:8 .8 c3.»
88.8 - $8.8 $8.8 $8.8
88¢. - :56 88c n86. @833 v8: 228:2. we om<
88.8 - :98 89.8 69.8
83 - 9.5.? an? 82. 33 2388: we 0?.
$8.8 - $8.8 88.8 88.8
«8.? - 8... 8a.? 2.18.? was...» E: be 8.2
E: .8 - 38.8 83.8 88.8
956 - wnud wand ...—and an. .8 A8.58.
68.8 - 63.8 63.8 :38
2.3 - an; 25 as. 2238.. 5 32a oz
68.8 - 39.8 890.8 88.8
3v.— - 38v. mad- «SN- 8838 8:88 Son
83.8 - G— _ .8 35.8 89.8
8:6 - .5388 236. End. @0838 85am
ASN.8 - 3.3.8 83.3 83.8
Bed- - 38v. Sad mum; 8883 8:82
33.8 - $3.8 $3.8 2.8.8
30.”- - £86 03 .8 1.8.2- 33 «0 220
GE .8 - 92 .8 ANS .8 Am: .8
31.3.? - agmd. 08.6.. - vm—d v0.3.5» 02
$3.3 - Ené 69.8 - $3.8
£3.0va— - «5.x awh— - «3.7 - 02
.md 30 .mm 30 .md 30 .mm 30 .833» bags
53 53 n 53 '53 n g 53 3'8 .la 8.3 v
9.3.x. 8.8m 8.5.: 88..»
a. £82 2-! nun—cal

 

 

 

885$ 38: so? 2%.»

97

3?. 23. a. 8.8 8m 532

 

 

 

 

 

 

 

 

89 89 82 82 «an 9.2 «an 2.5 36 233m
«85 8o... <8... 8o... 82. So... 8o... 83. £853 253 . $2.5
n8... «3.». ES 83. «85 89° 83 c8... flees 5» =3 582
89c - 5.... - 2.3 - 89c . 83.30220 2283:
and 83 a8... 83. 83.32220 32%...

”8:85.? 2.3.2..» 32¢
8». 8» 8» 8h 8» 8» 8» 8» 8:556 253 ... own—=5
8» 8» 8» 8» 8» 8» 8» 8» 88.35.. 8» Ba 5.82
8» 8» 8» 8» 8» 8» 8» 8» 8:556 its”; 8333
?F9 899 - 88.9 88. 9
$2. an: 83. 83 32 2288.. 255m
28.: 38.: 83.9 38.9
and. o8._ 92. :find 32 .o «5838 .o as.»
68.9 38.9 68.9 88.9
89o _8.o 386 «85 can? 32 3935.. B one.
$4.9 88.9 EH9 68.9
$3 n8... Red. 83. 12 2232 S on...
G89 68.9 38.9 38.9
Sod “8.... «8.9 82.. 853 is 3 83¢
:9 .9 639 98.9 38.9
E9 52. .32. :3"... 82 .o 8.2
639 88.9 - -
8a.»- 8a.». - - 2888.. .8 v8:
- - 88.9 6%. c
82. c3. _- 2288.. s 283 oz
Goad $2 .9
8a.. 33 3.88.. €23 58
$3. 9 so». 9
2.3 3...? v8.88 seam
38.9 88. c
- 33 n2 . 7 838.. .362
- 83.: 8:. 9
- - $2- :33. 32 s 220
33.9 28.9 68.9 88.9
53. 9a.? .38— .? 2:339 E39. 0?.
8:9 $9.9 EN. 9 93.:
55 2.3. - :3.” 533 u?
:9...» So .m.m 30 .ME 30 .9". So 335» 933
g 53 v 815.. ll 53 n 53 53 n QIS I gum
88: :5.» 88: gm
3.: 333. z-» 5.220

 

 

 

 

2.2.93 e8: 2...? use.

.0». a ...: 65. can ... .. ”$0. .a 8:855»... .8553... 8.865
.. .. o. 0 58. 8.5.5; ».556 .8 005...“. 8236 s 8.88.08 523 888.8. 850 .... .6 585 o... a 63.538 06:60 3.338..
5 00520 2 6882.8 58.. 8... 9:... 38.0508 .332...— 5 ”.28. 238.5: 05 .a 9.8830 .5... 688....8 88585... 5 A9.8.8
6.553... 5.3 .32.. 058 6855.8 9... 858M .m 0.5.... 5 63.88... 59.6 05 50¢ 63.5.0 8.55.8 o... 8.327. 038 m5... .882

 

98

 

 

 

 

 

 

£2 «a. 8.. 36 2.58
000.0 000.0 000.0 000.0 000.0 000.0 8.55.6 263 ... own—.5
000.0 000.0 000.0 000.0 000.0 000.0 8.5.56 ..8» 65. 6.82
000.0 000.0 000.0 000.0 000.0 000.0 85.88820 25.8.63
..wuauufiuumu 3.0.30.» 39¢
8» 8» 8» 8» 8» 8» 85.56 82.3 ... own—Er
8» 8» 8» 8» 8» 8» 8.5.56 .8» 65. 5.5.2
8» 8» 8» 8» 8» 8» 85556 303.25 05.808.»
2 .00.. $000.. PSO. $.00. :80. W80.
Am. . .00 G. . 00 300.00 £00.00 €00.00 900.00
08.0. 30.0. 000.0 300 500 v0. .0 68.. 6.5.8.5.. 0.58".
F000. $00.0. $000. .0000. H300. 5.00.
8.0.00 8.0.00 A2000 A2000 A2000 8.0.0.
*mmcd 3.36.». «:360 1896.0 .....ibmod ......imod 03.. ..o 9509.8 ..o 83>
8000.. 3000.. $000.. $000.. $000.. 3000..
3.000 3.00. 8.000 A2000 AS000 3.0.0.
«000. $00. .300. 08.0. 3.300. 800.0. 00. \ 62.53 68.. 25.8.5: .8 um...
$00.0. $000. 5000. $000. 5000. F000.
6.0.00 6.0.00 8.0.00 A3000 020.00 3.000
...wNod nflod ......Nmod uuwNod ..zznmod filmed .03: 20:339. no om<
$0000.. $0000.. 20000. 20000. 20000.. 20000..
28.9 08.9 $08.9 $08.9 $08.9 608.9
9.3800. 3380.0. .000 80.0 m0000. 8000. 62.53 6.5. .0 8.o<
.200. 3.00. 30000.. 30000.. $000. 5000.
@800 £8.00 A2000 9.0.00 6.0.00 6.0.00
22.3.0.0 .1300 .000. .000. v~00 n80 6.8. .6 88<
o n v m N . 8.8.23.83fl
850.. 8...: 850.. 0.5.8". 850.» 25.8.. .0 2a:

 

2:52 o .2.... 5 22.88.. was... .33. S .o 9.528..

6.? 03...».

99

.0. .0 .... as ”an a .100... 8080.000 .8005.
88...... - .. s 0 as. 8305. Sea. a. 005.0 329.... a 088.9. .00.; £88.08 .2.... .0 .0 52. 2.. a 3.305 00...... 00332..
... 00.5.5 0. 00.558 .500 8.0 0.5.. 0.00.0508 8.05.8 .05.... .8 .0550... ... 20.5. 0.000800 0... .0 00.5.8.0 5... 00.00.50 .500 0.5..
8.0808 .38.. o... 5.. 0.5 880.005.. ... 0.5 0.5.5 0.80.5.0 .0 0.00... ... 00.00800 .0050 0... 89.... 00.....5 8.0.5.8 0... 032.0 0.00. 0...... .852

 

 

 

 

0000 0. 00 0000 00.0 0.0800
.000 .000 000.0 000.0 000.0 000.0 8.8800 0.53 ... 00.0:;
000.. 000.. 000.0 000.0 000.0 000.0 8.8.5.0 .50» 0:0 .0002
000.0 0.0.0 000 .0 000 .0 000.0 000.0 800.00.00.20 0.2.00.5:
..00§0S00..0 3.0.0.5» 0300...
8» 8» 8.0 80 8» 8» 8.8500 02.3 .. 000....»
8» 8» 8» 8» 8» 8» 8.80.00 ...5» 0.5 5.52
8» 8» 8» 8» 8» 8» 8.0.050 30.20.... 050008..
- - - - .300. .300.
800.0000. 800.0000. .0000... 800.00. .0000. .0000.
3.0.0.002..- 3.800.000».- 000..0- 000.0- 00.0 00.0 000.. 0.2.00.5: 0.00.0".
- - - - ..000. ..000.
600.000. 200.000. 800.0. .020. 6.00. 6.00.
000.000- 03.000- 000. .- 000..- 230000 000.000 000.. .0 00:00.50 ..0 0.00»
- - - - .0000. .0000.
.0000. 3.0.0. .0000. .0000. .080. 8.0.8
0.3- 05.0- 30.0. 03.9 80.0. 0.0.0. 00. . 0233 08.. 20.88.. 00 00<
- - - - .300. .0000.
800.000. 800.000. .0000. .0000. .0800 .080.
000.00.. 000.02. 000.0 000.0 0.0.0 0.0.0 000.. 050080.. .0 00¢.
- - - - .0000.0.. .8000.
.0.-0.0. C .00. .0000. .0000. 200.0. ..000.
000.0. 000.0- 0.0.0. 000.0. ....000. 0.00.0- 8.00.00 00... .5 85¢.
- - - - .080. .080.
000.000. ..0...00. .020. .00. .0. 6.0.0. 6.00.
000.000 000.000 00. .0 ....0 ...:0000 ...:0000 000. 00 85<
0 0 0 0 0 . 830.53 08.023500
050...... 00...". 050080.. 0.8.0.... 00...". 0.2.00.5: 000.0
0.3.0.595 .05... 080.5... ..0 50:52 .30. 00.... »:<

 

 

 

 

.582 0 08.. ... .33 02... ”0.9. 2%..

100

Table A2.6: Land Cultivated in Last 6 Months, Household Fixed Effects

 

 

 

 

Area of land cultivated Area of land cultivated
per adult resident
Explanatory Variables l 2 3 4
Acres of land 0.573‘" 0575*" 0199*” 0199*”
(0.038) (0.038) (0.018) (0.018)
Acres of land squared 0003”" 0003‘” 0002““ 0002*"
(0.001) (0.001) (0.001) (0.001)
Age of household head 0249*" 0230"" 0095*" 0089*"
(0.069) (0.069) (0.032) (0.032)
Age of household head squared 0002‘" 0002"” -0.001”* -0.001**"
(0.001) (0.001) (0.0004) (0.0004)
Years of schooling of head -0.031 —0.025 -0.012 -0.011
(0.111) (0.111) (0.052) (0.052)
Female household head -2. 142*" -2.097"" -0.280 -0.257
(0.583) (0.584) (0.271) (0.272)
Past/future interview dummies yes yes yes yes
Month and year dummies yes yes yes yes
Village * wave dummies yes yes yes yes
P-value for joint significance:
Household characteristics 0.000 0.000 0.000 0.000
Month and year dummies 0.456 0.478 0.215 0.233
Village * wave dummies 0.000 0.000 0.000 0.003
Sample Size 2323 2323

 

 

 

Notes: This table shows the estimates omitted from the output presented in Table 8. Standard
errors are in parentheses. “ indicates statistical significance at 10%; " at 5%; and "" at 1%.

101

Table A2.7: Crops Farmed by Any Household Member in Last 7 Days. Probit

 

 

 

 

 

Coffee Banana Maize,
Explanatory Variables Cassava or
Beans
Acres of land 0094*" 0072*" 0.034“I
(0.015) (0.015) (0.014)
[0.037] [0.024] [0.011]
Acres of land squared -0.001"""* —0.001"* -0.001"
(0.0004) (0.001) (0.0005)
-0.001] {-0.0005] {-0.0003]
Age of household head [0.022“ 0.013 0026*"
(0.010) (0.010) (0.009)
[0.009] [0.004] [0.009]
Age of household head squared -0.0001 -0.0001 -0.0002***
(0.0001) (0.0001) (0.0001)
{-0.0001] {-0.00003] {-0.0001]
Years of schooling of head 0.0004 0.001 -0.003
(0.013) (0.012) (0.012)
[0.0002] [0.0002] {-0.001]
Female household head -0.274"* -0.176" -0.056
(0.079) (0.081) (0.072)
{-0.108] {-0.061} {-0.019]
Past/future interview dummies yes yes yes
Month and year dummies yes yes yes
Villaje * wave dummies yes yes yes
P—value for joint significance:
Household characteristics 0.000 0.000 0.000
Month and year dummies 0.000 0.000 0.000
Village "' wave dummies 0.000 0.000 0.000
Sample Size 2323 2889 2893

 

Notes: This table shows the estimates omitted from the output presented in Table 9.
Standard errors are in parentheses and have been corrected for clustering at the
household level; in brackets coefficients have also been converted to change in
probability (dp/dx) evaluated at the mean of all other regressors, which represents a
discrete change for dummy variables from 0 to l. " indicates statistical significance at

10%; *" at 5%; and “’" at 1%.

102

Table A2.8: Wage Employment and Nonfarm Self-employment of Men (20-50)

 

 

 

 

 

Nonfarm
Wage Ermloyment Self-employment
Explanatory Variables l 2 3
Age 0.058 0.058 0211*”
(0.044) (0.044) (0.049)
[0.020] [0.020] [0.053]
Age squared -0.001 -0.001 —0.003”*
(0.001) (0.001) (0.001)
{-0.0003] {-0.0003] {-0.001]
Schooling: 1-6 years 0247 -0.250 0768*”
(0.209) (0.208) (0.246)
{-0.083] {-0.084} [0.225]
Schooling: 7 years 0116 -0.117 0.868"“"'I
(0.227) (0.227) (0.275)
{-0.041} {-0.041] [0.208]
Schooling: 8+ years 0253 -0.255 0.540
(0.287) (0.287) (0.351)
{-0.083} {-0.084] [0. 162]
Head of household 0.281 0.282 0.124
(0.190) (0.191) (0.213)
[0.097] [0.097] [0.031]
Acres of land 0064" 0064" 0.005
(0.028) (0.028) (0.024)
{-0.022} {-0.022} [0.001]
Acres of land squared 0.004“ 0.004" 0.001
(0.002) (0.002) (0.001)
[0.001] [0.001] [0.00002]
Age of household head 0008 -0.008 0.010
(0.020) (0.020) (0.021)
{-0.003] {-0.003} [0.002]
Age of household head squared 0.0002 0.0002 -0.0002
(0.0002) (0.0002) (0.0002)
[0.00001] [0.0001] {-0.00004]
Years of schooling of head 0084*” 0.083"“”'I -0.043
(0.024) (0.023) (0.029)
[0.029] [0.029] {-0.011]
Female household head 0.216 0.213 -0.034
(0.181) (0.180) (0.220)
[0.078] [0.077] {-0.008]
Past/future interview dummies yes yes yes
Month and year dummies yes yes yes
Village “ wave dummies yes yes yes
P-value for joint significance:
Individual characteristics 0.077 0.073 0.000
Household characteristics 0.006 0.006 0.545
Month and year dummies 0.000 0.000 0.000
Villag“ wave dummies 0.000 0.000 0.000
Sample Size 1793 1622

 

 

 

Notes: This table shows the estimates omitted from the output presented in Table 10.
Probabilities are estimated with standard errors in parentheses corrected for clustering at
the household level; in brackets coefficients have also been converted to change in
probability (dp/dx) evaluated at the mean of all other regressors, which represents a
discrete change for dummy variables from 0 to 1. ‘ indicates statistical signifieance at

10%; " at 5%; and "“ at 1%.

APPENDIX 3

APPENDIX 3

Table A3. 1: Impact of Adult Household Member Deaths (15-50) on Farm Hours, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15—19 15-19 7-14 51-65
Female death past 0-12 months -1.779 0.558 -0. 125 -0.533 —0.882 -0.363
(1.258) (1.621) (1.383) (1.781) (0.683) (1.871)
Female death future 0-12 months -3.771“ -3.200 -2.822 -3.107 -1.197 -6.040"
(1.809) (2.006) (2.058) (2.012) (0.892) (2.914)
Male death past 0-12 months -3.010** 0.627 -3.536 -2.234 -3.296*** -3.185
(1.444) (1.971) (1.940) (1.984) (0.838) (2.091)
Male death future 0-12 months 1.572 —0. 134 -2.172 3.493 0.309 -1.972
(2.021) (2.932) (2.607) (2.658) 11 .042) (2.880)
R2 0.209 0.228 0.322 0.278 0.171 0.225
Breusch-Pagan test 38(1) 49.57 139.80 22.23 82.98 110.18 73.16
J-value 0.000 0.000 0.000 0.000 0.000 0.000
Hausman test xz 136.59 100.79 43.06 135.24 108.66 167.38
p-value 0.993 1.000 1.000 0.995 1.000 0.741
Household-member death variables:
Hausman test x2 2.228 0.365 33.481 2.775 3.630 3.604
p-value 0.694 0.985 0.481 0.596 0.458 0.462
P-value for joint significance:
Female death past & future 0.050 0.219 0.390 0.302 0.203 0.117
Male death past & future 0.067 0.943 0.159 0.149 0.000 0.288
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: Standard errors are in parentheses. * indicates statistical significance at 10%; *"‘ at 5%; and ""' at 1%. See

Table A32 for additional regressors included but not reported here.

103

104

Table A3.2: Farm Hours Equations, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Age 0.706" -0.435* 2.919 8.712” 3971*" -2.877
(0.284) (0.228) (4.076) (3.554) (0.769) (2.438)
Age squared 0013*" 0.004 -0.078 -0.279" 0133*" 0.022
(0.004) (0.003) (0.126) (0.109) (0.037) (0.021)
Female - - - - 2582*" 10.586"'"
- - - - (0.359) (1.551)
Child of head - - -0. 183 -0.569 -0.234 -
- - (1.762) (1.548) (0.744) -
Mother’s years of schooling - -0.019 -0.228* 0.024 -
- - (0. 152) (0. 136) (0.074) -
Father’s years of schooling - - -0.207 0.054 —0.047 -
- - (0.151) (0.132) (0.079) -
Mother deceased - - 1.243 0.319 0.566 -
- - (1.130) (0.938) (0.611) -
Father deceased - - 3063" -1.277 -1.130"I -
- - (1.248) (1.106) (0.600) -
Both parents deceased - - -0.259 1.942 —0.240 -
- - (1.975) (1.756) (0.994) -
No parent in household - - 0.236 0.912 1.247 -
- - (1.774) (1.596) (0.804) -
Schooling: 1-6 years -0.649 -0.356 - - - -2.408*"‘
(0.939) (1.085) - - - (1.145)
Schooling: 7 years 0474 -1.147 - - — -2.793
(1.022) (1.223) - - - (2.125)
Schooling: 8+ years -2.114 -l.605 - - - -4.134
(1.832) (1.618) - - - (3.092)
Head of household 1.863 1.925" - - - -0.052
(1.443) (0.961) - - - (1.404)
Acres of land -0.279“”‘l -0.183 0.299' -0.264** -0.028 -0.228
(0.128) (0.119) (0.174) (0.131) (0.071) (0.150)
Acres of land squared 0.004 0.010 0013" 0.005 0.001 0.007
(0.004) (0.006) (0.006) (0.004) (0.002) (0.004)
Age of household head -0.115 0.051 0303"" -0.022 -0.056 0.049
(0.123) (0.093) (0.149) (0.139) (0.077) (0.201)
Age of household head squared 0.001 -0.0002 0.002 -0.0003 0.0003 0.002
(0.001) (0.001) (0.001) (0.001) (0.001) (0.002)
Years of schooling of head -0.093 -0.279" -0.207 0223 0.032 0.112
(0.132) (0.127) (0.180) (0.164) (0.094) (0.248)
Female household head 2989" -1.750* 1.714 0.741 1.151“ 0.216
(1.190) (0.927) (1.263) (1.179) (0.608) (1.580)
Past/future interview dummies yes yes yes yes yes yes
Month and year dummies yes yes yes yes yes yes
Village * wave dummies yes yes yes yes yes yes
P-value for joint significance:
Individual characteristics 0.004 0.016 0.061 0.023 0.000 0.000
Household characteristics 0.017 0.018 0.023 0.041 0.176 0.000
Month and year dummies 0.017 0.011 0.435 0.048 0.126 0.812
VillaE“ wave dummies 0.000 0.000 0.000 0.000 0.000 0.000
Sample Size 2596 1894 1254 1204 3912 1393

 

Notes: This table shows the estimates omitted from the output presented in Table A3.1.

parentheses. * indicates statistical significance at 10%; " at 5%; and “* at 1%.

Standard errors are in

105

Table A3.3: Impact of Adult Household Member Deaths (15-50) on Farm Hours, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
ErgplanatolVariables 20-50 20-50 15-19 15-19 7-14 51-65
Female death past 0-6 months —2.779* —0.099 0869 -0.738 -1.468“ -0.960
' (1.538) (1.959) (1.775) (1.991) (0.822) (2.390)
Female death past 7-12 months 0025 1.644 0.928 -0.975 0.110 0.543
(1.959) (2.415) (2.043) (2.833) (1.050) (2.607)
Female death future 0-6 months 4.313" -1.286 -3.499 0.820 0.418 -3.439
(2. 120) (2.394) (2.577) (2.388) (1 .059) (3.632)
Female death future 7-12 months -2.752 6598" -1.847 -6.528*“‘ 4437*“ -9.834**
(2.857) (3.041) (3.235) (2.774) (1 .420) (4.303)
Male death past 0-6 months -3.396"' -0.114 -2.777 -2.190 -2.847"* -4.463"'
( l .785) (2.327) (2.440) (2.352) (1 .038) (2.602)
Male death past 7-12 months -2.051 2.066 -4.743* -0.954 -3.569"* -1 .352
(2. 199) (2.850) (2.832) (2.872) (1.242) (3 .006)
Male death future 0-6 months 5.450" —0.708 -2.567 4.700 1.843 -1.824
(2.371) (3.478) (2.831) (3.110) (1.245) (3.438)
Male death future 7-12 months -7.573"* 0.905 -0.296 -0.735 -2.627 -2.286
(3.515) (4.469) (5.825) (4.306) (1.683) (4.549)
R1 0.214 0.229 0.322 0.284 0.177 0.227
Breusch-Pagan test {(1) 50.03 139.61 22.46 84.31 114.09 73.73
p-value 0.000 0.000 0.000 0.000 0.000 0.000
Hausman test 11’ 5783.61 123.10 2259.72 197.31 74.03 97.17
p-value 0.000 1.000 0.000 0.254 1.000 1.000
Household-member death variables:
Hausman test 71’ 17.25 3.720 12.037 5.912 12.260 20.265
p-value 0.001 0.714 0.150 0.657 0.140 0.009
P-value for joint significance:
Female death past 0.194 0.785 0.788 0.896 0.193 0.891
Female death future 0.081 0.093 0.351 0.053 0.006 0.058
Female death past & future 0.082 0.234 0.630 0.204 0.010 0.212
Male death past 0.122 0.753 0.155 0.642 0.001 0.224
Male death future 0.005 0.953 0.663 0.297 0.078 0.786
Male death past & future 0.004 0.955 0.374 0.381 0.000 0.533
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: Standard errors are in parentheses. " indicates statistical significance at 10%; " at 5%; and ”" at 1%. See
Table A3.4 for additional regressors included but not reported here.

106

Table A3.4: Farm Hours Equations, Random Effects Estimates

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Age 0.311 -0.478 3.118 9.430" 4013*" -0.052
(0.243) (0.355) (3.951) (4.466) (0.703) (2.903)
Age squared -0.002 0.007 -0.064 -0.267"' 0146"" 0.001
(0.004) (0.005) (0. 122) (0. 137) (0.034) (0.025)
Female - - - - 1133”" 1.346
- - - - (0.312) (1.841)
Child of head - - -2.059 -0.912 -0.828 —
- - (1 .663) (2.237) (0.664) -
Mother’s years of schooling - -0.114 -0.025 0.160" -
- - (0.143) (0.210) (0.064) -
Father’s years of schooling - - -0.084 0.308 -0.126"‘ -
- - (0. 1430 (0.202) (0.069) -
Mother deceased - - 1.715 -0.807 -0.412 -
- - (1.068) (1.431) (0.537) -
Father deceased - - -2.143" -0.640 -0.931‘ -
- - (1.177) (1.661) (0.529) -
Both parents deceased - - 2.334 -0.624 1.140 -
- - (1.870) (2.590) (0.881) -
No parent in household - - -2.632 0.526 -0.271 -
- - (1 .675) (2.268) (0.720) -
Schooling: 1-6 years -0.294 -2.925“ - - - -1.075
(0.800) (1.704) - - - (1.351)
Schooling: 7 years 0.333 -3.474* - - - -1.947
(0.873) (1.911) - - - (2.509)
Schooling: 8+ years -l.401 -8.495"* - - - -2.243
(1.568) (2.514) - - - (3.653)
Head of household 1.606 1.007 - - - 2.732
(1.242) (1.479) - - - (1.670)
Acres of land 0209* -0.098 0.473‘" 0.357“ 0202*" 0.444”
(0.113) (0.176) (0.169) (0.169) (0.065) (0.184)
Acres of land squared -0.005 0.009 0017*" -0.001 -0.001 -0.008
(0.004) (0.009) (0.006) (0.005) (0.002) (0.005)
Age of household head -0.088 0.113 -0.094 0.126 0.033 «0.355
(0.106) (0.143) (0.141) (0.208) (0.038) (0.239)
Age of household head squared 0.001 -0.001 0.0005 -0.001 -0.001 0.004
(0.001) (0.001) (0.001) (0.002) (0.001) (0.002)
Years of schooling of head -0.062 -0.014 0.002 -0.112 -0.047 0.265
(0.1 13) (0.197) (0.170) (0.242) (0.083) (0.293)
Female household head 3210"" 0.663 1.102 -2.251 0.498 -0.560
(1.024) (1.436) (1.195) (1.739) (0.539) (1.879)
Past/future interview dummies yes yes yes yes yes yes
Month and year dummies yes yes yes yes yes yes
Villge * wave dummies yes yes yes yes yes yes
P—value for joint significance:
Individual characteristics 0.000 0.012 0.000 0.121 0.000 0.830
Household characteristics 0.008 0.714 0.053 0.019 0.000 0.053
Month and year dummies 0.015 0.013 0.500 0.206 0.108 0.826
Village * wave dummies 0.000 0.000 0.000 0.000 0.000 0.000
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: This table shows the estimates omitted from the output presented in Table A3.3.

parentheses. "' indicates statistical significance at 10%; “ at 5%; and "* at 1%.

Standard errors are in

APPENDIX 4

---v. an, '

APPENDIX 4

Table A4. 1: Impact of Adult Household Member Deaths (15-50) on House Hours, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Female death past 0-12 months 0973 1.966‘ -0.139 -1.781 -0.579 0.272
(1.425) (1.105) (1.430) (1.386) (0.741) (1.534)
Female death future 0—12 months -3.766"' 0.749 -0.688 -0.604 0.203 -1.621
(2.042) (1.368) (2.123) (1.585) (0.965) (2.387)
Male death past 0-12 months 0447 —0.208 0.759 -0.853 -1.164 -0.205
(1.636) (1.336) (2.009) (1.536) (0.913) (1.714)
Male death future 0-12 months 0.016 5.061" -0.499 1.711 1.139 0.140
(2.278) (2.004) (2.689) (2.093) (1. 129) (2.356L
12.2 0.126 0.181 0.287 0.234 0.123 0.201
Breusch-Pagan test {(1) 100.81 75.45 33.75 6.83 173.32 92.72
p-value 0.000 0.000 0.000 0.009 0.000 0.000
Hausman test 112 99.60 255.34 1051.18 106.85 954.26 42.70
p-value 1.000 0.000 0.000 1.000 0.000 1.000
Household-member death
variables:
Hausman test 12 13.653 6.687 7.484 3.212 2.006 6.608
p-value 0.008 0.153 0.112 0.523 0.735 0.158
P-value for joint significance:
Female death past & future 0.124 0.199 0.946 0.399 0.700 0.764
Male death past & future 0.466 0.033 0.903 0.600 0.204 0.989
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: Standard errors are in parentheses. " indicates statistical significance at 10%; “ at 5%; and "" at 1%. See

Table A42 for additional regressors included but not reported here.

107

108

Table A4.2: House Hours Equations, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Age 0.706" -0.435* 2.919 8.712“ 3.971‘” -2.877
(0.284) (0.228) (4.076) (3.554) (0.769) (2.438)
Age squared 0013"“ 0.004 -0.078 0279” 0133*" 0.022
(0.004) (0.003) (0.126) (0.109) (0.037) (0.021)
Female - - - - 2582"" 10.586”'"
- - - - (0.359) (1.551)
Child of head - - -0.183 -0.569 -0.234 -
- - (1.762) (1.548) (0.744) -
Mother’s years of schooling - - -0.019 -0.228" -0.024 -
- - (0. 152) (0. 136) (0.074) -
Father’s years of schooling - - -0.207 0.054 -0.047 -
- - (0.151) (0.132) (0.079) -
Mother deceased - - 1.243 0.319 0.566 -
- - (1.130) (0.938) (0.611) -
Father deceased - - 3063“ -1.277 -l.130"' -
- - (1.248) (1.106) (0.600) -
Both parents deceased - - 0259 1.942 -0.240 -
- - (1.975) (1.756) (0.994) -
No parent in household - - 0.236 0.912 1.247 -
- - (1.774) (1.596) (0.804) -
Schooling: 1-6 years 0649 -0.356 - - - -2.408"
(0.939) (1.085) - - - (1 . 145)
Schooling: 7 years -0.474 -1 . 147 - - - -2.793
(1.022) (1.223) - - - (2.125)
Schooling: 8+ years -2.114 -1.605 - - - -4.134
(1.832) (1.618) - - - (3.092)
Head of household 1.863 1.925” - - - -0.052
(1.443) (0.961) - - - (1.404)
Acres of land 0279” -0.183 0.299‘ -0.264" -0.028 -0.228
(0.128) (0.119) (0.174) (0.131) (0.071) (0.150)
Acres of land squared 0.004 0.010 0013“ 0.005 0.001 0.007
(0.004) (0.006) (0.006) (0.004) (0.002) (0.004)
Age of household head -0.115 0.051 -0.303” -0.022 -0.056 0.049
(0.123) (0.093) (0.149) (0.139) (0.077) (0.201)
Age of household head squared 0.001 -0.0002 0.002 -0.0003 0.0003 0.002
(0.001) (0.001) (0.001) (0.001) (0.001) (0.002)
Years of schooling of head -0.093 -0.279" -0.207 -0.223 0.032 0.112
(0.132) (0.127) (0.180) (0.164) (0.094) (0.248)
Female household head 2989" -1.750* 1.714 0.741 1.151" 0.216
(1.190) (0.927) (1.263) (1.179) (0.608) (1.580)
Past/future interview dummies yes yes yes yes yes yes
Month and year dummies yes yes yes yes yes yes
Village * wave dummies yes yes yes yes yes yes
P-value for joint significance:
Individual characteristics 0.004 0.016 0.061 0.023 0.000 0.000
Household characteristics 0.017 0.018 0.023 0.041 0.176 0.000
Month and year dummies 0.020 0.019 0.002 0.001 0.006 0.002
Village * wave dummies 0.000 0.000 0.000 0.000 0.000 0.031
Sample Size 2596 1894 1254 1204 3912 1393

 

Notes: This table shows the estimates omitted from the output presented in Table A4.1.

parentheses. ‘ indicates statistical significance at 10%; *“ at 5%; and ‘” at 1%.

Standard errors are in

109

Table A4.3: Impact of Adult Household Member Deaths (15-50) on House Hours, Random Effects Estimates

 

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Female death past 0-6 months —1.456 2.202 -1.121 ~2.033 -0.494 ~I.332
(1.730) (1.355) (1.822) (1.624) (0.882) (1.340)
Female death past 7-12 months -0.425 1.613 1.221 -1.416 -0.809 2.260
(2.201) (1.677) (2.094) (2.312) (1.122) (2.111)
Female death future 0-6 months -2.205 0.581 -1.088 0.123 0.517 2.635
(2.376) (1.658) (2.642) (1.973) (1.133) (2.994)
Female death future 7-12 months -5.547* 1.045 -0.423 -1.269 -0.425 -1.418
(3.202) (2.112) (3.315) (2.325) (1.515) (3.484)
Male death past 0—6 months 0.714 -0.541 0.525 -2.385 -0.663 -0.857
(2.010) (1.597) (2.510) (1.899) (1.119) (2.112)
Male death past 7-12 months -1.728 0.134 0.967 1.881 -1.847 0.414
(2.469) (1 .974) (2.904) (2.393) (1 .328) (2.434)
Male death future 06 months 1.864 2.174 -1.773 1.507 1.551 -3.177
(2.662) (2.406) (2.910) (2.531) (1.336) (2.787)
Male death future 7-12 months 4132 10.067"" 5.174 1.019 0.552 6.227“
(3.942) (3.110) (5.975) (3.599) (1.797) (3.684)
R2 0.128 0.183 0.290 0.238 0.123 0.209
Breusch-Pagan test 38(1) 101.95 75.72 34.31 7.19 172.69 94.93
p—value 0.000 0.000 0.000 0.007 0.000 0.000
Hausman test 1’ 415.49 65.08 1273.81 7422.35 404.55 34.58
p-value 0.000 1.000 0.000 0.000 0.000 1.000
Household—member death variables:
Hausman test xz 22.272 19.029 30.447 16.537 8.477 10.373
p-value 0.004 0.015 0.000 0.035 0.388 0.240
P-value for joint significance:
Female death past 0.698 0.193 0.676 0.393 0.695 0.389
Female death future 0.151 0.846 0.914 0.860 0.848 0.643
Female death past & future 0.371 0.495 0.915 0.707 0.892 0.601
Male death past 0.712 0.937 0.933 0.305 0.352 0.896
Male death future 0.426 0.005 0.546 0.809 0.503 0.094
Male death past & future 0.662 0.023 0.846 0.547 0.405 0.287
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: Standard errors are in parentheses. "' indicates statistical significance at 10%; " at 5%; and "'" at 1%. See
Table A4.4 for additional regressors included but not reported here.

110

Table A4.4: House Hours Equations, Random Effects Estimates

 

 

 

 

Women Men Women Men Children Adults
Explanatory Variables 20-50 20-50 15-19 15-19 7-14 51-65
Age 0.700“ -0.430* 3.032 8.774" 1979*" -2.978
(0.283) (0.226) (4.081) (3.562) (0.770) (2.442)
Age squared 0013"" 0.004 -0.082 0280” 0133"" 0.022
(0.004) (0.003) (0.126) (0.1 10) (0.037) (0.021)
Female - - - 2578*" 10.721”*
- - - (0.359) (1.555)
Child of head - - -0.132 -0.463 -0.236 -
- - (1.755) (1.558) (0.744) -
Mother’s years of schooling - - -0.014 -0.226* -0.023 -
- - (0.151) (0.137) (0.074) -
Father’s years of schooling - - -0.212 0.052 -0.048 -
- - (0.151) (0.133) (0.079) -
Mother deceased - - 1.267 0.347 0.561 -
- - (1.126) (0.945) (0.611) -
Father deceased - - 3026" -1.311 -1.134* -
- - (1.244) (1.113) (0.600) -
Both parents deceased - - -0.261 1.898 -0.212 -
- - (1.969) (1.766) (0.995) -
No parent in household - - 0.278 0.988 1.233 -
- - (1.767) (1.605) (0.805) -
Schooling: 1-6 years -0.656 -0.357 - - - -2.448**
(0.935) (1.078) - - - (1.148)
Schooling: 7 years -0.492 -1.189 - - - ~2.766
(1.019) (1.215) - - - (2.131)
Schooling: 8+ years -2.103 -1.673 - - - -3.988
(1.826) (1.608) - - - (3.099)
Head of household 1.920 1.941" - - - 0.110
(1.439) (0.956) - - - (1.408)
Acres of land -0.278*"‘ -0. 179 0.302“ -0.277"‘* -0.027 -0.225
(0.128) (0.119) (0.175) (0.132) (0.071) (0.150)
Acres of land squared 0.004 0.009 -0.013** 0.005 0.001 0.007
(0.004) (0.006) (0.006) (0.004) (0.002) (0.004)
Age of household head -0.114 0.056 -0.301" -0.034 -0.055 0.031
(0.122) (0.093) (0.148) (0.141) (0.077) (0.202)
Age of household head squared 0.001 -0.0003 0.002 -0.0002 0.0004 0.002
(0.001) (0.001) (0.001) (0.001) (0.001) (0.002)
Years of schooling of head -0.093 —0.274** -0.193 -0.224 0.035 0.120
(0.132) (0.126) (0.180) (0.165) (0.094) (0.248)
Female household head -3.024** -1.730"‘ 1.736 0.814 1.531" 0.063
(1.188) (0.922) (1.260) (1.188) (0.609) (1.585)
Past/future interview dummies yes yes yes yes yes yes
Month and year dummies yes yes yes yes yes yes
Village * wave dummies yes yes yes yes yes yes
P-value for joint significance:
Individual characteristics 0.004 0.014 0.063 0.024 0.000 0.000
Household characteristics 0.016 0.020 0.023 0.032 0.171 0.000
Month and year dummies 0.017 0.022 0.002 0.001 0.005 0.002
Village * wave dummies 0.000 0.000 0.000 0.000 0.000 0.020
Sample Size 2596 1894 1254 1204 3912 1393

 

 

Notes: This table shows the estimates omitted from the output presented in Table A4.3. Standard errors are in

parentheses. * indicates statistical significance at 10%; '"' at 5%; and "* at 1%.

APPENDIX 5

.22. 8:82 .8 :5 83282 ease“: 2.52.2.3 é a? 2.3 8m .2: a t. we in a t is a 3.28%? .8223”

8.828.. . 2962 2.32382 65 a 2233.6 5 88:8 83 2,2 8.8.228 30 65 he 2.5 88558 a pa sea 325m .662

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

«.82 4:. 82 4: $2 43 $2 «a 82m 2288
~85 $3 3.3 985 :3 :3 832 a: .o sea 3» ea 58.. as:
83 33 «86 89° 2.85 285 23 23. 658 6.2 ea 5.82. 63862

“8.283%?“ EBTQ 39¢

23 23 and 83 83 83 and $3 ”a
22.3 nuns 28.8 2.8.3 28.9 63.0 98.9 25$

:3 mafia 3.2.? 22.0 .28.? 82 .c 2.8%: on? 458:. N3 238 582. ca:
fine 92. c and 83$ 83$ 93.2 22 .3 83.9

.2:- $2.2- 82- 28.2 8m? 83 33 RS 22.85 «To as; 5320 632
228.5 83.5 63$ 2.33 53.2 28.0 88.9 28.0

EH? and- .08.»- «8.2 3.5.2 4.42.3.2 2.3- E: 3.88 «I. 65.2 582. 6:862
@943 82.. a $3.9 a: .c 23.9 68.3 $8.3 89.5

23.2- 83 $3 an.” 83 83 and. 83 €85 «To 23 582. 28.62
.mm 30 .2.”. 30 .22 30 .22 So $33» .2633
.953 .333 V .2533 .933 V .933 239‘ V 9.3 .933 V
8.5m 8.8% 8.5: 8.3m
293 :02 8.8 5:83

 

 

 

32.26 .8 652225 52282 328 3.29.82 26 295m

.28: 8 83: mean 386: 2288: .33. do case :2 .2 2,3

m Nun—lam:

111

112

.2 .2 035. .8 85: 98 .6832

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

$3- 2“ $3- 8: 82- 2m 8s an 3a 295m
mad $3.8 63 «8.8 28.8 83 83 on: sea a Ea 582. as:
8.8 «3.8 $38 a; 238 $3 83° 8nd 233 a. 28 582. 0.253
..ugokama ~50T§ Riki
83 8.3 :38 2.3 38 5.8 83c 38 Na
$8.3 $8.3 2:.3 $8.3 $8.3 $8.3 $8.3 38.3
=de 82 ~88 33m 3.38 2.2%.:- 35- 83 was... «I. see 582. 0:2
$5.3 $3.3 $8.3 $33 28.3 $8.3 $8.3 2 8.3
33- 88.8 82 .o- 9.3? an." 88.8 82- 82- $85 3.8 .23 582. on:
28.3 $8.3 $8.3 $8.3 EH3 88.3 833 $83
83- 82w 23- :3 and an”- 83- 33 €88 38 see 5.3 2283
$8.3 $8. 3 $8.3 22.3 22.3 23.3 2333 23.3
82.3 83 8... 33? 88.2 83- 83. an: .585 3.8 .23 532. 298m
.83 30 .mm 30 .mm 30 .mm 30 $8.5» $3.5
.333 9:23 V ...-33% .933 V 939‘ 939‘ V .933 939‘ V
canon-m 5.3m 3:0:
2-3 52 2-2 8:53

 

 

 

8:30 “on 8335qu 5:32 323 8.2.3.8: 3o 295m

.38: 8 633 382 .352 22383 :22 2o 33.5 x: .2 23

113

.2 .2 033. 8.. 88: 8m .6892

 

 

 

 

 

 

 

 

 

 

 

 

 

a? am am» mm 82 82 82 82 3a 295m
23 as; 33 83 «nod 33 ES and 833 a Ea 58o 0::
is 93 $3 $3 $3 $3 and 323 see a. .28 58: ”neon
..uuzuoKEu? 3.3ku 8.37%
:2 83 83 and 82 3: .o 33 a: .o R
3:5 68.8 3:9 38.9 23.3 aid $3.5 88. :
Rho 83. as..- $2 an? 32 8a.? 3a. _- 3.85 26 833 58.. 232
303 § fl .c 333 93.5 833 33.: $3.: $3.9
£2 3: ~25 £2- :8“? 83- 83- SN. 7 “.58... NZ :8 58.. 2a:
82 .3 $3.5 3.3.9 635 83. c 98.3 Eb: 28.3
$3. 93- GE. 33 a: .o. and 8.3. so; 5.55 NE 232 58a 28E
83.9 83.3 $2 .3 $3.9 can: 8%. c 38.: 63.9
83.. .53 3?. Sn? at? .83- 3Q? 25.? 2.8:. «To 23 58c 255m
.mm 30 .mm 30 .mm So m...“ So 333» €ng
.930: 939‘ V .933 Exec V .933 939‘ V 939‘ 335‘ V
owns: gum own—OE Shaw
3-: 933 2-5 5:220

 

 

 

«:96 5a uhaacogm 5:82 323 €353.30: .8 295m

.38: 8 63: mean .852 22.3.8: .32 .8 .895 2 .2 2.5

114

...: a .... .5 ”an a .. £2 a 8:855“ 8.8% 8.8.8. .. .32 .3233 2.. a 55.3% .8 38:8 83 9,2 .8558
30 o... 8. .65 88635.5 8: 828 ...—gm 3.2 0.895 68:08.... 5.562.. 52. .8on gang..." 038 «E... .832

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

8o. 3.. So. 3.. m3. v3 92 v3 ofim 0.953
one... coed «8.6 coed god 889 Sod coed 8:556 263 Uni...»
«mud 8c... mun... 8c... and 896 «8.0 coo... 8:556 :8» 65 5:22
N86 - n86 - cum... - Sod - 83.83855 6.2.83...
m8... - .98 - «and - 2nd - 8.588220 .5636...
..3§u.§.m.fi 5.8.5.8. §~§i
8» 8» 8» 8» 8» 8» 8» 8» 8:556 263 .. $2.5
8» 8» 8» 8» 8» 8» 8» 8» 8:556 53» 65 6.8.).
8» 8» 8» 8» 8» 8» 8» 8» 8:556 30.335 058:5
€8.28 - €348 - 8R8 - 63.8 -
«no... - 30. .N - Rod- - movd - 68.. 22.83.. 0.58".
88. .8 - Rec. 8 - $3.8 - $5.8 -
.26. - SYN- - Sod - m3... - 68.. ..o «5.8—.8 ..o 88>
85.8 - 68.8 - 98.8 - 68.8 -
.599 - ”8.9 - 89¢ - 8o... 83:3 22. 223:2. .o a?
63.: Sand - 33.8 - $3.8 -
.....NQN - .86 - mm: .c. - wand. - 68.. 6.2.3.3.. ..o om<
$8.9 - $8.9 - $8.9 - $8.9 -
.86 - v8.0. - .86. .86. - 85:3 65. .o 88<
$89 - $3.9 - 88.9 - $8.9 -
good. - ~36 - .86 - .2.... - 65.. ..o 8.8<
$2.28 - @368 - $9.8 - 38.8 -
...:mmodn - 8a.»..- - SQN - can... - 6.3%.. we 68:
88.8 38.8 - 88.8 - 68.8 -
Sod. - wmod - mad. - n86. - 623.3. om<
83. .V 38.8 - 63.8 - 8. ..8 -
o. md - 8nd. - 8nd. - i.e.. - om<
4...". 30 .md 30 .mm. 30 .m.n. 30 8.888» .0235me
2:3. .233 V .239. 2:9. V .239. 835. V has. .233 V
84.3. .5451“ 8:0: :Icfl
8.8 :02 3.3 5:63

 

 

 

8:50 .2. 8.53% .5602 30.3 6.2.83: .o 0.556 .8o_6=cm 8:3. ”3.2 as“...

.QN.W< OZGH. .Ou use 00% .60st

 

 

 

115

 

 

 

 

 

 

 

 

 

...m 3.. 2n 8.. m. an» n n 8am 0.3%
. 8o... men... 9...... 2...... 2...... ..8... 9...... 8:556 983 {85.5
2...... n... ... 2...... 2...... 8°... .3... 8.... 8.55.6 .8» 65. 65.2
- 8. ... - .nv... - 8...... - «3888.30 6.2.83...
3»... - n3... - ..8... - 8:882:30 .5656...
..u§§M.u 5.3.36» §~6>..~
8» 8» 8» 8» 8» 8» 8» 8.556 263 ... cum...»
8» 8» 8» 8» 8» 8» 8» 8.5.56 .8» 65 .552
8» 8» 8» 8» 8» 8» 8» 8.5.56 3o.>.85 8583....
- 98.... - 88.... - 8.8 -
n88. - 8. .N- - 3...? - 68.. 6.2.8.5.. 0.55.".
.2... - EN... - .3»... -
82.. - 8a..- - 8n... - 63.. ... 3.8.8 .o 33.»
...... - 8...... - ......8 -
2...... - .8... - 8...... - 605....Um 68.. 6.2.8.2. .6 om<
.8... - 68.... - 86...... -
«8.. - o5? - 3.... - 63.. 6.288.. 3 a...
n3... - .88 - .88 -
...... - 86... - 68.... - 683... E... .o .83.
non... - .3... .8 - $3.8 -
we. . .- - 8. ... - 8m... - 65.. .6 8.3..
n8... - ....od. - $8.8 -
”R... - 3.... - 3.... - 6.888.. ... 32... oz
n86 - .54.... - .82.. -
urn... - 8N... - $6.» - 68886 8:82. 6....
8m... - .858 - 8...:
n8... - 3nd- - 3nd- - 68886 .2...."—
cmmd. - 83.8 - 6.08 -
86.. - 3...... - 3...... - 68.306 .2662
866 - .328 - ......8 -
8...? - End - m8... - 68.. .6 6......
g... - .82.. - 88.... -
3n? - 82.. - m8... - 605... 02.
Sn.» - ..n. .... - .53.». -
.8... - .3... - 83 - 08
30 .m..n. 30 .m... 30 .m... 30 $.35. .0633
933V EaglugV 8 53V BacallusgV
5...". 83'... 5......
3.! g 5-5 5.5%

 

 

 

5.90 .2. 83.....on 5.60:. 320.. £288: .o 0.9.5. .8833 28.. ”3.2 2...:

116

8.? use. .9 8.8 8m .832

 

 

 

 

 

 

 

 

«9 a a9 «8 SQ 82 8a 82 Em 295m
:3 o8.o 83 Sod 82. 8o... 89° 89° 3858 933 . "as?
33 89° «:3 8o... _8.c 83. 23 c8... .2959. 8» as 582
as... - a; - «a .c - 83 - 8250820 223.8:
2.»... ~86 - «no... _8.c 8330220 .9633

”85.9%:an 3&3.» 33...»

8» 8» 8» 8h 8» 8» 8» 8» 8:556 933 . “flu—Er
8» 8» 8» 8» 8» 8» 8» 8» £8.56 :8» can 552
8» 8» 8» 8» 8» 8» 8» 8» 8:55.. 33:85 882:8:

98.3 68.3 98.9 98.9 -

38: 99.9 82”- 25.. - 32 2338.. one“:
98.: 92.9 93.9 92.9 -

and- v8.9 and 8nd - 82 .8 3628 “a 38>
98.9 98.9 - 98.9 98.9 -

on; 899 :85. 89c - 85.: 32 2288.. .o o?
93. 9 A8“. 9 - 98.9 83.9 -

E. 7 and «:6 an; - 32 2388.. S ou<
98.9 88.9 88.9 88.9 -

99°. 99° 89¢. n89 - 8.3.: is .... .22
92.9 93.9 99.9 9:9 -

”8... 96.9 3.3 x: .o - 2a. 3 8.3.
99 .9 89.99 - - - -

«3.9 23. - - - 2238.. .o .8:

- - 83.9 98.9 -
- .E .9. :23: - sense: 5 .55. oz
- 39.9 88.9 -
”8.”- :3: - 8,88: 9:23 58
98.9 939 -
83 8a.? . 9288:. 55“:
8.9.9 98.9 .
8N. _- 33..- - 8388 .282
92.9 95.9 -
- - £3 98.: - 82 .o 220

88.9 88.9 68.9 88.9 -

93.? 32v .583. :63? - 853 a?
89.9 95.99 - 2 2.: 93.: -

83 32.. - :93... ...:om.n - ou<
.m: 3o .m: 30 .m: 30 .m: 30 3g; 9%
394 333 glass; 3'3 8.93 9.3 ll 593

03.: =5". 3.8: :5:
$2 933. «I. 56:6

 

 

 

 

«:96 8a 8865qu 5:82 323 8.288: go 0385 .8335 use: unfit. 033.

APPENDIX 6

.22. 8:82 .8 5.. 8305 83.2? .9393; .9 9.3 23. 8m .5 a .... Ea new a .. ”99 a 8505:»: BEE...

8:835. ... .38. .3239: 05 a ”5.8320 .8 38:8 anon 26: .8398 SO 2: .5... 65 885:88 5 Ba 89:» Emu—am .882

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

nwm One man one mmfi 8w nnfl 8w “Em 295m
2nd mead wood wwod 3nd 96.». a2 .o wood 233 d 3.8 58v 222
vfl .o Sad «.36 Sad mood 31¢ owed ”and 233 Q 68 58c 0.95:
"3:335:59 EST...» 397m
Fmd 3nd fled 8nd «$6 «and wand #36 N:
99.9 93% 98.9 49.99 95.9 88.9 98.9 68.9
89.6 Sad. m9; 2an onfim- 38:6. Etc. 509 9:58 NT». 9.83 58». 032
23.8 Sand 83$ 23.8 88.8 Sand 9.328 ASN.8
mung- 936. com.“ owed Zed 3.2. .58.? .....imnnd. 2:58 «To 3.8 58: 032
985 63.8 83.9 93.8 wad 33$ 9900 88.8
.98.». 2%.: :56 99:9 «an..- 399 w~n9 as; 882: «To 2:5. 58v 0.980.:
936.9 9%.: Goad 89.9 88.8 83.0 Awmfimv 9%.:
55m an”. T 936 «and $36 «:6 :36. «who 2:58 «To 33 58: ...—«So:
.m.m 30 .md 30 .mm 30 .mm 30 838.299» 9.93.9393
8:3 .933 9 2:3 .233 9 8:3 .23: 9 2:3 .233 V
88: 8.9: 83:
3-8 :02 omém 5:53

 

 

 

3:90 89 8865qu 5:82 o>3< 8.288: as 295m

.38: 8 83: was: 93502 2288: .33 t .895 "2 .2 229

e Nae/Em:

117

118

.2 .e< 038,—. ..8 88: new .382

 

 

 

 

 

 

 

 

 

 

 

 

 

8v nmm one nmm 2n awn an own 08m 2986
«mod 89¢ :hd Nmmd mwod 55.0 “the omwd 233 3 man 586 232
$96 awed wand «2 .o sod NS .o ~36 086 283 a. 3.8 536 0.880%
”3:85:88. 3.53% 337$
Svd wand and X36 «and find End 286 NM
@349 88.9 83$ Cavév 63.8 38.9 33.8 9.3.3
n84 £09 956 v8.0 vwnd ~34..- 0N5? eavd. «5.88 «Ta 288 588 032
63.8 9.869 Good @929 9.3.0 A308 ASN.9 83.0
~34- an; Sad wag- mnmd. nnmd. 5nd- owed- 3808 «To 5.3 536 032
@349 88.3 8949 8503 CE .9 Awow$ 29 .9 A2: .vv
mad «3.7 21v wwwé. Em?” «35w- 13.x- Q _ .T 3:88 «To 888 58v 0880"“
92.9 92.9 23.9 9.8.9 92.9 9: .9 99.9 :39
Red 3.}. mad 8a.? .1399- wand 289 .w- wom.~ 2:88 NTo :8 58v 2.80m
.mm 30 .mm 30 .mm 30 .mm— 30 “£3.53 923353
.233 8:3. 9 .233 .933 V .233 2:3 9 8:3 2:3 9.
348: 88”— omzo: scam
2-2 52 2-2 5835

 

 

 

3880 ha 0855:3me .8622 o>on< 83:83: .8 298%

.28: 8 83: mean “one”: 2238a .33 8o 88.5 E .3 29¢

119

.23 2.3 5. 8.8 8m 532

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

vow mwv v8 mwv «mg 2: NR: 2.: 85 295m
Rod 5:. 3nd «Ed awvd 81c «Nod :66 838 a. man 58v 222
Ste «26 32. find on" .o aomd She Sod 288 a. 33 5%.. 295m
..uuauuwbgflm EBTQ 337m
«86 mmmd envd ~36 and 33 .o mmmd mafia NM
65$ 83$ 38.0 fined 88.3 92.8 a 3.6 33v
Sud. 3N? v56 mnvd. N2 6 won; .momé Sud 2:58 «To 288 58c 032
83.6 End $3.9 88$ 83.5 38.8 62 .NV 88. 3
mid. $86. «36. SN; mend «a: 93.7 3036- 9.82: «To “ma 53v 332
Avmmév 89.3 A2 08 fined AS adv Gan. c 82 .8 28.8
find. 8&6 3b.» emmd wand onud v2.7 9336 £52: NTo 233 58a canon
63 .9 63.8 $2 i 83.9 A89: $3.: 68. s 8%. 3
3nd- ~36 3a.». vmvd mad. anmé. mom. T 9: .o 2.58 ~70 3.3 58v 225m
.m.m 30 mm 30 .m.m 30 .md 30 §§I§ b.3353
.233 2:3 V .23: 933 V “(:3 2:3 V .223 .233 V
8.5: Eula..— 8.5:
3.2». 33v< 2-5 c.8220

 

 

 

.3an “on ohzficoaxm 5:32 32?. m2oaomaom .8 29:3

5.5: 8 83: mean “3.82 22.38: .33. Mo 5% a. .3 3.5

120

.fi—

3 ...... 65 run a 3. ”$2 a 8:85.36 .8658” 8:865 . .65: .3va5 06 6 555—0 Bu 68858 53 2:5 .8353 30
06 58 65 8868.3: 5 93 52:0 6.365%

.2 6< 035. 5 6888...: 5.55 06 59a 6355 855.8 06 «Bonn 038 «E... .582

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

 

man One mam one mnfi 8» $3 an ofim 2558
wood coed coed coed «Nod coed $56 coed 85:56 263 ... ems—5
685 896 Bod coed 89o «wad 036 coed 8:556 5n» 65 6:22
need - wand - Sod - god - 83:88.20 2288mm
8nd - Svd - 3nd - End - 8352856 56363

..uucuuKEwE 55.5.55. 55¢

8» 8» mom 8» no» mo» 8% 8» 8556 263 .. $2.5

no» 8» 8» 8a 8» we» 8» 8» 35.56 5....» 65 55.2

8» we» 8.» 8» no.» no» 8» 8» 85:56 33585 28833

83.8 - Sand - 89.8 - Gang -

86.6 - 2nd - god: - Fad. - 68: 22.83: 2.50..“
8.8.8 - 28.8 - $8.8 - 93.8 -

«a: .o. - mavd - vnmd - 686. - 68: no 558:8 6 58>
88.8 - $8.8 - 68.8 - $8.8 -

Sod - v8.6 - cad. - «8.? - 85$ 68.. 228:8 ho a?
35.8 - 83.8 - Sam .8 - 9% .8 -

63 .c. - End. - ....SNA - mend - 68: 228:3 we om<
88.8 - $8.8 - 98.8 - 68.8 -

mad - wad - mood - Sod. - 623$ 65— .8 86.x
$59 - 63.8 - age - 98.8 -
13d. - .686. - 8nd. - n26 - 6:2 .8 v.23:
38.0 - 98.8 - €56 - $8.8 -
.636 - ENS- - $6.2- - mm: - 22.852. .6 63m
88.8 - $8.8 - $8.8 - 28.8 -

«8.? - zed - mad. - L3? - B53 a?
38.8 - €88 - Amend - 83.8 -

mead - Sod. - 8w; - nVN.N - om<

.mfl 30 .mm 30 .md 30 .m.m 30 35.56» 55535.5
5:3 5:3 V 5:3 5:3 V 5:3 5:3 V 5:9: 5:3 V

83: an 83: 5.8m
3.8 5.2 onén 55o?

 

 

 

8:60 “on 23655.".— 5602 o>on< mEOsomsom ho 2.58m .5555 58m 5N6< 22¢.

121

8.3. 03.... 3.. 88 8m .832

 

 

 

 

 

 

 

 

 

 

 

8e n8 8: m an a an n can a
83 8a... 8.3 8?: 28.: Bad «8... 89¢ 355% 2.3 _. of?
83 89° 29.9 8o... 28.9 89¢ one... Sod 853 .8» Ba 58:
89c - «8.: - $2. . 82 .9 - 85305.20 22.88:
:8... 53 - :3 n5... ”35285.0 .335:

323995.... 2.3.5 5 32.:

8» 8V 8» 86 8» 8» 8» 8» 85:56 263 .. ou.:=_>
8» 8» 8» 8» 8» 8» 8» 8» 85:56 8» 65 6:02
8» 8» 8» 8» 8» 8» 8» 8» 85:56 302585 555:8:—

293: 2839 222.9 - 88.9
35.? «:3 28.2- 98.9 33 2388.. 2.80".
88.9 2.5.9 82".: 89.9
95.? 82 .? :83- at? 32. 9o 958:3 .o as»
$8.9 68.9 88.9 88.9
as... 83. m8? 2?? 2.25.: 32 29.3.8: 9.. om<
33.9 88.9 38.9 33.9
83. 8n.“ 2.3 «86 32. 22392.... om<
2.8.9 68.9 2 8.9 88.9
83 28.? «86 2?? 2.233 E: 3 .22..
38.9 2.89 89.9 2.3.9
.5 .? «and ”5.? 32. E: .6 85..
929 2.5.9 5.39 88.9

«8.? a; .985 .83.— 2288: 5 .852 oz

28.29 25.3: 839 89.9
83.- 8a.? .3998- R? 2.388.. €25 58
85.29 88.”: 23.9 82.9

3.2 a; «8.2 was? 2.8882. .25".

55: 25.5 85.9 22.9

on?“ $3- .82 .2 $3 2.888.. .332
2839 93.29 2899 92.9
8a.: N32 ~98 368.: 32. 9o .56
98.9 68.9 83.9 28.9
:80? 22.? EN? 83 2.233 0?.
28.9 25.9 25.9 89.9
...§.- - on: 9.3 32.? - 02
.mm 80 .mm 80 .mm 80 .mm 80 33.5: 3355:
5:3 5:3V $.3Il5:3V 5:3 5:3V §I3V
8:0: 5.3m 825: g"—
25 52 2-5 3885

 

 

 

 

 

 

:96 i 235.25 .382 222 8.288: 3 2:58 .3835 55: ”£3 05¢

122

83. 03.... a. 8.8 8m .632

 

 

 

 

 

 

 

 

 

 

«8 as. .8 a: an: 2.: «a: 2.: 3a 295m
82. :85 83. o8.o 89¢ 83. Sod 82. £55.... 253 . 05$
8o... 83. and 83. 83 ES 23 o8... 85.5. :2 =5 582
«85 - and - «a .o - ES - 83.9050 2238:
a; 93 23 v85 835853 .3355
..uxaugcu: 3.3.x 3» 32¢
8» 8» 8» 8» 8» 8» 8» 8» 85556 963 o§>
8» 8» 8» 8» 8» 8» 8» 8» 82.5.6 .36» 55 5:02
8» 8» 8» 8» 8» 8» 8» 8» 85.52. 8385 832:8.—
aa .3 - 95.3 - 83.8 5nd -
5.. «Non 3." Sad 32 22.38.. 223m
2%.: 826 $929 fine
53 c3... .83. ..3n.._ 32 ..o "5.85. .o as»
68.9 28.9 68.9 G88
22.. as... 88.9 .23. B53 32 2238.. .8 a?
33.: 98.: $3.9 $9.8
an: 83. 82. .53.? 82. 2238.. .o a?
98.9 $8.9 $8.8 $8.8
n86 .08.? n8? 8...? 85.3 Es .0 8.3
83.9 88.9 a: .9 $2 .8
Rod .23 Ed 3&3 Ba. 3 8.3.
83.5.: 88.»: - -
831 n2 .2 - - 2288.. .o as:
- - 68d std
- - and $3”. 2235.. 5 528 oz
- - ch .3 88.8
- - 83 an. _- 3.88.” 2:23 58
- - 899 580.3
- - 21¢ 83- 838.. §=fl
- - :39 83.0
- - $3” an... 838“. .252
- - 63.9 85.9
- - 2.9m- 599816- 5: .«0 EEO
$8.9 $8.9 68.9 88.8
«8° v3.9 53. :5”? 323 a?
83.8 63°: $8.: 5a. a
a: - an; «8.? .258 of
.mm 8o .mm 30 .mm 30 .mm 30 §§§ 33:33
§ 33 v $34. $3 a g 53 n 33 So: V
881': glam 8.5: Inclalm.
$8 333 3-» 5320

 

 

 

536 E aaiaxm .58: 26.2 3.288m .o 2.53 .8835 can 3.3 2.3.

LIST OF REFERENCES

LIST OF REFERENCES

Ainsworth, Martha, Susmita Ghosh and Innocent Semali. 1995. “The Impact of Adult
Deaths on Household Composition in Kagera Region, Tanzania.” draft, Policy Research
Department, World Bank, Washington, DC.

Ainsworth, Martha and Mead Over. 1994a. “AIDS and African Development.” World
Bank Research Observer 9(2): 203-240.

Ainsworth, Martha and Mead Over. 1994b. “The Economic Impact of AIDS in Africa.”
in M. Essex, S. Mboup, P. Kanki, M. Kalengayi, eds., AHDS in Africa. New York:
Raven Press.

Ainsworth, Martha, Godlike Koda, George Lwihula, Phare Mujinja, Mead Over and
Innocent Semali. 1992. “Measuring the Impact of fatal Adult Illness in Sub-Saharan
Africa.” Living Standards Measurement Study no. 90, World Bank, Washington, DC.

Alderman, Harold and Christina Paxson. 1992. “Do the Poor Insure? A Synthesis of the
Literature on Risk and Consumption in Developing Countries.” WPS 1008, Agriculture
and Rural Development Department, World Bank, Washington, DC.

Barnett, Tony. 1994. The Efi’ects of HIV/ADS on Farming Systems and Rural Livelihoods
in Uganda, Tanzania and Zambia. Final report to Food and Agricultural Organization.
Norwich, UK: Overseas Development Group, University of East Anglia.

Barnett, Tony and Piers Blaikie. 1992. AIDS in Africa: Its Present and Future Impact.
New York: Guilford Press.

Barnett, Tony, Joseph Tumushabe, Grace Bantebya, Ruth Ssebuliba, Juma Ngasongwa,
David Kapinga, Monica N delike, Michael Drinkwater, Godfrey Mitti and Martina
Haslwimmer. 1995. “Final Report: The Social and Economic Impact of HIV/AIDS on
Farming Systems and Livelihoods in Rural Africa: Some Experience and Lessons from
Uganda, Tanzania and Zambia.” Journal of International Development 7(1): 163-176.

Becker, Gary. 1965. “A Theory of Time Allocation.” Economic Journal 75: 496-517.

Bloom, David and Ajay Mahal. 1997. “Does the AIDS epidemic threaten economic
growth?” Journal of Econometrics 77(1): 105-124.

Bongaarts, John. 1996. “Global Trends in AIDS Mortality.” Population and
Development Review 22(1): 21-45.

123

124

Brown, Lynn, Patrick Webb and Lawrence Haddad. 1994. “The Role of Labour in
Household Food Security: Implications of AIDS in Africa.” Food Policy 19(6): 568-573.

Chamberlain, Gary. 1980. “Analysis of Covariance with Qualitative Data.” Review of
Economic Studies 47(1): 225-238.

Cleave, John. 1974. Afi'ican farmers: Labor Use in the Development of Smallholder
Agriculture. New York: Praeger Publishers.

Cleveland, William. 1979. “Robust Locally Weighted Regression and Smoothing
Scatterplots. ” Journal of American Statistical Association 74(368): 829-36.

Collier, Paul, Samuel Radwan and Samuel Wangwe. 1990. Labour and Poverty in Rural
Tanzania. Oxford: Clarendon Press.

Cuddington, John. 1993a. “Modeling the Macroeconomic Effects of AIDS, with an
Application to Tanzania.” World Bank Economic Review 7(2): 173-189.

Cuddington, John. 1993b. “Further Results on the Macroeconomics Effects of AIDS: The
Dualistic Labor Surplus Economy.” World Bank Economic Review 7(3): 403417.

Cuddington, John and John Hancock. 1994. “Assessing the Impact of AIDS on the
Growth Path of the Malawian Economy.” Journal of Economic Development 43(2): 363-
368.

Deaton, Angus. 1997 . The Analysis of Household Surveys: Microeconometn'c Analysis for
Development Policy. Washington, D.C.: Johns Hopkins University Press.

Dercon, Stefan. 1996. “Risk, Crop Choice, and Savings: Evidence from Tanzania.”
Economic Development and Cultural Change 44(3): 485-513.

Evenson, Robert, Barry Popkin, and Elizabeth Quizon. 1980. “Nutrition, Work, and
Demographic Behaviour in Rural Philippine Households.” in H. Binswanger, R.E.
Evenson, C.A. Florencio and EN. White, eds., Rural Households in Asia. Singapore:
Singapore University Press.

Fafchamps, Marcel. 1993. "Sequential Labor Market Decisions Under Uncertainly: An
Estimable Household Model of West African Farmers. " Econometrica 16(5): 1173-1197.

Gillespie, Stuart. 1989a. “Potential Impact of AIDS on farming systems.” Land Use
Policy 4: 301-312.

Gillespie, Stuart. 1989b. “The Potential Impact of AIDS on Food Production Systems in
Central Africa.” Food and Agriculture Organization, Rome, Italy.

125

Glomm, Gerhard and Michael Palumbo. 1993. “Optimal Intertemporal Consumption
Decisions Under the Threat of Starvation.” Journal of Development Economics 42(2):
271-291.

Greene, William. 1990. Econometric Analysis. New York: Macmillan Publishing
Company.

Grossman, Michael. 1972. “On the Concept of Human Capital and the Demand for
Health.” Journal of Political Economy 80(2): 223-255.

Jacoby, Hanan. 1993. “Shadow Wages and Peasant Family Labor Supply: An
Econometric Application to the Peruvian Sierra.” Review of Economic Studies 60(4): 903-
921.

Jacoby, Hanan and Emmanuel Skoufias. 1995. “Risk, Financial Markets, and Human
Capital in a Developing Country.” draft, Department of Economics, University of
Rochester.

Juster, P. Thomas and Frank Stafford. 1991. “The Allocation of Time: Empirical
Findings, Behavioral Models, and Problems of Measurement.” Journal of Economic
Literature 29(2): 471-522.

Kambou, Gerard, Shantayanan Devarajan, and Mead Over. 1992. “The Economic Impact
of AIDS in an African Country with a Computable General Equilibrium Model of
Cameroon.” Journal of African Economies 1(1): 109-130.

‘ Khandker, Shahidur. 1988. “Determinants of Women's Time Allocation in Rural
Bangladesh.” Economic Development and Cultural Change 37(1): 111-126.

Killewo, J aphet, Klinton Nyamuryekunge, Anita Sandstrom, Ulla Bredberg—Raden, Stig
Wall, Fred Mhalu and Gunnel Biberfeld. 1990. “Prevalence of HIV-1 Infection in the
Kagera Region of Tanzania: a Population-based Study.” AIDS 4(11): 1081-1085.

Killewo, Japhet, L. Dahlgren and A. Sandstrom. 1994. “Socio-Geographical Patterns in
HIV-1 Transmission in Kagera Region, Tanzania.” Social Science and Medicine 38(10):
129-134.

Kochar, Anjini. 1995. “Explaining Household Vulnerability to Idiosyncratic Income
Shocks.” American Economic Review Papers and Proceedings 85(2): 159-164.

Kochar, Anjini. 1996. “Smoothing Consumption by Smoothing Income: Hours-of-Work
Responses to Idiosyncratic Agricultural Shocks in Rural India.” draft, Department of
Economics, Stanford University.

126

Konde-Lule, J., R. Ssengonzir, R. McNamara, C. Li, M. Wawer, J. Edmundson, R.
Kelly 1995. “HIV-Related Mortality and Household Economic Burden, Rakai, Uganda.”
IX International Conference on AIDS and STD in Africa, Kampala, Uganda.

Mukoyogo, Christian and Glen Williams. 1991. AIDS Orphans: A Community
Perspective From Tanzania. London: ACTION AID.

Mueller, Eva. 1984. “The Value and Allocation of Time in Rural Botswana.” Journal of
Development Economics 15(1-3): 329-360.

National Research Council. 1996. Preventing and Mitigating AIDS in Sub-Saharan
Africa. Washington, D.C.: National Academy Press.

Newman, John and Paul Gertler. 1994. “Family Productivity, Labor Supply and Welfare
in a Low Income Country.” Journal of Human Resources 29(4):989-1027.

O’Dell, Michael, Helen Hubert, Deborah Lubeck, and Peter O’Driscoll. 1996. “Physical
Disability in a Cohort of Persons with AIDS: Data from the AIDS Time-Oriented Health
Outcome Study.” AIDS 10(6): 667—673.

Over, Mead. 1992. “The Macroeconomic Impact of AIDS in Sub-Saharan Africa.”
African Technical Department, Technical Working Paper No. 3 , World Bank,
Washington, DC.

Over, Mead and Daniel Dorsainvil. 1996. “The Impact of Adult Mortality on Income,
Consumption and Assets, ” draft, Policy Research Department, World Bank, Washington,
DC.

Over, Mead and Peter Piot. 1993. “HIV Infection and Sexually Transmitted Diseases.” in
D. Jamison et al. Disease Control Priorities in Developing Countries. New York: Oxford
University Press.

- Philipson, Tomas and Richard Posner. 1995. “The Microeconomics of the AIDS
epidemic in Africa.” Population and Development Review 21(4): 835-848.

Pitt, Mark, Mark Rosenzweig and MD. Nazmul Hassan. 1990. “Productivity, Health,
and Inequality in the Intrahousehold Distribution of Food in Low-Income Countries.”
American Economic Review 80(5): 1139-1156.

Pitt, Mark and Mark Rosenzweig. 1986. “Agricultural Prices, Food Consumption and the
Health and Productivity of Farmers.” in Inderjit Singh, Lynn Squire and John Strauss,
eds. , Agricultural Household Models: Extensions and Applications. Baltimore: Johns
Hopkins Press.

127

Pitt, Mark and Mark Rosenzweig. 1990. “Estimating the Intrahousehold Incidence of
Illness: Child Health and Gender-Inequality in the Allocation of Time.” International
Economic Review 31(4): 969-989.

Quisumbing, Agnes. 1996. “Modeling Household Behavior in Developing Countries:
Discussion.” American Journal of Agricultural Economics 78(5): 1346-1348.

Reardon, Thomas. 1997 . “Using Evidence of Household Income Diversification to
Inform Study of the Rural Nonfarm Labor Market in Africa.” World Development 25(5):
73 5-747. ,

Rose, Elaina. 1995. “Ex Ante and Ex Post Labor Supply Responses to Risk in a Low-
Income Area.” draft, Department of Economics, University of Washington.

Rosenzweig, Mark. 1980. “Neoclassical Theory and the Optimizing Peasant: An
Econometric Analysis of Market Family Labor Supply in a Developing Country.”
Quarterly Journal of Economics 94(1): 31-55.

Rosenzweig, Mark. 1988. “Risk, Implicit Contracts and the Family in Rural Areas of
Low-Income Countries.” Economic Journal 98(4): 1148-1170.

Rosenzweig, Mark and Kenneth Wolpin. 1985. “Specific Experience, Household
Structure and Intergenerational Transfers: Farm Family Land and Labor Arrangements in
Developing Countries.” Quarterly Journal of Economics 100(5): Supplement, 961—987.

Ryder, Robert and Roy Mugerwa. 1994. “The Clinical Definition and Diagnosis of AIDS
in African Adults.” in M. Essex, S. Mboup, P. Kanki, M. Kalengayi, eds., AES in
Africa. New York: Raven Press.

Sen, Amartya. 1966. “Peasants and Dualism with and without Surplus Labor.” Journal
of Political Economy 74: 425-450.

Singh, Inderjit, Lynn Squire and John Strauss. 1986. Agricultural Household Models:
Extensions and Applications. Baltimore: Johns Hopkins Press.

Skoufias, Emmanuel. 1993. “Labor Market Opportunities and Intrafamily Time
Allocation in Rural Households in South Asia.” Journal of Development Economics
40(2): 277-310.

Strauss, John and Duncan Thomas. 1995. “Human Resources: Empirical Modeling of
Household and Family Decisions.” in H. Chenery and TN. Srinivasan, eds.,
Handbook of Development Economics, Volume 3. Amsterdam: North Holland.

Tibaijuka, Anna. 1984. An Economic Analysis of Smallholder Banana-Cofiee Farms in
the Kagera Region, Tanzania: Causes of the Decline in Productivity and Strategies for

128

Revitalization. Published Ph.D. dissertation, Swedish University of Agricultural
Sciences, Department of Economics and Statistics, Uppsala.

Tibaijuka, Anna. 1997 . “AIDS and Economic Welfare in Peasant Agriculture: Case
Studies from Kagabiro Village, Kagera Region, Tanzania.” World Development 25(6):
963-975.

Townsend, Robert. 1994. “Risk and Insurance in Village India.” Econometrica 62(3):
539-591.

Wooldridge, Jeffrey. 1997 . “Asymptotic Properties of Weighted M-Estimators for
Variable Probability Samples.” Econometrics and Economic Theory Paper N o. 9604,
Department of Economics, Michigan State University.

World Bank. 1992. Tanzania AIDS Assessment and Planning Study. World Bank,
Washington, DC.

World Bank. 1993. “Report of a Workshop on: the Economic Impact of Fatal Illness in
Sub-Saharan Africa.” World Bank, Washington, DC. and the University of Dar es
Salaam.

World Health Organization. 1995. Unpublished data compiled by the Global Programme
on AIDS and cited in John Bongaarts. 1996. “Global Trends in AIDS Mortality.”
Population and Development Review 22(1): 21-45.

Chapter 2
EXAMINING REPORTING ERRORS IN RETROSPECTIVE
MORTALITY DATA
I. Introduction

Empirical analysis of many aspects of household and individual behavior
typically relies on retrospective data as reported by individuals. Even with the
development of panel surveys to collect household survey data, at least in the initial
interview, retrospective information will be counted on for identification of past events.
Researchers field surveys and utilize recall data pertaining to a wide range of topics
from retrospective life histories (e. g. marriages and births) to consumption expenditures
in the recent past.

The quality of retrospective data is, however, often difficult to discern. Some
experimental studies have been designed and implemented to explore the extent to
which different types of recall errors plague retrospective data. These sorts of analysis
can hopefully improve methods of collecting retrospective data and minimizing recall
errors. For example, Scott and Amenuvegbe (1990) examine how varying duration of
recall periods is associated with errors in recalling expenditures. Loftus and Marburger
(1983) assess the improvement in accuracy of retrospective accounts when surveys use
a highly salient landmark event (the eruption of Mt. St. Helens) to mark the beginning
of the reference period. Van Der Vaart et al. (1995) explore possible reductions in

recall bias when surveys provide a checklist to respondents.
129

130
Because we anticipate that the nature of recall bias will vary with the type of

retrospective data collected, it is difficult to make generalized statements about
appropriate survey techniques, let alone accuracy. Unfortunately, researchers are often
limited in their ability to assess the quality of retrospective data in a specific data set.
Auxiliary sources, such as marriage or birth certificates or receipts for expenditures,
are usually not available to validate retrospective data. For some types of recall error,
researchers can look for indicators of quality within a data set. For example,
pronounced spiking or heaping of variables such as age, job tenure or duration of
unemployment may suggest rounding on the part of respondents. Even in this case,
though, without further information it is not possible to distinguish rounding from a
true heaping process. When available, panel data can present an alternative means of
assessing the quality of retrospective data by comparison of retrospective reporting in
reinterviews with that in the original interview when recall periods overlap.

Using the Kagera Health and Development Survey (KHDS), this paper attempts
to assess the quality of the reporting of deaths of household members in two ways.
First, the pattern of incidence of deaths over the 2—year recall period in the first
household interview is examined. Second, the data collected in the first full interview
of the KHDS are compared with information reported in the enumeration (baseline)
survey that preceded it. Although the enumeration survey does not use the same survey
instrument as the full KHDS questionnaire, basic event reporting can be compared

across SUIVCYS.

131
This analysis is useful for several reasons. In general, we expect the reporting

of death, a major event, to be less prone to event recall as compared to retrospective
data on minor events such as incidental expenditures. The degree to which deaths are
inconsistently reported may underscore the need to pursue longitudinal studies that
collect data over shorter retrospective periods and later can be used to piece together
more reliable event histories.

Within the more narrow context of this specific data set, this analysis may be
insightful to researchers anticipating utilizing the KHDS with its extensive and rich
data. The expressed objective of this survey was to examine the impact of deaths on
households and communities; thus, the quality of this component of the KHDS is
especially important. As a consequence of the following analysis, data for the
retrospective period reported in the first household interview is truncated when used in
Chapter 1 analysis.

This paper is organized as follows. Section II discusses the various aspects of
recall error. Section III describes the KHDS data and the degree to which the
enumeration survey can be compared with the first interview of the panel. The
incidence of deaths reported in the KHDS is presented in Section IV. Section V covers
the comparison across the enumeration survey and the first KHDS full interview.

Section VI concludes.

II. Recall Error

132
There is a large body of work in survey literature about reporting errors in

survey data; Sudman and Bradbum (1974) summarize the extensive work in this area
from the 19508 and 1960s. Retrospective data, in particular, introduces various data
quality issues linked to recall errors. Over time, respondents may forget that some
event occurred, and thus underreport events. Remembrance of events will depend not
only on the duration of the recall period, but also on the nature of event being reported.
For example, with respect to expenditures and income, Deaton (1997) notes that a
longer recall period may be plagued with greater measurement error because
respondents have difficulty recalling expenditure for minor items several months later.
This suggests that estimates of these variables, expenditures or income, fall as the recall
period is lengthened. It also suggests that less meaningful events are more likely to be
forgotten. That is, important life events may be associated with improved recall. As
well, events linked to other events may improve reporting accuracy, particularly if the
other event is a salient one. For example, migrations tied to marriage may be more
accurately reported than other, less significant moves. Smith and Thomas (1997)
explore these issues and others relating to recall error in reporting migration histories in
the first and second Malaysian Family Life Surveys conducted 12 years apart. They
find that salient moves linked with meaningful events or moves that last a long time are
more accurately reported.

Aside from errors in the recollection of an event, the quality of retrospective
data depends on the remembrance of the details of the event. Recall bias will, no

doubt, depend on the nature of the event being recalled. Blane (1996) asserts that

133
details of emotional laden events are more likely to be recalled inaccurately while

everyday activities are recalled more accurately.1

A pivotal detail of retrospective reporting is the dating of the event. Forward
telescoping, for example, is the reporting of an event more recently that it actually
occurred, thus downward biasing time from the survey to the event. Forward
telesc0ping may even be deliberate on the part of the respondent. Shorter recall periods
may be subject to a “ ‘ boundary’ or ‘ start-up’ problem, whereby respondents report
items just before the reporting period, in an effort to be helpful and to ensure that the
enumerator does not miss events that he or she would ‘ surely’ want to know about
(Deaton, 1997).” landmark events that clearly mark the beginning of the reference
period may reduce the incidence of unintentional forward telescoping (Loftus and
Marburger, 1983). Along this same line, Blane (1996) discusses the concept of a “Life
Grid” whereby all dates of any changes in areas of interest are cross-referenced. These
events can include events external to the households such as wars and earthquakes.

Heaping is one possible type of recall error in which we observe an unusual
concentration of responses at certain durations with respect to prior expectations about
the frequency distribution.2 The problem of heaping has a range of applications from
reporting of duration of breastfeeding (Klerman, 1993) to unemployment duration
(Torelli and Trivellato, 1993). Ureta (1992) finds pronounced spiking at five-year

intervals in reported job tenure from several Current Population Surveys; furthermore,

 

‘ Blane (1996) goes on to describe these everyday events as the hum-drum details of life that include
“barely noticed background routines of life” which would seemingly make them less accurately recalled.
2 Evidence that KHDS households may approximate dates can be drawn from the sample of children
under 60 months of age (not reported here); the distribution of children by reported age displayed very
distinct heaping at 6-month intervals.

134
she finds that the probability of rounding increases with tenure. Respondents may heap

as an approximation of duration, such a rounding to the nearest year. Researchers have
implemented methods to smooth data as a correction for reported tenure (see, for
example, Ureta (1992) and Diebold et a1. (1997) with respect to tenure distributions in
the Current Population Surveys). Such corrections are typically based on the
presumption that the heaping is an artifact of rounding by respondents which may not
be the case. Klerman (1993) tests the presumption that heaping moves answers to the
closest heaping point (e.g. year or half-year) with respect to breastfeeding durations in
Malaysia and finds that in most cases the heaped data is not pulled from adjacent
responses.

Errors in recalling the date or other characteristics of events may be noisy but
not necessarily associated with a pattern like telescoping or heaping. With respect to
birth histories, women may recall all births but year-of-birth may be reported with
noise, particular in settings where women have a large number of births, and low levels

of schooling/literacy.

III. Data

The data for this study are drawn from a research project conducted by the World
Bank and the University of Dar es Salaam in the region of Kagera in Northwest
Tanzania.3 The explicit objectives of the project were to measure the economic impact of

prime—age (15-50) adult fatal illness in the region and to propose cost-effective strategies

 

3 For further description of the project and data see Ainsworth et al. (1992) and World Bank (1993).

135
to help survivors. Kagera is estimated to be one of the regions in Tanzania most affected

by the HIV/AIDS epidemic (World Bank, 1992). The issue of the quality of the
retrospective data about incidence and characteristics of deaths in the KHDS is
important because the objectives of the survey rely not only on the reporting of deaths,
but also on the characteristics of the event (i.e. date and classification of deaths with
respect to household membership and age).

Begun in 1991, the KHDS is a panel survey that consists of four lengthy
interviews, referred to as “waves” , approximately 6—7 months apart on average. The
KHDS included 838 households in the first wave; 757 households complete all four
waves. In the first wave of the KHDS, households were asked to report deaths in the
24-month period preceding the interview. In the following waves, the reference period
is the last 6 months (or the time since the last interview). Information on deaths was
fielded in two survey modules. The first (section 20a) is designed for the deaths of
members of the household (not necessarily related to the respondent) and the second
(section 20b) captures the deaths of any non-household member (i.e. non-coresident)
relatives. The deaths reported in these sections of the household survey should be
recorded only once. Thus, there is no overlap in the recall periods of the four waves of
data and sections 20a and 20b are mutually exclusive.

The initial 838 households were selected on the basis of a baseline survey
(referred to as the enumeration survey) fielded to all households in the 51 survey
villages less than a year before the start of the KHDS panel. The enumeration survey

asks about presence of adults in the household who are too sick to work and about

136
deaths of household members in the last 12 months. Thus, the retrospective period

covered in the enumeration survey lies within the 24-month recall period of the first
wave.4 The consistency of the reporting of household member deaths can therefore be
examined by linking the enumeration data with that of the first wave. While very
similar, the survey instruments are not, however, the same. Inconsistencies across the
two surveys can not necessarily be attributed to recall bias as opposed to the differences

in the survey instruments.

IV. Incidence of Death in Kagera Households

A priori, in any random sample of households we expect to observe a (fairly)
smooth distribution in the percentage of households reporting a prime-age death over
each month in a 24 month period. An exception to this uniform distribution would be
the occurrence of deaths due to a contagion (in the community or the household), in
which case the proportion of households reporting death may increase as dates
approach the interview date. Another exception would be in the event of seasonal
patterns in the incidence of deaths; certain months may be associated with higher
mortality. This would be the case if, for example, certain seasons are related to higher
morbidity or disease rates for environmental or economic reasons. Examination of the

timing of reported deaths confirms that death rates, in general, are highest in the

 

4 For households in 2 villages (out of 51), the start date of the wave I recall period follows that of the

recall period for the enumeration. In one of these cases, the difference is only a few days in the same
month. In the other village (with 17 households), the wave I recall period starts 2 months after the start
of the enumeration recall period. For other households in which the entire enumeration recall period lies
within the wave I recall period, the length between the start of the wave I recall and the start of the
enumeration survey recall varies by household.

137
months of April, August and July. July and August are the driest months of the year --

part of the “kiangazi” season. The month of April in the “masika” rainy season is one
of the rainiest months of the year. However, this is not likely to be an issue in the
KHDS where the sample of households interviewed span a seven-month period.

If we examine the reporting of prime-age adult household, member deaths among
the sample of all 837 households interviewed in the first wave of the KHDS, we notice
two important features consistent with the sampling strategy employed (Figure 1).5
Despite the seemingly low levels of adult mortality in Figure 1, the levels of mortality
in this sample of households are actually much higher than would be predicted for the
average household from the 1988 census. Using the census data, Over (1995) estimates
the risk of an adult death in the average Kagera households to be about 1.88 percent per
year. By his calculations, we observe more than four times the expected number of
households experiencing at least 1 prime-age death. This outcome is directly attributed
to the stratified choice based sample employed which was based on a short enumeration
survey collected, on average, 8-9 months earlier.

The second feature of Figure 1 is the increase in the proportion of households
reporting a death after about 7-8 months in the past. This is also consistent with the
sampling strategy employed. The objective of the sampling procedure was to increase

the project’s chances of observing serious illness or death in the sampled households.

 

5 The xth month refers to deaths reported anywhere from exactly x months in the past to just under

(x+ 1) months. Multiple deaths are accounted for in the event that the deaths occurred in different
months in the past. Only a few households reported 2 deaths in the same month; these cases of multiple
deaths are not taken into account here. Households report only the month and year of death; day of death
has been set to the 15th of the month. One household is dropped from all analysis due to inability to link
the household with its respective enumeration data.

138
One part of the criterion for those households over-sampled was the reporting of an

adult death within the 12 months preceding the enumeration survey, which predates the
wave 1 interview by 7 months, on average.

To account for the sample strategy, the data are weighted using sample weights
defined as:

QG)/H(i)

where Q0) is the population fraction of some outcome j and H0) is the sample fraction
of the same outcome (see Manski and Lerman, 1977). In a simple random sample we
expect to observe Q(j)= H(j). In this analysis case j is either “sick” or “well”, where
“sick” households were oversampled in every community. “Sick” households are those
that report either a prime-age adult too sick to work or a prime-age adult death in the
preceding 12 months at enumeration. Weights are calculated for both groups within
each of 51 clusters (or villages); thus there are 102 distinct sample weights. This
formula does not account for first stage sampling, which involved the selection of the
51 clusters among all 529 primary sampling units in the region. Therefore, these
weights cannot be used to infer regional statistics.

The enumeration data, collected over all household within the cluster, is used to
calculate the population proportions that are “sick” and “well”. For most clusters in
our sample, where j=“sick”, H0) is 14/ 16; conversely, H0) is 2/ 16 where j is “well”.
Population proportions of “well” households are considerable higher than the “sick”

households; the mean and median percent of the population “sick” is 8 percent.

139
The weighted probability of observing a death in any month over the 24 month

period is substantially lower (Figure 2). The bumps in this distribution, in particular at
19 months, are driven by a few deaths in “well” households which get considerable

weights as noted above.

V. Consistency in Reporting of Deaths

Although the KHDS was not designed to assess the consistency of reporting of
deaths, we can glean some information from comparison of the first wave with the
enumeration data. As aforementioned, the 12 month retrospective period covered at
enumeration is within the 24 month recall period covered in wave 1. Unfortunately,
the survey modules used at enumeration and wave 1 are not the same. Specifically, at
enumeration, households identified the age and cause of death (illness, accident,
childbirth, or other) in response to the question: “Has any adult in this household died
in the past 12 months?” In wave 1, households are asked: “Has any member of your
household who was residing with you died in the past 24 months?”. For non-household
members, households were asked: “Do you have any relatives who were living away
from your household and who died in the last 24 months?” Detailed information is
collected about each death reported in wave 1, including the age of the deceased, cause

of death, and the date of the death.6

 

6 For the enumeration survey, only adult deaths due to illness are considered. The adult deaths reported
in wave 1 are almost all reported as caused by illness, but some are due to other causes. Because deaths
due to non-illness causes are included in the wave 1 reported deaths, this favors fewer inconsistencies in
underreporting among households that report a death at enumeration and potentially results in more
inconsistencies among those reporting no death due to illness at enumeration.

140
Because of the lack of details for deaths reported at enumeration, confirmation

of re-reporting of a specific person can not be done.7 Rather, a more rudimentary
comparison can be done be assessing the reporting of at least one prime-age adult death
(15-50) in only one or both surveys. Since specific deaths are not matched in the two
surveys, the match rates for reporting of a prime-age death in both surveys may
understate misreporting if some matches for reporting a prime-age death in this analysis

do not capture the same person.

Comparison of the Two Surveys

Comparison of these data reveals marked inconsistency in reporting. Table 1
reports the cross-tabulation of the reporting of deaths by households at enumeration to
subsequent reporting in wave 1. From the enumeration survey, households can fall into
one of two categories based on the reporting of any death in the recall period. From
the wave 1 survey, households are divided into four groups depending on any reported
death over the preceding 24 months and the timing of the reported death with respect to
the enumeration recall period.

Among the 837 households surveyed at enumeration, 369 reported a death due
to illness of at least one adult 15-50 years during the 12-months preceding enumeration.
Less than half of the households that identified a death during the enumeration survey
(159 of 369) also identified a prime-age adult death of household members occurring in

the enumeration recall period when re-surveyed in wave 1 (Table 1, Panel A). On the

 

7 Although collected at enumeration, the data available at this time do not consist of the specific age
reported but rather the age as under 15, 15-50 and 51-65.

141
other hand, 155 households reported a death at enumeration in the 12 months preceding

enumeration and, subsequently reported no death over the entire 24 month recall period
at wave 1. Another 55 households reported a death at enumeration, but the death(s)
reported at wave 1 occurred outside of the 12 month enumeration recall period.
Among households reporting no death in the last 12 months at enumeration (n=468),
we observe 16 households reporting at least one death over the same period in the
subsequent wave 1 interview. Thus, in most cases, the inconsistency in reporting
surrounds the under-reporting of deaths in the subsequent survey.

There are several possible explanations for these apparent inconsistencies:
changes in classification of membership of deceased, recall errors in age of deceased
and errors in dating of deaths. These explanations assume that households do not
forget about the event altogether, which, given the nature of the event, seems unlikely.
With the exception of errors in dating of deaths which places them out of the range of
the recall period, these explanations assume that in some fashion, all deaths were
rereported. That is, the mismatch arises from the category that deaths at reinterview
were considered (15-50 year-old household members). An additional possibility for
inconsistencies between these data is that the respondents supplying this information for
the households change from the enumeration survey to the wave 1 interview. In both
cases, the respondent should have been the household head. However, in some
instances the household head is not in the household at the time of the survey or
headship changed from enumeration to wave 1. For both surveys, we do not have

information on who is responding for the household, although it presumed to be the

142
head when present. We do, however, know if headship changed from enumeration to

wave 1 and this information will be used below. The maintained, yet untested,
hypothesis is that the deaths reported at enumeration actually occurred.

Some deaths may be systematically re-classified by households as non—household
members. This would occur if, for example, a fraction of these deaths are not long-
term members but rather individuals who returned to the household shortly before
dying, perhaps returning in light of a fatal illness. About 30 percent of all prime-age
deaths observed during the panel (deaths occurring between wave 1 and wave 4) are
individuals who moved into the household between waves. Thus, these individuals
resided in the household for less than 7 months before they died. Households may be
re-classifying such short-term member deaths over time into the category of non-
coresident. Other research has suggested that AIDS-related deaths may be associated
with afflicted adults returning to their native areas shortly before dying (National
Research Council, 1996). Among other reasons, this may be to seek terminal care and
to save the family the cost of shipping the body.

The structure and sequence of the household survey offers a second explanation
for the potential reclassification of deaths. Respondents in wave 1 may have a more
narrow or restrictive sense of membership when reporting deaths in the final survey
module after completing the full household survey. The in-depth coverage of the

household elicited in wave 1 may lead respondents to be more discriminating in

143
classifying deceased persons as former household members as opposed to the very short

enumeration survey.8

Taking into account the re-classification of deaths and the possibility of errors in
age-reporting, the data are re-examined in two ways. Because the date of deaths was
not collected in the enumeration survey, we are not able to assess the extent to which
dating errors are driving these observed inconsistencies. Table 1, Panel B considers
death of individuals 15-65 who are reported in wave 1 as either household members or
non-member relatives who died in the same village. Table 1, Panel C broadens this to
include all adult deaths aged 15-65 (reported as either member or non-member)
reported in wave 1. Under this broader definition, a substantial number of
inconsistencies observed in Panel A are removed. In fact, about half of all households
reporting no member deaths in wave 1 despite reporting such death at enumeration do
report a death during the enumeration recall period if we expand deaths to cover non-
member relatives who died in the same village. Nevertheless, even with the broadest
description in Panel C, we still find that about one-third of households reporting a death
in the 12-month recall period from enumeration fail to report a death in this same
period at wave 1 (127 of 369). These broader set of characteristics of deaths reported
in wave 1 also introduce numerous inconsistencies for households reporting no death at

enumeration.

 

8 Another possibility is that at enumeration households were motivated to classify deceased non-member
relatives as household members if they perceived that the enumeration survey was tied to some sort of
social welfare or assistance program.

144
Probability of Underreporting in Wave 1

The interesting question posed by these inconsistencies is to what extent can we
identify whether certain households are more likely to report contrary information or
which household characteristics are linked to misreporting across these two surveys.
Table 2 presents results from probit estimation of the probability of reporting no deaths
in wave 1 among the subsample of 369 households that reported at least one death 15 -
50 at enumeration.9 By this classification, households that rereported a death outside
the enumeration recall period are not considered to be inconsistent. Thus, this narrows
the focus to the subset of households that reported a death at enumeration and do not
rereport at any time during the 24 month period at wave 1. The dependent variable is 0
for households reporting any adult 15-50 member death in the 24 month wave I recall
period and 1 for households reporting no death. The explanatory variables include
characteristics of the household head (age, age squared, and dummy variables for
highest level of education completed and sex) in wave 1, log per capita expenditure
reported in wave 1 (in Tanzanian shillings), and a dummy variable indicating whether
headship changed between enumeration and wave 1.

The set of characteristics of the household head, individually and jointly, do not
significantly influence the probability of misreporting. Jointly, the education of the

head is significant, although the interpretation of the coefficients is not clear.

 

9 We could expand our coverage of deaths reported at enumeration to include deaths 15-65. This,

however, adds only 2 more households and does not qualitatively change the results of this analysis.
Standard errors have been corrected for clustering at the household level.

145
Per capita expenditure (logged) is positively associated with misreporting,

although the effect is quite small. 10 This suggests that among households with a death
in the distant past, wealthy households, for whom the death may be smaller economic
crisis for the household, are more likely to report the event inconsistently. If these
inconsistencies are driven by re-classification of deaths, then deaths in wealthier
households are more likely to be re-classified from members to non-members than
deaths in less wealthy households. Additional support of this supposition is drawn from
the deaths reported in waves 24. Households reporting an adult 15-50 death of a
household member who joined the household between waves (individuals who join the
household and die within 7 months) have, on average, higher levels of per capita
expenditure. This may indicate that wealth is identifying those households in which the
deceased where short-term members and, thus, are more likely to be re-classified as
non-household members.

A change in headship between enumeration and wave 1 is associated with more
errors in reporting. Probability of misreporting is 20 percentage points lower in
households where the head did not change from the enumeration survey to wave 1.

This is ahnost half of the mean probability of misreporting in the sample (.42).

VI. Conclusions
In a wide range of social science research, retrospective questions are used to

inform researchers about past behaviors and events that are otherwise unobtainable.

 

‘0 The probability of misreporting increases by 1 percent for a 14 percent increase in expenditure per
capita (10,000 Tanzania shillings), evaluated at the mean level of expenditure per capita.

146
The quality of these data, no doubt, depend on the nature of the event being recalled

and the survey design. Additionally, the accuracy may depend on the characteristics of
the respondents themselves.

This study has explored the consistency of recall of prime-age adult deaths as
reported by surviving household members. This was undertaken in two ways. First,
the reporting of prime-age deaths by month over the 24 month recall in the first full
household interview was examined. Incidence of deaths did not appear to be heaped
and, when weighted, gave the low probabilities that would otherwise be expected from
regional census data. Second, reporting in an initial interview was compared with a
subsequent, and more detailed reinterview. More specific analysis was confounded due
to the limited amount of data collected in the initial contact. Nevertheless, comparison
of households reporting any death in past 12 months at first interview with reporting of
any adult death in past 24 months at the reinterview 7-8 months later revealed stark
underreporting in the second survey. Several alternative interpretations could explain
these results, but lack of data prohibit exploring many of these. One outstanding caveat
to these results is the possibility that differences in the survey instruments rather than
recall errors can explain inconsistencies across surveys. Further analysis found that
some characteristics of the household, in particular per capita wealth and change in
household head between interviews, were associated with a higher probability of
underreporting among those households reporting a death in the first survey.

Even if a sizable number of these inconsistencies were driven by changes in

headship and differences in survey modules, the discrepancies are large enough to be

147
surprising given the landmark nature of a prime-age death in households. They suggests

that even for major lifetime events, details surrounding them are variable at least among
the set of characteristics including date, age and household membership. In terms of
informing methodology in the specific area of incidence of deaths, researchers may
need to emphasize shorter recall periods or at least not emphasize in analysis those
deaths reported in the more distant past.11

With respect to the analysis undertaken in Chapter 1, the retrospective data on .
deaths reported in the first household interview is truncated so that only the information
within 6 months of the household interview is used. Without further information, it is
not possible to assess what portion of the recall period in wave 1 is reliable. By
truncating these data at 6 months past, however, the recall period for all waves is

approximately 6 month.

 

“ For example, Tibaijuka (1997) collects data on adult (over 14 years) deaths in households over the
period 1980-1989 which may suffer substantial recall errors.

Percent of Households Reporting

Percent of Households Reporting

148

 

 

l 3 5 7 9 11 13 15 17
MomhsSinceInterVieW

Figure l: Households Reporting a Prime-Age
Death, Unweighted (n=837)

 

0 r l l l l l l i r l i l
l 3 5 7 9 11 13 15 17 19 21 23
Months Since Interview

 

Figure 2: Households Reporting a Prime-Age
Death, Weighted (n=837)

149

Table 1: Reporting of Deaths During Enumeration and at Wave 1 Interview,

Percent of Households (n=837)

 

 

 

 

 

 

 

PANEL A
Wave I :
Enumeration: Reporting of death of adult 15-50 (household member)
Reporting of death in 24 months Jareceding wave 1
15-50 in 12 months After enumeration Enumeration recall More than 12 months
precetflrg enumeration None & before wave 1 period before enumeration
None (n=468) 401 48 16 3
At least 1 death (n=369) 155 20 159 35
PANEL B
Wave 1 :

Enumeration:
Reporting of death
15-50 in 12 months

Reporting of death of adult 15-65 (members & non-members
in same villageLin 24 months preceding wave 1

 

After enumeration Enumeration recall More than 12 months

 

 

 

 

 

 

 

preceding enumeration None & before wave 1 period before enumeration
None (n=468) 296 90 62 20
At least 1 death (n=369) 88 41 201 39

PANEL C

Wave 1 :

Enumeration: Reporting of death of adult 15-65 (members and non-members)
Reporting of death in 24 months preceding wave 1
15-50 in 12 months After enumeration Enumeration recall More than 12 months
preceding enumeration None & before wave 1 period before enumeration
None (n=468) 173 119 142 34
At least 1 death (n=369) 43 41 242 43

 

 

Note: Deaths reported at enumeration consist of only those due to illness. Deaths reported at wave 1 are
those due to any cause. The cause of death reported in the vast majority of cases for both surveys was
illness. Households with multiple deaths over the 24—month recall period, including at least one during
the 12 months before enumeration, are assigned to the category “12 months before enumeration” ;
households with multiple deaths after enumeration and before the 12 month recall are assigned to the last
category “more than 12 months before enumeration”.

 

150

Table 2: Probability of Reporting no Adult Deaths in Past 24 Months
at Wave 1, Among Households Reporting an Adult Death
in Last 12 Monthsat Enumeration (n=369)

 

 

 

 

Mean
Dependent variable 0.420
Explanatory variables
Age of household head 50.280 0.003
(0.022)
[0.001]
Age of head squared (/100) -0.004
(0.020)
{-0.002}
Education of head 1-6 years? 0.439 0.281
(0.184)
[0. 1 10]
Education of head 7+ years? 0.304 -0.113
(0.223)
{-0.044}
Female head 0.317 -0.215
(0.162)
{-0.083}
Log expenditure per capita 11.171 0.184*
(0.101)
[0.072]
No change in household head 0.938 -0.519*
(0.274)
{-0.205}
Constant -1.845
(1.152)
Pseudo R2 0.028

 

Note: B are estimated using probit estimation with standard errors in
parentheses corrected for clustering at the household level. In brackets,
coefficients have been converted to change in probability evaluated at the
mean of other regressors. The dependent variable is equal to 1 for households
reporting no household member 15-50 death over the 24-month recall period,
and equal to 0 for households reporting any such death(s) in the 24-month,
regardless of the date reported. Log expenditure per capita is expenditure in
last 12 months (in Tanzanian shillings) reported in wave 1 measured per
capita, replacing actual number of members in household with a calculation
of the number of residency-adjusted adults. 9 education variables are jointly
significant at 5%. * indicates statistical significance at 10%.

LIST OF REFERENCES

LIST OF REFERENCES

Ainsworth, Martha, Godlike Koda, George Lwihula, Phare Mujinja, Mead Over and
Innocent Semali. 1992. “Measuring the Impact of fatal Adult Illness in Sub-Saharan
Africa.” Living Standards Measurement Study no. 90, World Bank, Washington, DC.

Blane, DB. 1996. “Collecting Retrospective Data: Development of a Reliable Method
and a Pilot Study of Its Use.” Social Science and Medicine 42(5): 751-757.

Deaton, Angus. 1997. The Analysis of Household Surveys: Microeconometric Analysis for
Development Policy. Washington, D.C.: Johns Hopkins University Press.

Diebold, Francis, David Neumark, Daniel Polsky. 1997. “Job Stability in the United
States.” Journal of Labor Economics 15(2): 206-233.

Klerman, Jacob. 1993. “Heaping in Retrospective Data: Insights from the Malaysian
Family Life Surveys’ Breastfeeding Data.” RAND, Santa Monica, California

Loftus, Elizabeth and Wesley Marburger. 1983. “Since the Eruption of Mt. St. Helens,
Has Anyone Beaten You Up? Improving Accuracy of Retrospective Reports with
Landmark Events.” Memory and Cognition 11(2): 114-120.

Manski, Charles and Steven Lerman. 1977. “The Estimation of Choice Probabilities from
Choice-based Samples.” Econometrica 45(8): 1977-1988.

National Research Council. 1996. Preventing and Mitigating ADS in Sub-Saharan
Africa. Washington, D.C.: National Academy Press.

Over, Mead. 1995. “Adults Deaths in the Sample of Households: Survey Methods and
Results.” draft, Policy Research Department, World Bank, Washington, DC.

Scott, Chris and Ben Amenuvegbe. 1990. “Effect of Recall Durations on Reporting of
Household Expenditures.” Social Dimensions of Adjustment in Sub-Saharan Africa,
Working Paper 6, World Bank, Washington, DC.

Smith, James and Duncan Thomas. 1997. “Migration in Retrospect: Remembrances of
Things Past.” RAND, Santa Monica, California.

151

152

Sudman, Seymour, and Norman Bradburn. 1974. Response Effects in Surveys: A Review
and Synthesis. Chicago, Il: Adline.

Tibaijuka, Anna. 1997. “AIDS and Economic Welfare in Peasant Agriculture: Case
Studies from Kagabiro Village, Kagera Region, Tanzania.” World Development 25(6):
963-975.

Torelli and Trivellato. 1993. “Modeling Inaccuracies in Job-search Duration Data.”
Journal of Econometrics 59: 187-211.

Ureta, Manuelita. 1992. “The Importance of Lifetime Jobs in the US. Economy,
Revisited.” American Economic Review 82(1): 322-335.

Van Der Vaart, Wander, Johannes Van Der Zouwen and Wil Dijkstra. 1995.
“Retrospective Questions: Data Quality, Task Difficulty, and the Use of a Checklist.”
Quality and Quantity 29: 299-315.

World Bank. 1992. Tanzania AIDS Assessment and Planning Study. World Bank,
Washington, DC.

World Bank. 1993. “Report of a Workshop on: the Economic Impact of Fatal Illness in
Sub-Saharan Africa.” World Bank, Washington, DC. and the University of Dar es
Salaam.

MICHIGAN 51'an UNIV , LIBRARIES
11111 HWIIWIIHI 1111111111 ”11111 ”1111111111111111111111|
31293018419030