Noncooperative Decision Making in the Household:
Evidence from Malawi
Selma Walther
August 2016
Abstract
This paper proposes a novel test of productive e¢ ciency in the household that also allows
a test of noncooperative decision making. I extend the collective model (Chiappori 1988, 1997)
to allow labor choices to a¤ect future bargaining power by raising the value of outside options.
Even if household consumption sharing is e¢ cient, labor choices are no longer e¢ cient. Using
data on Malawi, where there is predetermined variation in land rights that determine outside
options in marriage, I show that individuals spend more time on agricultural labor and less
time on wage labor when household land is theirs. They also have lower overall income and
consumption. The results are inconsistent with the fully e¢ cient collective model but consistent
with a noncooperative model with limited commitment, where individuals allocate their labor
supply to maximize future bargaining power. The empirical results are not driven by di¤erences
in agricultural productivity, wages or savings. Limited commitment can lead to ine¢ cient
allocations that reduce household welfare.
Keywords: collective model, noncooperative model, production, Malawi
JEL Classi…cation: D12, D13, J12
Department of Economics and Nu¢ eld College, University of Oxford, [email protected]. This
paper was previously circulated under the title "Kinship and Consumption" with the author’s maiden name Telalagic.
The author thanks Toke Aidt, Wiji Arulampalam, Martin Browning, Lucia Corno, Ian Crawford, Tom Crossley,
James Fenske, Michael Keane, Pramila Krishnan, Hamish Low, Costas Meghir, Imran Rasul, Christine Valente,
Ansgar Walther and seminar paticipants at the University of Cambridge, University of Oxford, Université Libre de
Bruxelles, KU Leuven, University of Groningen, Institute for Fiscal Studies and the Barcelona Summer Forum in
Family Economics for helpful comments.
1
1
Introduction
Becker’s approach to marriage (Becker 1973, 1974) rests on the simple idea that individuals marry
when their expected surplus from marriage is greater than the sum of their utilities while single.
This surplus can depend on a number of factors, such as the gains to specialisation, or the scope
for economies of scale. The collective model (Chiappori 1988, 1992) assumes that this surplus is
generated and shared in an e¢ cient way: nothing is left on the table. Tests of intra-household
allocation of consumption have failed to reject e¢ ciency (Browning and Chiappori 1998, Bobonis
2009, Rangel and Thomas 2012). However, Udry (1996) rejects e¢ ciency in household production
by comparing plot yields of male- and female-controlled plots and Udry and Goldstein (2008) show
that individuals who have weak tenure security of their plots are less likely to leave their land
fallow, which is ine¢ cient. There are alternative models that explain ine¢ ciency as a result of
noncooperative behavior with limited commitment (Iyigun and Walsh 2007, Lundberg and Pollak
2003), but these models have not been tested.
In this paper, I propose and implement a novel test of production e¢ ciency in the household,
that also tests whether ine¢ ciency is driven by limited commitment. The empirical results reject
production e¢ ciency and are instead consistent with an extension to the collective model that
incorporates noncooperative choices in labor supply.
I consider a standard model of the household, which makes two decisions: …rst, labor choices
are made, which generate consumption; second, this consumption is shared between spouses. In
the collective framework, both choices are e¢ cient. In the …rst step, labor is allocated in order
to maximise the total amount of consumption generated (Chiappori 1997, Apps and Rees 1997).
In the second step, consumption is shared e¢ ciently between spouses according to a sharing rule
that depends on spouses’outside options (known as distribution factors). In this case, bargaining
power should not a¤ect how labor is allocated in the …rst period, as long as the disutility from all
types of labor (e.g. agricultural labor and wage labor) is the same. I compare this prediction with
a model where the …rst period choice is noncooperative because spouses are unable to commit to
a particular labor allocation. Bargaining power in the second period depends on previous labor
allocation, because labor supply increases the value of future income, which determines outside
options. In this version of the model, labor choices are no longer independent of distribution
factors and are in general ine¢ cient: individuals spend more time on the type of labor that most
improves their outside option. This also results in lower overall consumption available for sharing
in the second period.
I show that the empirical results are inconsistent with the fully e¢ cient collective model but
consistent with the limited commitment collective model. I test productive e¢ ciency by estimating
how time allocation across di¤erent types of labor, and overall output, depend on predetermined
variation in land ownership in Malawi. Malawi is a useful laboratory for this test because of the
coexistence of two descent rules, which are determined by ethnic group at birth. Some ethnic
groups practise matrilineal descent, where ancestry is traced through the mother. In matrilineal
communities, women inherit land from their mothers when they marry, and keep this land if they
2
divorce. Other ethnic groups practise patrilineal descent, with ancestry and therefore property
inheritance traced through the father. It is well established in the literature that matrilineal women
have strong outside options due to their access to land, while men have strong outside options in
patrilineal communities (Lamphere 1974, Johnson 1988, Davison 1997).1 Social norms dictate that
men divide their working time between agricultural labor and wage labor, while women specialise
in agricultural labor. This asymmetry means that women are constrained in the labor choices they
can make: in the data, their wage hours are negligible (see Table 3). In the context of the theoretical
model, this means that only men can make a choice in the …rst period.
According to the collective model, the identity of who owns the family land should not a¤ect
labor supply choices nor overall production levels, conditional on the productivity of di¤erent labor
types. Instead, I …nd that men spend more time on agricultural labor (+1:5 hours per week) and
less time on wage labor ( 1:4 hours per week) when they are patrilineal, controlling for a rich set
of covariates, including temperature, rainfall and household-level soil quality. They generate 10%
less consumption overall than matrilineal men. These …ndings are di¢ cult to reconcile with an
e¢ cient model, but are consistent with a mechanism where men make noncooperative choices in
labor supply because this raises their bargaining power in future periods. In the case of patrilineal
households, household income could be increased if men reallocated their time away from their
land and towards wage labor. This …nding is also inconsistent with unobserved di¤erences in
agricultural productivity, which would imply higher agricultural labor for patrilineal men but also
higher income. The di¤erence in labor supply and consumption between patrilineal and matrilineal
households is not observed for a placebo group of households that do not own any land.
To address concerns over omitted variables that correlate with descent, labor and consumption,
I pursue several strategies. First, I control for a rich set of geographical, community and individual
characteristics. I also control for various measures of plot quality and restrict the sample to areas
where patrilineal and matrilineal groups are well mixed. To account for residual colonial in‡uence, I
construct a variable that measures the distance to the nearest railway station built during colonial
times using the GPS coordinates of stations and villages. I also take account of ethnic group
characteristics using the Murdock Ethnographic Atlas (1967).
Next, I implement the strategy described in Oster (2015) and Altonji, Elder and Taber (2005),
which uses selection on observed variables to inform on selection on unobserved variables. In the
most conservative estimates, I …nd that selection on unobservables would have to be eleven times
as strong as selection on observables to explain the gap in wage work hours between patrilineal and
matrilineal men, and four times as strong to explain the gap in household consumption. Observables
and unobservables would have to be negatively correlated in order for unobservables to explain the
gap in agricultural work hours. The estimates also provide bias-adjusted bounds on the coe¢ cients
of interest, with patrilineal households losing 7:6
1
9:9% consumption compared to matrilineal
I de…ne an individual’s outside option as his or her utility when divorced because divorce is frequent and not
stigmatised in Malawi: lifetime divorce probabilities are between 40-65% and over 40% of women remarry within the
…rst two years after a divorce (Reniers 2003). For societies where divorce is uncommon, an alternative outside option
is within-household noncooperation, as in the separate spheres model of Lundberg and Pollak (1993).
3
households, as a result of spending 13:6
17:2% more time on agriculture and 25:4
27:1% less
time on wage work, compared to baseline.
To provide more evidence on the mechanism, I analyse husbands’wages and agricultural productivity. I …nd that, consistent with an overinvestment story, patrilineal men have substantially
lower agricultural productivity than wages, and this wedge is smaller for matrilineal men. I also
show that the results are not driven by high discount factors, as measured by saving behavior, nor
the inheritance of political rank, which sometimes goes with land inheritance.
The …ndings produce several other interesting implications. First, the results in this paper
suggest an e¢ ciency gain to matrilineal descent. Matriliny has remained surprisingly prevalent in
Malawi, despite e¤orts by British colonialists and religious missionaries in the early 20th century to
convert communities to patrilineal descent (Peters 1997, 2002). Matriliny weakens men’s incentives
to overinvest in the land.2 This is a potential reason for the persistence of matrilineal descent to
this day.
Second, endogenous bargaining power leads to ine¢ cient decisions when there is limited commitment. This suggests that there may be e¢ ciency gains from commitment devices in marriage,
such as prenuptial contracts that condition on labor supply choices. This coheres with recent …ndings on the potential e¢ ciency bene…ts of prenuptial contracts in Voena (2015) and Bayot and
Voena (2015).
This paper contributes to the family economics literature by proposing a new test of e¢ ciency
in production decisions, rejecting this test, and …nding evidence for an extension to the collective
model that allows for limited commitment in the …rst stage. Allowing bargaining power to be
endogenous to choices has precursors in the family economics literature. For example, in a model
where bargaining power is increasing in income and decreasing in fertility for women, and education
levels a¤ect subsequent income, women are predicted to overinvest in their education because it
makes up for the negative e¤ect of having a child on the sharing rule (Iyigun and Walsh 2007). In
an alternative model with limited commitment, couples make location choices in the …rst period,
which advantage one spouse in their subsequent consumption share. Even if consumption sharing
is e¢ cient, the …rst period location choice may not be (Lundberg and Pollak 2003). Finally,
bargaining power has also been endogenised in a fully e¢ cient collective model of consumption
with commitment (Basu 2006). However, none of these papers provides empirical evidence for
ine¢ cient labor supply allocations due to endogenous bargaining power. The theory is also related
to a growing literature on noncooperative models of the household (Cherchye, Cosaert, Demuynck
and De Rock 2016, Lechene and Preston 2011, Chen and Woolley 2001).3
There is empirical evidence for noncooperative behavior in various contexts, including Kenya,
where men use their power over migration decisions to reduce the ability of their wives to earn their
2
This is consistent with a related literature on tenure insecurity and long-term investments in land: e.g. Besley
(1995), Place and Otsuka (2001), Kishindo (2010).
3
The model also relates to Rainer (2007), who discusses theoretically the role of prenuptial contracts in a model
where individuals invest in a relationship-speci…c asset that determines individuals’ outside options, and therefore
individuals’bargaining power over the asset during marriage.
4
own income (McPeak and Doss 2006), and Northern Cameroon, where women overinvest in those
crops whose income they control (Jones 1983). There is also evidence of a lack of consumption
smoothing in the household, with allocations depending on whether shocks a¤ect the husband
or wife (Doss 2001, Du‡o and Udry 2004) and wives bearing a larger burden of adverse shocks
(Dercon and Krishnan 2000). In experimental contexts, spouses are observed to engage in income
hiding (Ashraf 2009), even if this reduces their expected income (Jakiela and Ozier 2016). This
paper provides a more general test of e¢ ciency in production by taking into account choices across
several types of labor income, rather than focusing only on agricultural work.
The paper proceeds as follows. In Section 2, I discuss the two descent types in Malawi. Section 3
outlines the collective model of labor supply and consumption, as well as a noncooperative extension
to the collective model. In Section 4, I describe the data and present some summary statistics.
Section 5 presents the main empirical analysis, while Section 6 addresses selection on observables
and unobservables. Section 7 provides further evidence on the mechanism and Section 8 concludes.
2
Background
In rural Malawi, individuals belong to ethnic groups (sometimes referred to as tribes), whose
rules are important for family life. Ethnic groups follow either matrilineal or patrilineal descent,
and of the eleven main ethnic groups, six are matrilineal and …ve are patrilineal (Spring 1995),
corresponding to around 60% of rural households being of matrilineal descent. In a matrilineal
household, the woman traditionally receives land from her mother when she marries, which she
keeps if the couple divorce (Berge et al. 2014, Peters 2010, Davison 1997).4 The husband has no
rights to this land.5 Note that this di¤ers from other matrilineal societies studied in the literature
(e.g. La Ferrara 2007), where land passes from brother to sister’s son.6 In patrilineal households,
the opposite happens: men receive land from their families on marriage, which they keep on divorce,
with the woman returning to her family. Divorce matters: Malawi has one of the highest divorce
rates on the continent, with one in two marriages dissolving (Reniers 2003).
I interpret descent as a measure of outside options, because its most important role is in determining land inheritance at time of marriage. However, there may be other economically salient
features of descent that could drive labor supply and consumption allocations. According to Tyler
(1889), descent determines two rules: property inheritance and succession to rank or status. Adams
(1999) describes the matrilineal Fanti of West Africa, where in addition to property inheritance
and succession to rank, descent also guarantees use of clan land and a proper funeral or burial.
When considering household choices about consumption and labor supply, it is probably safe to
ignore the latter, and use of clan land is similar to property inheritance. This leaves succession to
4
Matrilineal households also tend to be matrilocal, with the husband locating in the wife’s village on marriage,
while patrilineal households tend to be patrilocal (the wife moves to the husband’s village).
5
There are cases of men purchasing land in matrilineal communities, but this is overwhelmingly to give it to their
daughters when they marry (Berge et al. 2014, Peters 2010).
6
Although land is shared based on descent following divorce, consumption goods tend to be shared equally on
divorce. Child custody tends to follow descent rules, so going to matrilineal women and patrilineal men.
5
political o¢ ce, in addition to property inheritance, as being a key factor determined by descent. It
will be important to show that the results obtained in the empirical section are not driven by the
role of descent in succession to political o¢ ce, so that descent can be interpreted as a measure of
outside options.
Figure 1, a map of Malawi, depicts the dispersion of descent by district.7 Darker shading
represents districts where matriliny is more common. In the Southern region, most districts are
over 50% matrilineal. In the Central region, there is a similar number of matrilineal and patrilineal
people, while the Northern region is mostly patrilineal. This somewhat uneven dispersion of descent
will be addressed in the robustness checks in the empirical part of this paper.
Labor allocation in rural Malawi is a¤ected by gender-based social norms. Almost all households
derive a substantial amount of their income from agriculture, with women tending to engage in
agricultural labor and performing many tasks on their own (Hirschmann and Vaughan 1983).8
According to the FAO (2011), 93% of women are subsistence farmers.9 Men usually work both
on the land and for wages.10 It is rare for women to work for wages, unless they are unmarried
(Spring 1995), which implies that men are predominantly responsible for providing the household’s
consumption goods (Schatz 2002). On the other hand, domestic labor is predominantly carried out
by women (Spring 1995). These patterns guide the assumptions made in the theoretical framework
in the next section.
3
Theoretical Framework
3.1
The Collective Model
In order to determine whether household behaviour in Malawi is e¢ cient, I describe the canonical
‘collective’ model of the household (Chiappori 1998, 1992) and derive predictions on how labor
supply and output relate to descent rules. I then describe an extension of the collective model
that allows individuals to invest in their outside options, which generates alternative predictions
for behaviour.
The collective model assumes that household decisions are made e¢ ciently, so that allocations
can be modelled as the result of maximising household social welfare with Pareto weights on individual utilities. These Pareto weights represent the relative power of individuals in the decision-making
process and are driven by ‘distribution factors’, which are variables that do not a¤ect income or
preferences but do a¤ect relative weights.
7
The prevalence of matriliny and patriliny by district is calculated based on the Living Standards Measurement
Study data used in the empirical section of this paper. For the purposes of this map, in those villages where both
types of descent are practised, half of the households are apportioned matrilineal descent while half are apportioned
patrilineal descent. The …gures are weighted based on the sampling strategy of the data (see footnote 24).
8
This is di¤erent from the setting in Udry (1996), where men and women control separate plots.
9
Maize is the most commonly farmed crop, followed by pigeonpea and tobacco (in terms of the number of households reporting that they farm it). Harvesting, storage and seed selection are mostly done by women, while ridging,
planting and weeding are done jointly by men and women (Kerr 2005, Hirschmann and Vaughan 1983).
10
Wage labor refers to any work for a salary, commission or in-kind payment, excluding agricultural activities on
other farms. A typical task might be brick-laying.
6
Figure 1: A map of Malawi depicting the prevalence of matriliny and patriliny by district.
7
A household consists of a husband (a) and wife (b) who both enjoy consuming a vector of private
goods x. Their utilities are ui (xi ) for i = a; b. The prices of private goods are a vector p.
Each household member spends ni hours working for wages wi and hi hours working on the
family land.11 Agricultural labor yields output AF (h), where A is a measure of agricultural productivity. This output can be sold to the market at price q. The household’s income is then
y = w n + qAF (h), and its budget constraint for private consumption is
p
X
xi = w n + qAF (h).
(1)
Normalizing total time to one, the time allocation constraint is
ni + hi = 1; 8i:
(2)
The collective model of household choices is
max
fx;h;ng
where
(z)ua (xa ) + ub (xb ) subject to (1) and (2),
(3)
(z) is the husband’s Pareto weight and z is a vector of distribution factors that re‡ect
bargaining power.
Following Chiappori (1997), a necessary condition for solving (3) is pro…t maximization by a
hypothetical "household …rm" that produces agricultural output and hires labor:
= max qAF (h)
h
w h:
(4)
Hence, optimal choices of agricultural labor h are fully determined by the price of agricultural
products q, productivity A, and the level of outside wages w. In particular, the choice of h is
independent of the sharing rule
and any distribution factors z. This result is very general; see
Chiappori (1997) and Apps and Rees (1997).
3.1.1
Empirical predictions of the collective model
Descent systems determine land rights and therefore are a good candidate for a distribution factor.
However, distribution factors must not a¤ect productivity or prices. A central empirical prediction
of the collective model then is that, controlling for productivity and prices, descent systems will not
a¤ ect the share of time spent on agricultural labor, nor the total output of the household.
Moreover, the model predicts that the share of each household member’s time spent on agricultural labor will (i) increase with agricultural prices or productivities, and (ii) decrease with market
wages.
A challenge in testing the above prediction is that we cannot perfectly measure agricultural
productivity or land quality. If there are residual (unobserved) productivity di¤erences across
11
I write n = (na ; nb ), h = (ha ; hb ) and w = (wa ; wb ) for the vectors of allocations and wages.
8
descent systems, we might wrongly attribute resulting di¤erences in labor allocations to descent
itself, leading to a false rejection of e¢ ciency.
This issue can be addressed by conducting an additional test on overall levels of household
consumption.12 The collective model predicts that households in more productive regions dedicate
more time to agricultural labor, and also enjoy a higher level of overall household consumption due
to their superior productivity. In other words, in the presence of residual di¤erences in agricultural
productivity, the collective model predicts that agricultural labor shares and overall consumption
will be positively correlated across descent systems. This is ruled out in the empirical section.
In addition, several approaches will be taken in the empirical part of the paper to address
potential unobserved di¤erences in agricultural productivity.
With no di¤erences in agricultural productivity but in the presence of wage di¤erences across
descent systems, wages, wage labor shares and overall consumption will be positively correlated
across descent systems. This is also ruled out in the empirical section.
Note that these predictions are robust to introducing leisure in the model. Suppose that individuals enjoy consuming goods x and leisure l, with utility ui (xi ; li ) and time allocation constraint
n i + hi = 1
li ; 8i:
(5)
The household …rm’s problem (4) is unchanged, so it remains the case that agricultural labor
choices h are fully determined by prices and productivity.13 In the presence of residual productivity
di¤erences, it also remains the case that agricultural labor shares and consumption will be positively
correlated, as long as private consumption is a normal good.14
3.2
Noncooperative choices and limited commitment
As an alternative to the collective model, suppose that e¢ cient household bargaining occurs after
agricultural output and wage income has been earned. In an initial stage, the household members
unilaterally choose their labor allocation. This setup describes a situation where it is not possible
to make binding commitments about one’s labor supply. Assume further that only the husband
makes a labor supply decision. This is justi…ed by the empirical fact that women in Malawi engage
12
I assume that savings are negligible, so that output equals consumption. Brune, Giné, Goldberg and Yang (2016)
study farmers in Malawi and …nd average savings of approximately 8.4 days worth of household expenditure.
13
Note that the result generalises to di¤erent e¤ort costs of wage and agricultural labour, but not if these e¤ort
costs also vary between patrilineal and matrilineal households.
14
It is more di¢ cult to obtain predictions for leisure choices l (and therefore wage labour choices n). These variables
i
i
i
depend on bargaining power in general. Let M RSjl
= @u
= @u denote the marginal rate of substitution between
@li @xi
j
leisure and private good j for member i. We have the …rst-order condition
i
M RSjl
=
wi
; 8i:
pj
The MRS is therefore uniquely determined by prices. Note also that the MRS is, in general, a function of both private
consumption xi and leisure li . When bargaining power shifts, private consumption xi will generally be reallocated.
Therefore, it will generally be necessary for leisure choices to also be reallocated to restore the e¢ ciency condition
above.
9
in little to no wage labor; see Table 3.
At the bargaining stage, the household takes its income y = wn + qF (h) as given; as only the
husband makes a choice, n, h and w refer to the husband’s allocation and wages. Since bargaining
is e¢ cient, the household’s choices solve
xi = arg max ui (x) subject to p x = si (z)y ;
x
where si is the share of income allocated to member i, which is commensurate with i’s bargaining
power (and which satis…es sa (z) + sb (z) = 1). This is equivalent to the maximisation problem in
(3) (see e.g. Browning, Chiappori and Weiss 2014).
Bargaining power depends on the incomes which spouses would enjoy outside of marriage. These
are determined by the formal laws and social norms governing divorce. Upon divorce, the husband
obtains a share
2 [0; 1] of household wage income, and a share
2 [0; 1] of agricultural income.
To formalize the e¤ect of outside options on bargaining power, let z = [ (wn) + qF (h)] =y
denote the share of household income obtained by the husband in case of divorce, and assume that
the sharing rule at the bargaining stage satis…es sa =
(z) and sb = 1
(z), where d =dz >
0. Relative income is commonly used as a distribution factor in the literature (e.g. Browning,
Bourguignon, Chiappori and Lechene 1994, Hoddinott and Haddad 1995), justi…ed by empirical
results that reject income pooling (Lundberg, Pollak and Wales 1997, Thomas 1990). Divorce laws
and control of land have also been used as distribution factors (Chiappori, Fortin and Lacroix 2002,
and Udry 1996, respectively).15
The husband’s indirect utility, as a function of bargaining power z and household income y, is
therefore
V (y; z) = max fua (x) subject to p x = (z)yg :
x
At the labor allocation stage, he solves
max V (y; z) subject to y = wn + qF (h), z =
h;n
(wn) + qF (h)
and h + n = 1 :
wn + qF (h)
The …rst-order condition for optimality16 can be written as
dy
dy
=
+y
dn
dh
dz
dh
dz
dn
0 (z)
(z)
:
(6)
The e¢ cient choice is to equalize the marginal products of wage and agricultural labor: dy=dn =
dy=dh. In the current setting, however, agricultural labor yields an additional bene…t for the
husband if it increases his bargaining power by more than wage labor does. This e¤ect is captured
by the second term in (6), and it leads the husband to oversupply agricultural labor.17
15
See the discussion on distribution factors in chapter 5 of Browning, Chiappori and Weiss 2014.
Obtained by maximising (z)y with respect to h and n subject to the constraint h + n = 1.
dy
17
Note that this implicitly assumes that agricultural labour generates agricultural income (i.e. dh
> 0), and descent
( ) determines how agricultural income a¤ects the outside option. This is obvious when we consider that the value
16
10
The division of output upon divorce is now crucial for household e¢ ciency. Evaluating dz=dh
and dz=dn and substituting into (6), we have
dy
dy
=
+(
dn
dh
) F (h) + nF 0 (h)
qw 0 (z)
:
y (z)
It is clear that the husband over-supplies agricultural labor if and only if
case
dy
dh
<
dy
dn ,
(7)
>
, because in this
so that a marginal reallocation from h to n would increase y. Further, the e¢ cient
choice maximises household output and therefore consumption. Any deviation from the e¢ cient
choice results in lower overall consumption, even if one spouse obtains a higher share of it.
This result predicts that limited commitment and noncooperative decision making can lead to
ine¢ cient behavior. Limited commitment has been studied extensively in the recent literature;
Voena (2015) is an important example. The model derived here is similar in spirit to Iyigun and
Walsh (2007), who model pre-marital investment in education and its relationship to subsequent
fertility; they …nd that women overinvest in their education to protect their marital surplus, which
declines when they have children. Lundberg and Pollak (2003) model the location decision of
spouses, where similar to the set-up here, the …rst period location choice is made noncooperatively
but the second period consumption allocation is e¢ cient. They show that dynamic e¢ ciency does
not necessarily result. However, neither of these papers test the mechanism empirically.
3.2.1
Empirical predictions of the noncooperative model
A central empirical prediction of the collective model with investment in outside options is that
ine¢ cient choices can occur if post-divorce resource allocation of one type of income is not equal to
post-divorce resource allocation of another type of income. In this case, labor allocation is skewed
towards earning the type of income that most improves the spouse’s outside option.
In Malawi, the two key types of male labor are agricultural labor and wage labor (see Table 3).
Agricultural labor improves the value of land, both because of future crops that can be harvested,
and because conservation e¤orts such as tree planting can reduce erosion and thus improve the
fertility of the land. Therefore, in patrilineal households, where men keep most of the land on
divorce, higher levels of agricultural labor are expected, relative to matrilineal households, where
men own a smaller share of the land. These land shares are discussed in more detail in the next
section and can be seen in Table 2.
With an additional assumption on how wage income is shared on divorce, the model also
has empirical predictions regarding household consumption. Suppose that wage income is shared
equally on divorce, such that
is close to one half for both patrilineal and matrilineal households -
this is not unreasonable, as recent judicial changes have resulted in courts ordering equal division of
purchases such as movable property and personal belongings following divorce (Mwambene 2005).
of agricultural labour is in fact the value of future harvested crops. Clearly these are tied to the land. Agricultural
labour can also raise the value of the land indirectly through soil conservation measures, for example. In fact, there
is a substantial literature documenting that matrilineal men invest less in household land (Place and Otsuka 2001,
Lovo 2016, Kishindo 2010).
11
Table 1: Empirical predictions of theoretical models
Agricultural labor (Men)
Di¤erence between Patrilineal and Matrilineal HHs
Collective model
Noncooperative model
0
+
Wage labor (Men)
0
-
Consumption (HH)
0
-
Then, using the average share of land owned by the husband in Table 2 as a measure of , greater
ine¢ ciency is expected for patrilineal households, because the wedge between
and
is greater.
This implies lower household consumption for patrilineal than matrilineal households.
The predictions of the fully e¢ cient collective model and the limited commitment model with
noncooperative decision making are summarized in Table 1.
The remainder of the paper is devoted to examining the consistency of the data with these two
models.
4
Data Sources and Description
The data are from the Malawi Third Integrated Household Survey (IHS), conducted by the World
Bank and the Malawi National Statistical O¢ ce (NSO). In the …rst stage, 768 communities were
selected based on probability proportional to size within each district. In each community or village,
16 households were randomly selected, leading to a sample of 12271 households. I restrict the sample
to rural households headed by a married couple, which yields a sample of 7203 households.
4.1
Variables
Descent is identi…ed based on the following question, which was asked to community representatives:
"Do individuals in this community trace their descent through their father, their mother, or are
both kinds of descent traced?" I de…ne a household to be patrilineal if the answer is "father",
matrilineal if the answer is "mother", and mixed if the answer is "both" and generate dummy
variables based on these answers.18
In the main analysis, I focus on consumption rather than income. This is for two reasons: …rst,
the consumption data I use are far richer than the income data available, and second, consumption
is less susceptible to shocks than income, hence painting a more accurate portrait of households’
18
Although identi…cation of descent is e¤ectively at the community level, this is unlikely to result in measurement
error because of two reasons: …rst, the most common reason for migration is marriage, and second, marriage across
descent types is extremely rare. In the IHS, approximately 50% of spouses live in the village of their birth, and
of those that have moved, 45% report marriage as the reason, while the next most common reason is moving with
parents, presumably when they were children. For these households, descent in the current village of residence is
the same as descent at birth. For the remaining households, it could be that more productive spouses move from
patrilineal to matrilineal villages. This possibility is addressed in the analysis of individual wages in Section 7.1,
which shows that patrilineal men’s predicted wages are higher than matrilineal men’s predicted wages at all points
in the distribution.
12
economic well-being (Deaton and Zaidi 2002). However, I do report results on income to verify
that they are consistent with the consumption results.19
In order to provide a balanced picture of household’s consumption levels, I consider several measures of consumption. I begin with aggregate household expenditure. This consumption aggregate
includes the value of food purchased, produced for own consumption and received as a gift, various
household items, the rental value of durables, the rental value of accommodation and expenditure
on health and schooling. I then consider per capita household expenditure and expenditure that
has been divided by an equivalent measure of household size, which gives children a lower weight
than adults. Next, I use detailed information on consumption to construct three alternative measures of consumption: consumption from purchases, public consumption and private consumption.
Consumption from purchases is interesting because it requires cash income, which is traditionally
earned by the husband working for wages. All consumption expenditure is de‡ated by a temporal
and spatial price index.
Labor supply information is provided in the form of time allocation in the previous week,
including agricultural work on household land and working for wages. To measure total working
hours, I sum all income-earning activities (agricultural work on household land, wage work, running
a business, helping with a business, and ganyu, which is low-paid agricultural work on others’land).
4.2
Summary Statistics
Table 2 shows the characteristics of the sample. We can see that matrilineal and patrilineal households are similar on basic characteristics. The average landholding size, household size and number
of children are not signi…cantly di¤erent across any of the descent types, although patrilineal household heads are signi…cantly better educated with fewer of them having no education, and are older
(p-value
0:01). There is evidence of some regional dispersion of descent, particularly relating
to the Northern region, where there are almost no matrilineal households. In the Southern and
Central regions, there is a fairly even share of patrilineal and matrilineal households. Consistent
with the anthropological literature, divorce rates are highest in matrilineal communities.20
The table also reports individual land ownership. As expected, women own signi…cantly more
land and men own signi…cantly less land, on average, in matrilineal compared to patrilineal households (p-values
0:001). Since the total amount of land owned by households is not signi…cantly
di¤erent across these descent types, this implies that women own a greater share of household land
on average in matrilineal compared to patrilineal households, while the opposite is true for men. To
show this more clearly, I calculate
from the model in Section 3.2 as the amount of land owned by
the husband divided by the total amount of land solely owned by the husband or wife (so, excluding
19
I use the terms consumption and expenditure interchangeably. Aguiar and Hurst (2005) distinguish between food
expenditure and food consumption by considering time spent on food preparation and the nutritional value of food
consumed.
20
The divorce rate is measured at the district level. It is calculated from the full IHS sample and represents the
proportion of household heads who report being separated or divorced. The …gures are consistent with those in
Reniers (2003), calculated from the 2001 Demographic and Health Survey of Malawi.
13
Table 2: Summary statistics by descent
Patrilineal
Matrilineal
Mixed
HH size
5:08
(2:14)
5:02
(1:94)
5:34
(2:18)
# Children
2:97
(1:91)
2:92
(1:74)
3:13
(1:93)
0:47
Age of head
41:36
(15:59)
39:99
(14:08)
42:29
(16:53)
0:00
% Divorced
8:23
(2:73)
12:38
(3:19)
11:05
(3:55)
0:00
Land (rainy, acres)
2:04
(1:63)
1:95
(1:80)
2:25
(2:24)
0:21
Land (dry, acres)
0:09
(0:36)
0:12
(1:30)
0:08
(0:43)
0:50
Husband’s land (all types, acres)
0:78
(1:42)
1:06
(1:58)
1:17
(2:35)
0:00
Wife’s land
0:56
(1:15)
0:32
(0:89)
0:38
(0:80)
0:00
Jointly owned land
0:34
(1:94)
0:28
(0:91)
0:35
(1:15)
0:27
Others’land
0:38
(1:05)
0:46
(1:15)
0:44
(1:13)
0:07
0:51
0:73
0:62
0:00
% Heads with no education
70:59
76:08
68:90
0:01
% Residing in North
42:07
0:10
10:75
0:00
% Residing in Centre
33:86
46:50
47:05
0:01
% Residing in South
24:08
53:40
42:20
0:00
Number of observations
2472
4448
283
6920
Implied
y
P-value
Pat=Mat
0:41
This table reports mean (standard deviation). y Reported for households that own land. P-value is for the hypothesis
that the matrilineal and patrilineal group means are equal, excluding the mixed group.
14
Table 3: Labor hours per week
Labour supply
Total income-earning labour
Agricultural
Wage
Domestic labour
Number of observations
Patrilineal
Matrilineal
Mixed
P-value
Pat=Mat
0:15
Husband
21:08
(20:26)
22:50
(19:74)
18:78
(18:55)
Wife
15:38
(16:57)
13:61
(14:10)
11:77
(13:07)
0:03
H
12:17
(14:07)
11:65
(12:54)
10:52
(13:07)
0:46
W
11:80
(13:13)
10:87
(11:71)
9:62
(11:44)
0:19
H
3:73
(13:04)
5:26
(14:69)
4:89
(14:28)
0:01
W
0:26
(2:96)
0:47
(4:35)
0:39
(3:51)
0:06
H
0:62
(2:83)
0:84
(2:98)
0:59
(2:83)
0:05
W
8:53
(7:12)
8:33
(6:91)
7:49
(7:30)
0:46
2472
4448
283
6920
This table reports mean (standard deviation). Total income-earning labour consists of agricultural labour, wage labour,
time spent running a business or helping with a business, and time spent working on others’farms. Domestic labour
consists of time spent fetching water and …rewood. For domestic labour, there is a reduced sample size of 2466 for
patrilineal women, 2465 for patrilineal men, 4435 for matrilineal men and 282 for mixed communities men. P-value is
for the hypothesis that the matrilineal and patrilineal group means are equal, excluding the mixed group.
15
Table 4: Real household consumption expenditure (’000s MWK) by descent and region
North
N
Centre
N
South
N
Patrilineal
197:16
(182:52)
1233
Matrilineal
209:43
(129:74)
9
Mixed
263:17
(340:86)
79
241:16
(162:34)
631
252:78
(195:97)
1892
331:89
(299:70)
111
140:96
(104:83)
608
198:19
(168:18)
2547
175:16
(120:31)
93
223:59
(185:20)
4448
258:36
(262:74)
283
All
198:53
(166:55)
Number of observations 2472
P-value Patrilineal = Matrilineal: 0.003
This table reports mean (standard deviation). Exchange rate: MWK 1000 = $1.39.
joint land). The table reports the average
in the sample, which again shows that men own a
greater share of land in patrilineal than matrilineal communities (p-value < 0:001). Mixed communities are, as expected, between these two values, as they have both matrilineal and patrilineal
households. This shows that land is an important determinant of patrilineal men’s outside options,
but a less important determinant of matrilineal men’s outside options.
Next, I discuss the summary statistics of labor supply, which are in Table 3.21 This table shows
the number of hours each spouse spends per week on each type of activity.22 Women tend to
work harder when they are patrilineal. When we look at the share of total labor time spent on
di¤erent activities, we …nd a pattern that is in line with the predictions of the theoretical framework:
patrilineal men spend more time on agricultural labor and less time on wage labor, on average,
than matrilineal men; the wage labor di¤erence is signi…cant at the 1% level. Women spend on
average half an hour per week on wage labor, which is less than a tenth of the time that men spend
on wage labor on average. Women and men spend similar amounts of time on agricultural labor;
the rest of women’s time is devoted to domestic work.
Finally, I discuss the expenditure of patrilineal and matrilineal households. Table 4 summarises
real aggregate household expenditure by descent and region. This table shows that matrilineal
households are better o¤ than patrilineal households in every region. In Malawi, the di¤erence in
mean expenditure between matrilineal and patrilineal households is 11% and statistically signi…cant
(p-value = 0:003).
21
Domestic labour is time spent fetching water and …rewood. The questionnaire did not ask about more typical
domestic tasks like cooking and cleaning. In addition, there is no data on leisure, which is why the total number of
hours is not equal to the number of hours in a week.
22
See Telalagic Walther (2016) for a detailed analysis of how husbands and wives in Malawi allocate their time.
16
5
Estimating Equations and Empirical Results
The fully e¢ cient collective model predicts that descent should have no direct e¤ect on the labor allocation and consumption outcomes of households. In contrast, the noncooperative model
predicts that patrilineal men should spend more time on agricultural labor and less time on wage
labor, compared to matrilineal men, and that patrilineal households should have lower consumption
overall, because of the ine¢ cient labor allocation. In this section I test the consistency of the data
with these two models.
5.1
Estimating Equations
To evaluate the e¤ect of descent on labor allocation, I estimate the following equation:
hi;c;d;r =
0
+
1 Pc;d;r
+
2 ln wi;c;d;r
+ G0i;c;d;r
3
+ X0i;c;d;r
4
+ "i;c;d;r :
(8)
I estimate this equation for …ve di¤erent dependent variables hi;c;d;r : weekly total hours of
income-earning labor, agricultural hours, wage hours, the di¤erence between agricultural and wage
hours and their sum. The coe¢ cient of interest in each case is
1:
The unit of observation is the
household i, and communities are indexed by c, districts by d and regions by r. Pk;d;r is a dummy
variable that equals one if the household’s community is patrilineal and zero if it is matrilineal,
which, as explained in the previous section, is assumed to be equal to the household’s descent.
I also include a vector of geographical controls, Gi;c;d;r . In addition to several measures of soil
quality at the household level, I also include community-level measures of rainfall, temperature
and greenness, which measures the onset and duration of spring. The intention is to control for
characteristics that a¤ect the productivity of agriculture and the suitability of the land for di¤erent
crops, and may also be correlated with descent.
The vector Xi;c;d;r consists of individual-level, community-level, district-level and region-level
covariates. At the individual level, I control for the age, age squared, gender and highest education
level of the head of household, the amount of land farmed by the household, the number of male
children, female children, male adults, female adults, male elderly and female elderly in the household, dummy variables measuring the month and year of the interview and the household’s distance
to the nearest road. Community-level variables control for the strength of the local economy and
farming patterns; they are dummy variables for the existence of a women’s group, the presence
of immigrants and the presence of wage or business labor opportunities, and the proportion of
households that farm maize, tobacco, groundnut, rice and mango in the community excluding the
respondent household. The regressions also include the district-level divorce rate and a set of region
dummy variables.
This speci…cation also controls for the logarithm of husbands’ wages, ln wi;c;d;r . If husbands
only accept wage work when the wage is high enough then the observed wage will overestimate
the underlying distribution of wage o¤ers. I estimate a Heckman selection model (Heckman 1979)
for observed wages, using descent, geographical variables such as temperature and soil quality, and
17
the number of adults and number of children in the household as instruments for participation in
wage labor. These variables satisfy the exclusion restrictions because they do not a¤ect the wage
an individual receives but do a¤ect the return to agricultural labor, which a¤ects participation
in agricultural labor and thus participation in wage labor.23 The results yield an estimate of
= 0:46
, suggesting the presence of selection bias. Using the Heckman estimates, I calculate
the predicted wage for each husband in the sample.
The baseline estimating equation for consumption outcomes is
ln Ci;c;d;r =
0
+
1 Pc;d;r
+ G0i;c;d;r
2
+ X0i;c;d;r
3
+ "i;c;d;r ;
(9)
where I use the natural logarithm of various measures of household expenditure as the dependent
variable, Ci;c;d;r . The coe¢ cient of interest is
1,
which measures the percentage di¤erence in
household expenditure between patrilineal and matrilineal households.
5.2
Key Results
The e¤ect of descent on husbands’labor supply is reported in Table 5. I cluster all standard errors
at the community level, weight the regressions based on the sampling strategy and take account of
the fact that I am using a subpopulation of the full sample.24
The results are consistent with the noncooperative model, and inconsistent with the fully e¢ cient collective model. Patrilineal men spend, on average, 1:45 hours more per week on agricultural
labor and 1:39 hours less per week on wage labor, which represents a 13:6% increase and 27:1%
reduction relative to baseline, respectively.
These di¤erences appear to be driven by substitution between wage and agricultural labor. We
cannot reject that these coe¢ cients are the same but opposite sign. This is con…rmed by the result
that the coe¢ cient on patriliny in a regression of the sum of agricultural and wage labor hours is
not signi…cantly di¤erent from zero, and the coe¢ cient in the regression of total labor hours is not
only insigni…cant but close to zero.
Sensitivity to the inclusion of controls is reported in Table 14 in Appendix A, which shows that
the inclusion of controls increases the gap in agricultural hours between matrilineal and patrilineal
households, but does not signi…cantly change the magnitude of the coe¢ cient on patriliny for wage
work hours.
The e¤ect of descent on household consumption is reported in Table 6. Regression (1) estimates
the e¤ect of patriliny on the natural logarithm of real household expenditure, which is the variable
that was summarised in Table 4.
As with labor supply, there is a signi…cant di¤erence in the consumption levels of patrilineal and
23
In a regression of log wage on covariates and these exclusion variables, the latter are not signi…cant.
I cluster at the community level because while communities are selected randomly, the variable of interest (descent)
is measured at the community level, which may induce correlation in errors between households in a given community.
With regard to the use of weights, I follow the guidance in Deaton (1997, p. 72), which suggests the use of an ‘auxiliary’
regression to test whether slope parameters vary with weights, which is the case here.
24
18
Table 5: The e¤ect of descent on labor supply
(1)
Total
(2)
(3)
Agricultural Wage
Patrilineal
0.099
(1.065)
1.453
(0.682)
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
R2
Y
Y
Y
7203
628
0.138
Y
Y
Y
7203
628
0.177
Husband’s Labour Supply
-1.385
(0.831)
(4)
Agric.
-Wage
2.839
(1.141)
(5)
Agric.
+Wage
0.068
(1.005)
Y
Y
Y
7203
628
0.184
Y
Y
Y
7203
628
0.200
Y
Y
Y
7203
628
0.153
Standard errors are clustered by community and reported in parentheses.
5% level and
denotes signi…cance at 1% level,
at
at 10% level. The geographical controls are temperature, rainfall, greenness and soil quality, which are
listed in Appendix C. The individual controls are the natural log of the husband’s predicted wage from a Heckman
selection model, acres of land farmed, the head’s age, age squared, education level and gender, the number of male
children, female children, male adults, female adults, male elderly and female elderly in the household, dummy variables
measuring the month and year of the interview, and the household’s distance to the nearest road. The community
controls are a dummy variable for the existence of a women’s group, the presence of immigrants, the presence of wage or
business labour opportunities, and the proportion of households that farm maize, tobacco, groundnut, rice and mango.
All regressions also include dummy variables for the Southern and Central regions and the district-level divorce rate.
matrilineal households, which is inconsistent with the fully e¢ cient collective model. Patrilineal
households have 9:9% lower household expenditure than matrilineal households, and this result
is highly signi…cant. This …nding is consistent with the hypothesis that patrilineal households
make less e¢ cient labor supply choices, resulting in lower overall consumption. The geographical
variables in regression (1) are jointly signi…cant with a p-value < 0:0001 and an F-statistic of 17:41.
Sensitivity to control variables is reported in Table 15 in Appendix A, which shows that the
consumption gap with only sparse household head controls is 14:5%, which decreases to 14:3% when
household-level controls are added. Adding community and geographical controls gives regression
(1) in Table 6. The movement in R2 shows that community and geographical controls explain
almost 10% of the variation in log household expenditure.
These …ndings are robust to using di¤erent measures of consumption. I estimate the e¤ect of
descent on the natural logarithm of per capita expenditure in regression (2). The coe¢ cient on
patriliny is e¤ectively unchanged from regression (1), with a statistically signi…cant 10% gap in
per capita consumption. Regression (3) shows that this gap also holds for equivalent expenditure.
Regression (4) reports the e¤ect of patriliny on consumption from purchases, which excludes consumption from own agricultural production. The di¤erence between matrilineal and patrilineal
households’consumption from purchases is 13:2%, larger than the di¤erence for all consumption.
This is interesting because the noncooperative model predicts that matrilineal husbands engage
in more wage labor, which generates cash income. The di¤erence in purchases is consistent with
this hypothesis. Finally, I split consumption into public and private expenditure in regressions (5)
19
Table 6: The e¤ect of descent on measures of consumption
(1)
Ln(exp)
Patrilineal
-0.099
(0.031)
(2)
Ln(pc
exp)
-0.100
(0.032)
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
R2
Y
Y
Y
7203
628
0.370
Y
Y
Y
7203
628
0.368
(3)
Ln(equiv
exp)
-0.097
(0.031)
(4)
(5)
(6)
Ln(purchase)Ln(public) Ln(private)
-0.132
(0.036)
-0.127
(0.031)
-0.089
(0.034)
Y
Y
Y
7203
628
0.349
Y
Y
Y
7203
628
0.322
Y
Y
Y
7203
628
0.352
Y
Y
Y
7203
628
0.325
Standard errors are clustered by community and reported in parentheses.
5% level and
denotes signi…cance at 1% level,
at
at 10% level. The geographical controls are temperature, rainfall, greenness and soil quality, which are
listed in Appendix C. The individual controls are acres of land farmed, the head’s age, age squared, education level
and gender, the number of male children, female children, male adults, female adults, male elderly and female elderly
in the household, dummy variables measuring the month and year of the interview, and the household’s distance to
the nearest road. The community controls are a dummy variable for the existence of a women’s group, the presence of
immigrants, the presence of wage or business labour opportunities, and the proportion of households that farm maize,
tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern and Central regions
and the district-level divorce rate.
and (6) respectively. Patrilineal households are worse o¤ on both measures of consumption but
particularly on public expenditure, which includes items such as durables that are typically shared
equally on divorce.
The signi…cant e¤ect of descent on labor supply allocations and household consumption is inconsistent with the fully e¢ cient collective model, which predicts that distribution factors should
not matter for the e¢ cient allocation of household production. Instead, these results are consistent with the noncooperative model, suggesting that individuals invest in their outside options to
increase future bargaining power.
5.3
Robustness to Alternative Speci…cations
To check the robustness of these results to alternative speci…cations, I re-estimate all labor supply
equations using a tobit model (see Table 16 in Appendix A). The results are consistent with those in
Table 5, with only the wage labor coe¢ cient being smaller in magnitude and insigni…cant. Instead
of using the natural logarithm of expenditure, I estimate the e¤ect of descent on the levels of each
of the consumption categories in an OLS and tobit model (Tables 17 and 18). The results are
…rmly robust to this alternative model. Next, I estimate the e¤ect of descent on income in both
an OLS regression and a tobit model (Table 19).25 I …nd that, consistent with the predictions of
25
Both income and wages are de‡ated by the same price index used to de‡ate the consumption aggregates. The
construction of the income aggregate follows Hoddinott and Haddad (1995): it includes salaries, income from crop
sales, pro…t from business and remittances from children and others.
20
the noncooperative model, the income gap between patrilineal and matrilineal households is MWK
31002 ($43:30). This is interesting, because the average income of matrilineal households is MWK
86446 ($120:70), so that this gap represents approximately 36% of average income, which is not
inconsistent with a 10% gap in consumption. I also estimate the e¤ect of descent on husbands’wage
earnings (Table 19), where I …nd that patrilineal husbands earn MWK 11881 ($16:60) less than
matrilineal husbands, which represents a 12% reduction in baseline income due to husbands’wage
earnings. This is surprisingly close to the estimated 10% gap in consumption. Notice that average
income is far lower than average consumption, showing that income in this survey is underreported
and hence less reliable than consumption in measuring household welfare.
6
Identifying Causal Relationships
The consumption gap between patrilineal and matrilineal households, as well as the evidence that
patrilineal husbands spend more time on agricultural labor and less time on wage labor than
matrilineal husbands, is inconsistent with the fully e¢ cient collective model but consistent with a
model where spouses make ine¢ cient labor supply choices due to limited commitment. However,
the empirical results could also be driven by omitted variables that are correlated with descent,
consumption and labor supply. To address this issue, I pursue two strategies in this section. First,
I control for observable characteristics that may be correlated with these variables. Second, I use
the strategy in Oster (2015) and Altonji, Elder and Taber (2005), which addresses selection on
unobserved variables using selection on observed variables.
6.1
Controlling for Observables
Due to the history of settlement in Malawi in the past, descent today is not evenly spatially
distributed. The regressions in Tables 5 and 6 already control for a number of geographical variables,
which are calculated based on the household’s GPS coordinates and provide extensive information
on distance, climatology, soil and terrain.
To show the robustness of the estimates, I control for a number of additional variables relative
to the baseline regressions of household expenditure and husband’s labor supply. In column (1) in
Table 7 and regression (1) in Table 8, I control for the GPS coordinates of the community: the
latitude, longitude, their squares and interaction.26 The inclusion of GPS coordinates does not
change the estimate of the consumption gap between patrilineal and matrilineal households, and
the e¤ect of descent on labor supply is in line with the previous results.
Geography drives land productivity, which is particularly important for labor supply and consumption allocations. To address this, I control for the land quality of plots, with plot-level responses weighted by plot size to give a household-level measure.27 I control for these variables in
26
The GPS coordinates are provided in the data and are calculated as the average of household GPS coordinates
in each community, with a random o¤set (within a pre-speci…ed range) applied to maintain con…dentiality. Latitude
and longitude are measured in decimal degrees.
27
Respondents were asked to rate the quality of each plot that they farm as good, fair or poor. They were also
21
column (2) in Table 7 and regression (2) in Table 8. Patrilineal households still have signi…cantly
lower consumption than matrilineal households.28 At …rst glance, it appears that plot quality can
explain 0:7% of the consumption gap between the two groups of households. However, this is
misleading, because the sample size is smaller when plot quality is controlled for. The correct comparison is with the coe¢ cient on patriliny in the baseline regression without plot quality restricted
to the sample of households that have plot quality information. In such a regression, the coe¢ cient
on patriliny is
0:081 (see regression (1) in Table 21). In fact, the inclusion of plot quality increases
the consumption gap by 1:1 percentage points. This suggests that patrilineal households are positively selected on plot quality. A similar pattern emerges in the labor supply results: although
the coe¢ cient on wage labor is no longer signi…cant, the size of the coe¢ cient does not change.
In column (3) in Table 7 and regression (3) in Table 8, I restrict the sample to the Southern and
Central regions. The results are again in line with the baseline estimates.
Next, I consider the role of colonialism. Colonialists may have encouraged the formation of
industry, making their settlement areas wealthier today. Alternatively, the slave trade may have
damaged trust and development (Nunn and Wantchekon 2011). To control for the residual impact
of colonial in‡uence, I construct a variable that measures the community’s distance to the closest
railway station built during colonial times.29 Column (4) in Table 7 and regression (4) in Table
8 control for this variable and show that colonialism fails to explain the consumption and labor
patterns of matrilineal and patrilineal households in the data.
There may be ethnic group characteristics that correlate with descent and consumption but
that do not a¤ect outside options, such as work ethic. To explain the results, these characteristics
would have to be true of all matrilineal ethnic groups or all patrilineal ethnic groups: for example,
all matrilineal ethnic groups are harder working than all patrilineal ethnic groups, and this is not
caused by descent. This cannot be directly tested with ethnic group …xed e¤ects as these would
be collinear with the patriliny variable. Instead, I add a series of indicator variables measuring the
most spoken language in the community. Language is highly correlated with ethnic group but not
perfectly correlated with patriliny. Column (5) in Table 7 and regression (5) in Table 8 report the
results, which are robust to these …xed e¤ects.30
Considering further the role of ethnic group traits, I control for various ethnic group characteristics from the Ethnographic Atlas (Murdock 1967). In Appendix A, column (2) in Table 20
and regression (2) in Table 21 control for the major crop type and predominant form of animal
husbandry from the Murdock Atlas, while column (3) and regression (3) control for class structure
asked to rate the plot’s erosion as no erosion, low, moderate or high, and the plot’s slope as ‡at, slight slope, moderate
slope and steep, hilly. I construct household-level measures of plot quality, plot erosion and plot slope.
28
In the regression equation for consumption, the three plot quality variables are jointly signi…cant with a p-value
< 0:0001% and an F-statistic of 14:28.
29
I use the station list on http://www.cear.mw and exclude stations built after Malawi achieved independent rule.
I calculate the distance to the closest railway station for each village. The list of stations and their GPS coordinates
are in Appendix D.
30
Another alternative explanation is that more able patrilineal men migrate to matrilineal areas. However, marriages between matrilineal and patrilineal individuals appear rare - in the LSMS, village headmen were asked about
the types of marriages observed in their village, but a mixed marriage was not listed as an option.
22
Table 7: Controlling for observables in labour supply regressions
Estimated coe¢ cient on Patrilineal
1. Agricultural
(1)
1.560
(0.701)
(2)
1.183
(0.705)
(3)
1.457
(0.692)
(4)
1.473
(0.679)
(5)
1.458
(0.687)
2. Wage
-1.529
(0.874)
-1.179
(0.814)
-1.534
(0.883)
-1.393
(0.830)
-1.341
(0.839)
3. Wage - Agric.
3.089
(1.195)
2.362
(1.160)
2.991
(1.201)
2.866
(1.136)
2.799
(1.153)
4. Wage + Agric.
0.031
(1.040)
Y
Y
Y
Y
N
N
N
7203
628
0.004
(0.988)
Y
Y
Y
N
Y
N
N
6781
624
-0.077
(1.037)
Y
Y
Y
N
N
N
N
5882
518
0.080
(1.005)
Y
Y
Y
N
N
Y
N
7203
579
0.116
(1.011)
Y
Y
Y
N
N
N
Y
7203
628
Geographical controls
Individual controls
Community controls
GPS
Plot quality
Railway distance
Tribes
Number of observations
Number of community clusters
Standard errors are clustered by community and reported in parentheses.
level and
denotes signi…cance at 1% level,
at 5%
at 10% level. This table reports the value of the coe¢ cient on the variable Patrilineal in regressions where
the dependent variables are husbands’agricultural labour (row 1), wage labour (row 2), the di¤erence between wage and
agricultural labour (row 3) and the sum of wage and agricultural labour (row 4). Compared to the baseline regressions
in Table 5, regression (1) adds GPS co-ordinates, (2) adds plot quality, plot erosion and plot slope, (3) restricts the
sample to the Southern and Central regions, (4) adds distance to nearest colonial railway station, and (5) adds language
dummy variables. (2) has a reduced sample size due to missing plot quality or colonial information for some households.
The geographical controls are temperature, rainfall, greenness and soil quality, which are listed in Appendix C. The
individual controls are the natural log of the husband’s predicted wage form a Heckman selection model, acres of land
farmed, the head’s age, age squared, education level and gender, the number of male children, female children, male
adults, female adults, male elderly and female elderly in the household, dummy variables measuring the month and
year of the interview, and the household’s distance to the nearest road. The community controls are a dummy variable
for the existence of a women’s group, the presence of immigrants, the presence of wage or business labour opportunities,
and the proportion of households that farm maize, tobacco, groundnut, rice and mango. All regressions also include
dummy variables for the Southern and Central regions and the district-level divorce rate.
23
Table 8: Controlling for observables in the consumption regression
(1)
(2)
Patrilineal
-0.098
(0.032)
-0.092
(0.030)
Geographical controls
Individual controls
Community controls
GPS
Plot quality
Railway distance
Tribes
Number of observations
Number of community clusters
R2
Y
Y
Y
Y
N
N
N
7203
628
0.374
Y
Y
Y
N
Y
N
N
6781
624
0.368
(3)
(4)
Ln(expenditure)
-0.088
-0.100
(0.033)
(0.031)
Standard errors are clustered by community and reported in parentheses.
5% level and
Y
Y
Y
N
N
N
N
5882
518
0.366
Y
Y
Y
N
N
Y
N
7203
579
0.371
(5)
-0.095
(0.031)
Y
Y
Y
N
N
N
Y
7203
628
0.373
denotes signi…cance at 1% level,
at
at 10% level. Compared to the baseline regression (1) in Table 6, regression (1) adds GPS co-ordinates,
(2) adds plot quality, plot erosion and plot slope, (3) restricts the sample to the Southern and Central regions, (4) adds
distance to nearest colonial railway station, and (5) adds language dummy variables. (2) has a reduced sample size
due to missing plot quality or colonial information for some households. The geographical controls are temperature,
rainfall, greenness and soil quality, which are listed in Appendix C. The individual controls are the natural log of
the husband’s predicted wage form a Heckman selection model, acres of land farmed, the head’s age, age squared,
education level and gender, the number of male children, female children, male adults, female adults, male elderly and
female elderly in the household, dummy variables measuring the month and year of the interview, and the household’s
distance to the nearest road. The community controls are a dummy variable for the existence of a women’s group, the
presence of immigrants, the presence of wage or business labour opportunities, and the proportion of households that
farm maize, tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern and
Central regions and the district-level divorce rate.
24
and marital structure. Controlling for these variables does not alter the coe¢ cient estimates.
Finally, in the same tables, I address the role of religion, which may be correlated with descent,
spouses’labor choices and consumption outcomes. I control for dummy variables that measure the
self-reported religion of the husband and wife in column (4) and regression (4). The coe¢ cient on
patriliny is unchanged from the baseline regressions.
6.2
Using Selection on Observables to Assess the Bias from Unobservables
Although the results are robust to controlling for observables, there may be further unobservable
factors that a¤ect selection into descent type and economic choices that, if observed, would explain
away the result. In this section, I adapt and implement the strategy proposed by Oster (2015) and
Altonji, Elder and Taber (2005) to verify the robustness of the results to selection on unobservables.
The method proceeds in three steps:
1. Record the impact of controlling for observable variables on the coe¢ cient of interest. I use
the coe¢ cients in Tables 5 and 6. Also required is a regression with a sparse set of individual
controls that are hypothesised not to be related to unobservables; I estimate regressions with
dummy variables for patriliny, mixed community, the age, age squared and education level of
the household head.
2. Quantify selection on unobservables. In the simple case with dependent variable y, one observable control multiplied by its true coe¢ cient X and a linear combination of unobservable
controls multiplied by their true coe¢ cients captured by vector Z, we write
=
Cov(y; Z) Cov(y; X)
=
:
V ar(Z)
V ar(X)
Intuitively, the bias introduced by omitting unobservables is
times the bias introduced by
omitting observables; the former is measured in step 1. In practice, I implement Oster’s
general estimator for multiple observable controls (see Oster 2015). To quantify selection, I
…nd the coe¢ cient of proportionality
that would produce a coe¢ cient of zero on patriliny.
3. Construct set estimates for the coe¢ cient of interest that are robust to selection: assuming
that the R2 in a regression with all unobservables could be at most Rmax , the coe¢ cient is
bounded on one side by the estimate with observable controls, and on the other side by a
bias-adjusted e¤ect that assumes
= 1:
I carry out these steps for the regressions on log household expenditure and husband’s labor
supply.31 Figures 2, 3 and 4 plot the set estimates of the coe¢ cient on patriliny for the regressions
on husband’s agricultural hours, husband’s wage hours, and log household expenditure, respectively,
as a function of Rmax . Oster (2015) argues that assuming Rmax < 1 is usually justi…ed, e.g. due
31
I adapt Oster’s code to survey data. The adapted Stata code and instructions can be found on the author’s
website.
25
to measurement error, and suggests that as a rule of thumb, Rmax = 1:3 Rf2 ull , where Rf2 ull is the
achieved R2 in the regression with observable controls (as in Tables 5 and 6).
The results indicate that the coe¢ cient estimates are robust. Figure 2 shows that, for the
e¤ect of patriliny on agricultural hours, the identi…ed coe¢ cient set moves away from zero as Rmax
increases. This is because the addition of controls moves the coe¢ cient on patriliny away from
zero, as seen in Table 14. The identi…ed coe¢ cient set at Rmax = 1:3 Rf2 ull is [1:453; 1:790], or an
increase of 13:6
17:2% compared to baseline. The identi…ed set for the coe¢ cient on patriliny in
the wage regression does not include zero until Rmax exceeds 0:85, and the estimated bounds on
patriliny in the wage regression with Rmax = 1:3 Rf2 ull are [ 1:385; 1:270], implying a reduction
of 25:4
27:1% in the wage hours of patrilineal husbands. The set estimate of the coe¢ cient on
expenditure in Figure 4 does not contain zero until Rmax exceeds 0.85. For Rmax = 1:3 Rf2 ull , the
estimated bounds on patriliny are [ 0:099; 0:0764], implying a consumption wedge of between
7.6% and 9.9%. Thus, in both the expenditure and wage regressions, the required Rmax for the
identi…ed set to include zero is greater than the recommended 1:3
Rf2 ull . Similar fan charts for
the regressions on Agriculture - Wage hours and Agriculture + Wage hours in Appendix B show
the robustness of these coe¢ cients.
Turning to the coe¢ cient of proportionality, a value of
> 1 implies a robust result: especially
in survey data, where questions and control variables are not chosen at random, unobservable
variables are unlikely to carry more explanatory power than observables (Altonji et al. 2005).
Note that a negative value of
implies that controlling for observables increases the value of the
coe¢ cient of interest away from zero, so that unobservables would have to be related to the variable
of interest in the opposite direction to observables. This is the case here for all outcomes except
expenditure and wage hours.
Figure 5 plots the value of , the coe¢ cient of proportionality, for di¤erent values of Rmax , for
those regressions where
is positive. The …gure shows that
exceeds one for both coe¢ cients
for most values of Rmax . Taking as a benchmark the values of Rmax = 1:3
Rf2 ull , the estimated
coe¢ cient of proportionality is 10.81 for the wage regression and 3.94 for the expenditure regression.
Note that this is for a model that already controls for a rich set of observable measures of geography,
so that selection on unobservables would have to be almost eleven times as strong as selection on
observables such as soil quality, rainfall, plot erosion and temperature to explain the e¤ect of
patriliny on wage labor hours, and almost four times as strong to explain the e¤ect of patriliny on
expenditure.
26
Coefficient on Patrilineal
6
5
4
3
2
1
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rmax
Figure 2: Estimated bounds on the coe¢ cient on patriliny in a regression of husband’s agricultural
labor supply on controls.
27
Coefficient on Patrilineal
0
−0.5
−1
−1.5
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rmax
Figure 3: Estimated bounds on the coe¢ cient on patriliny in a regression of husband’s wage labor
hours on controls.
28
0.02
Coefficient on Patrilineal
0
−0.02
−0.04
−0.06
−0.08
−0.1
−0.12
0.4
0.5
0.6
0.7
0.8
0.9
1
Rmax
Figure 4: Estimated bounds on the coe¢ cient on patriliny in a regression of log of household
expenditure on controls.
29
Expenditure
Wage
5
4
δ
3
2
1
0
0
0.2
0.4
0.6
0.8
1
Rmax
Figure 5: The estimated (ratio of selection on unobservables to selection on observables) required
to explain the coe¢ cient on patriliny as being driven by unobservables, as Rmax increases, for log
of household expenditure and wage labor hours. The lowest possible Rmax is that achieved in the
regression with all observable controls, which is why there is no estimated for expenditure for low
values of Rmax . The Rmax when = 1 is 0:854 for expenditure and 0:848 for wage hours.
30
7
Testing for Mechanisms: Wages, Land Ownership, Political Succession and Savings
The estimated negative e¤ect of patriliny on consumption and wage hours, and positive e¤ect on
agricultural hours, is inconsistent with a fully e¢ cient collective model, but consistent with an extension to the collective model in which there is limited commitment and noncooperative decision
making in labor supply choices. I provide further evidence for the mechanism in the noncooperative
model by showing that there is a larger positive wedge between the wages and agricultural productivity of patrilineal than matrilineal men, which is consistent with overinvestment in agriculture by
patrilineal men. I also show that the estimated e¤ects of descent are not observed for a placebo
group of households that own no land.
Next, I consider other mechanisms. One possibility is that matrilineal husbands have higher
productivity on the labor market, and hence higher wages, or that patrilineal husbands have higher
agricultural productivity. I address this by showing that, …rst, matrilineal husbands have lower
wages than patrilineal husbands, and, second, that the theoretical framework shows that higher
(unobserved) agricultural productivity of patrilineal households cannot explain both the labour allocation and consumption results, as it would predict higher income and consumption for patrilineal
households.
I also consider whether descent could be capturing factors other than land inheritance. I rule out
inheritance of political succession by excluding households who have a member that is involved in
local or government politics. Finally, I address the possibility that matrilineal households discount
the future at a higher rate which, combined with an assumption that wage work generates income
more immediately than agricultural work, is consistent with the empirical results. However, I show
that matrilineal households have higher savings, which is inconsistent with a higher discount rate.
7.1
The Wedge between Wages and Agricultural Productivity
Table 9 reports the mean and median values of husbands’ predicted wages from the Heckman
selection model discussed above, as well as estimates of the average agricultural product (APAL).32
Recall that equation (7) predicted a wedge between the marginal agricultural product and the
wage that was increasing in (
), suggesting that patrilineal men should have a bigger di¤erence
between the agricultural product and wage than matrilineal men. The last row of Table 9 shows
that this is indeed the case. I report the ratio of the predicted wage to the average agricultural
32
To calculate the agricultural product, I divide the sum of the estimated value of consumption from own production,
and agricultural sales revenue in the last year, by the number of hours of own-farm agricultural labour by all household
members in the last year. As labour supply information is recorded for the past week, I multiply this by 52, but this
is likely to be a¤ected by seasonality. To overcome this problem, I calculate the median agricultural hours across
households in each Traditional Authority area (TA) for each month (i.e. weekly hours multiplied by 4.3), on the
condition that there are at least four households observed in every month of the year in that TA. I then calculate
the annual agricultural hours for each TA as the total of the median hours in each month; this is used in place of the
household’s weekly hours. Where there are less than four observations for certain months of the year, I calculate the
district-TA-level median agricultural hours, again conditioning on there being at least four households observed in
every month in every group. Where this is not satis…ed, I take the region-level median monthly agricultural hours.
31
Table 9: Measures of wages and agricultural productivity
Labor type
Predicted Wage, husbands
Number of observations (N )
Return (MWK per hour)
Means
Medians
Patrilineal Matrilineal Patrilineal Matrilineal
67:11
64:00
52:21
48:33
2472
4448
2472
4448
Average Product of Agricultural Laboury
N
52:34
2386
54:07
4300
36:17
2386
39:74
4300
MASAF Wage, malesyy
N
49:05
2472
48:83
4448
50
2472
50
4448
Ratio: Predicted Wage / APAL
N
9:29
2002
4:91
3938
1:31
2002
1:18
3938
y of household member. yy Hourly rates are obtained by dividing the daily rate by four (tasks typically
take four hours; see Goldberg 2015).
product, where both the average and median ratios are higher for patrilineal than matrilineal men.33
There is also a non-zero wedge between matrilineal men’s wage and agricultural product, suggesting
ine¢ cient labor allocations among matrilineal households as well.
7.2
Land Ownership
I test whether the e¤ects of descent are present in a placebo group with no land ownership by
re-estimating the expenditure and labor supply regressions for this subsample. The theory predicts
that the estimated di¤erences between patrilineal and matrilineal households are driven by incentives to invest in land, so that no di¤erences should be observed between matrilineal and patrilineal
households that do not own land. This is what Table 10 shows. Among households who own no
land, there is no signi…cant di¤erence between patrilineal and matrilineal households in terms of
expenditure and labor allocation. The magnitudes of the coe¢ cients in regressions (1) and (2) are
close to zero, while the coe¢ cient in regression (3) is of the opposite sign to that observed in the
baseline regressions.
7.3
Di¤erences in Productivity
The collective model predicts that households in more productive regions dedicate more time to
agricultural labor, and also enjoy a higher level of overall household consumption due to their
superior productivity. In other words, in the presence of residual di¤erences in agricultural productivity, the collective model predicts that agricultural labor shares and overall consumption will be
positively correlated across descent systems. This is clearly not the case in the empirical results:
33
I assume that the average agricultural product is the same for all household members. This is a simplifying
assumption that ensures identi…cation of the agricultural product, because it is not possible to identify how much of
consumption from own production came from the labour of each individual household member.
32
Table 10: Consumption and labor supply in placebo group
(1)
Agricultural
0.240
(1.796)
Y
Y
Y
422
191
0.419
Patrilineal
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
R2
(2)
Wage
6.303
(3.997)
Y
Y
Y
422
191
0.497
(3)
Agric.-Wage
-6.062
(4.533)
Y
Y
Y
422
191
0.509
Standard errors are clustered by community and reported in parentheses.
level and
(4)
Agric.+Wage
6.543
(4.225)
Y
Y
Y
422
191
0.456
(5)
Ln(exp)
-0.018
(0.134)
Y
Y
Y
422
191
0.637
denotes signi…cance at 1% level,
at 5%
at 10% level. These regressions exclude households who do not own any land. The geographical controls
are temperature, rainfall, greenness and soil quality, which are listed in Appendix C. The individual controls are the
natural log of the husband’s predicted wage from a Heckman selection model (only in regressions (1) to (4)), acres of
land farmed, the head’s age, age squared, education level and gender, the number of male children, female children, male
adults, female adults, male elderly and female elderly in the household, dummy variables measuring the month and year
of the interview, and the household’s distance to the nearest road. The community controls are a dummy variable for
the existence of a women’s group, the presence of immigrants, the presence of wage or business labour opportunities,
and the proportion of households that farm maize, tobacco, groundnut, rice and mango. All regressions also include
dummy variables for the Southern and Central regions and the district-level divorce rate.
5
Patrilineal
Matrilineal
Ln(Predicted Wage)
4.5
4
3.5
3
1
2
3
Quintile
4
5
Figure 6: The average logarithm of the predicted wage of patrilineal and matrilineal husbands at
each quintile.
33
patrilineal households have lower consumption, although they engage in more agricultural labor.
Columns 3 and 4 in Table 9 show that the median agricultural product is lower in patrilineal than
matrilineal households. This is in line with the previous …nding that patrilineal households spend
more time on agriculture than matrilineal households, hence lowering their average product.34
Alternatively, suppose that matrilineal husbands have higher productivity in wage labor. Although the labor supply regressions control for the husband’s predicted wage, we can exclude this
mechanism more clearly by analysing the predicted wage in Table 9. The estimates show that
matrilineal husbands have similar wages to patrilineal husbands, on average, and the median patrilineal wage is higher than the median matrilineal wage. Figure 6 is a plot of the natural log of
husbands’ wages, where I report the average wage at each quintile separately for matrilineal and
patrilineal husbands. This graph shows that patrilineal husbands’average wages are higher than
matrilineal husbands’ average wages at every quintile. This directly contradicts the hypothesis
that the results can be explained by higher wages of matrilineal husbands.35 Table 9 also shows
the median hourly wage for men paid by the Malawi Social Action Fund (MASAF) public works
programme for comparison.36 The MASAF wage shows that the estimates of the predicted wage
are sensible: imputed wages are similar to an existing, observed wage that is a lower bound on what
could be achieved in the labor market.
7.4
Succession to Political O¢ ce
Adams (1999) argues that in addition to determining property inheritance, descent also determines
succession to political o¢ ce. This could be an alternative mechanism that drives consumption
and labor supply di¤erences between matrilineal and patrilineal households. To show that this
is not the case, I exclude any household with a village headman (this is the key descent-related
position in villages) and estimate the e¤ect of descent on consumption and agricultural and work
hours. The coe¢ cients on patriliny in regressions (1), (2) and (3) in Table 11 are similar to
those estimated in the baseline regressions. Note that there are only 32 households with a village
headman, so that only a small proportion of the sample could potentially inherit local political
o¢ ce. To consider government roles more widely, I exclude all households with a member who
works for the government in regressions (4), (5) and (6). Although the coe¢ cient on wage hours
loses signi…cance, the magnitudes of the coe¢ cients are unchanged. Succession to political o¢ ce
cannot explain the e¤ect of descent on consumption and labor choices.
34
Assuming that agricultural production exhibits diminishing marginal returns and that patrilineal and matrilineal
households can be described by the same production function, then the average agricultural product can be used
instead of the marginal agricultural product to compare this wedge across the two descent types. In this case the
wedge will be a lower bound on the true wedge.
35
This is also true when comparing observed wages, which are higher for patrilineal than matrilineal men.
36
The MASAF programme o¤ers a social safety net. It operates di¤erently to employment guarantee schemes
elsewhere, such as the National Rural Employment Guarantee of India, which sets wages above the market clearing
wage.
34
Table 11: Controlling for succession to political o¢ ce
Patrilineal
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
R2
HHs without village
(1)
(2)
Agric.
Wage
1.483
-1.386
(0.685)
(0.827)
Y
Y
Y
Y
Y
Y
7178
7178
628
628
0.178
0.186
headman
(3)
Ln(exp)
-0.098
(0.031)
Y
Y
Y
7178
628
0.371
Standard errors are clustered by community and reported in parentheses.
level and
HHs without government workers
(4)
(5)
(6)
Agric.
Wage
Ln(exp)
1.450
-1.317
-0.100
(0.682)
(0.813)
(0.032)
Y
Y
Y
Y
Y
Y
Y
Y
Y
7136
7136
7136
628
628
628
0.177
0.185
0.367
denotes signi…cance at 1% level,
at 5%
at 10% level. The geographical controls are temperature, rainfall, greenness and soil quality, which are listed
in Appendix C. The individual controls are the natural log of the husband’s predicted wage from a Heckman selection
model (only in regressions (1), (2), (5) and (6)), acres of land farmed, the head’s age, age squared, education level and
gender, the number of male children, female children, male adults, female adults, male elderly and female elderly in
the household, dummy variables measuring the month and year of the interview, and the household’s distance to the
nearest road. The community controls are a dummy variable for the existence of a women’s group, the presence of
immigrants, the presence of wage or business labour opportunities, and the proportion of households that farm maize,
tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern and Central regions
and the district-level divorce rate.
7.5
Savings
The results could be consistent with a discount factor argument, where matrilineal households are
more impatient and so spend more today at the expense of saving for tomorrow, and allocate their
time towards more immediate income-earning work. This relies on the assumption that wage work
generates income more quickly than agricultural work. Given the absence of information on cash
savings in the data, an alternative way to measure savings is through the number of livestock,
which 67% of households have. Table 12 shows that patrilineal households own signi…cantly fewer
livestock. The survey also asked the household head about their subjective assessment of whether
the household income is su¢ cient for building household savings. Their answers are reported
in Table 13. Matrilineal household heads are 8% more likely to say they are able to build their
savings, and report a higher average answer to this question, which corresponds to a better …nancial
situation. Both of these results contradict a story of higher discount factors among matrilineal
households.
35
Table 12: Savings: livestock
(1)
Number of livestock
-0.566
(0.334)
Patrilineal
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
R2
Y
Y
Y
7203
628
0.169
Standard errors are clustered by community and reported in parentheses.
denotes signi…cance at 1% level,
at 5% level and
at 10% level. The
geographical controls are temperature, rainfall, greenness and soil quality,
which are listed in Appendix C. The individual controls are the acres of land
farmed, the head’s age, age squared, education level and gender, the number
of male children, female children, male adults, female adults, male elderly
and female elderly in the household, dummy variables measuring the month
and year of the interview, and the household’s distance to the nearest road.
The community controls are a dummy variable for the existence of a women’s
group, the presence of immigrants, the presence of wage or business labour
opportunities, and the proportion of households that farm maize, tobacco,
groundnut, rice and mango. All regressions also include dummy variables
for the Southern and Central regions and the district-level divorce rate.
Table 13: Ability to build savings
Patrilineal
0.136
(0.352)
Matrilineal
0.218
(0.395)
Di¤erence
0.082
How able?y (1-5)
2.429
(1.175)
2.692
(1.189)
0.262
N
2471
4448
Able to build savings? 1 = Yes, 0 = No
This table reports mean (standard deviation). y The possible answers to this question are: 1 = Your current
income is not su¢ cient, so you have to borrow to meet expenses; 2 = Your current income is not su¢ cient, so you
have to use savings to meet expenses; 3 = Your current income only just meets your expenses; 4 = Your current
income allows you to save just a little; 5 = Your current income allows you to build your savings.
36
8
Conclusion
In this paper, I propose and implement a new test of production e¢ ciency in the household, that
also tests whether ine¢ ciency is driven by limited commitment. I extend the collective model
to allow a noncooperative …rst stage, which results from limited commitment in labor allocation.
Labor supply a¤ects future bargaining power because it determines the value of future outside
options. Even if sharing of consumption in the second period is e¢ cient, labor allocations are no
longer e¢ cient. Instead, individuals spend more time on the type of labor that most improves their
outside option. This also results in lower overall consumption available for sharing in the second
period.
I test for production e¢ ciency in the household by exploiting predetermined variation in descent
in Malawi, which governs land rights. I show that labor supply and consumption depend on
descent, which is inconsistent with the fully e¢ cient collective model after any di¤erences in labor
productivity are taken into account. Patrilineal men spend more time on agricultural labor and
less time on wage labor than matrilineal men. The outside options of patrilineal men are increasing
in the value of their land, so that this …nding is consistent with the limited commitment collective
model, where individuals invest in their outside options to raise future bargaining power. Patrilineal
households also have lower consumption and income than matrilineal households, which is consistent
with an ine¢ cient labor allocation. The empirical results are robust to a wide variety of checks,
and together these …ndings point towards a mechanism driven by ine¢ cient investment in outside
options.
The results in this paper suggest that there may be ine¢ ciencies in how resources are allocated
across time within households. This relates to the literature on limited commitment and dynamic
models of the household (Mazzocco 2007, Robinson 2012, Dubois and Ligon 2011). Although static
consumption e¢ ciency is usually not rejected in the data, there is growing evidence that e¢ ciency
should be reconsidered when looking at dynamic decisions.
This paper rejects production e¢ ciency in the context of the collective model, and so is coherent
with similar results in other papers, mostly in developing countries. As consumption e¢ ciency is
often not rejected in the literature, this suggests that there may be more serious constraints to
achieving production e¢ ciency, and that these two measures of e¢ ciency should be considered
separately, as in Rangel and Thomas (2012).
More generally, this paper demonstrates that the size of the household ‘pie’may not be invariant
to spouses’outside options. When spouses make noncooperative decisions, sel…sh choices can have a
negative impact on household welfare. This suggests that there may be a role for binding contracts
within marriage that condition on labor supply. This is resonant with a recent literature on the
e¢ ciency gains of prenuptial contracts (Voena 2015, Bayot and Voena 2015).
37
References
[1] Adams, B. N. "Cross-Cultural and U.S. Kinship." Handbook of Marriage and the Family. Ed.
M. B. Sussman, S. K. Steinmetz and G. W. Peterson. 2nd ed. New York: Plenum Press, 1999.
pp. 77-92.
[2] Altonji, J. G., T. E. Elder and C. R. Taber. "Selection on Observed and Unobserved Variables:
Assessing the E¤ectiveness of Catholic Schools." Journal of Political Economy 113.1 (2005):
pp. 151-184.
[3] Aguiar, M. and E. Hurst. "Consumption versus Expenditure." Journal of Political Economy
113.5 (2005): pp. 919-948.
[4] Apps, P. F. and R. Rees. "Collective Labour Supply and Household Production." Journal of
Political Economy 105.1 (1997): pp. 178-190.
[5] Ashraf, N. "Spousal Control and Intra-Household Decision Making: An Experimental Study
in the Philippines." American Economic Review 99.4 (2009): pp.1245–1277.
[6] Basu, K. "Gender and Say: A Model of Household Behaviour with Endogenously Determined
Balance of Power." The Economic Journal 116 (April, 2006): pp. 558-580.
[7] Bayot, D. and A. Voena. "Prenuptial Contracts, Labor Supply and Household Investments."
Mimeo (2015).
[8] Becker, Gary S. "A Theory of Marriage: Part I." Journal of Political Economy 81.4 (1973):
pp. 813-846.
[9] Becker, Gary S. "A Theory of Marriage: Part II." Journal of Political Economy 82.2 Part 2:
pp. S11-S26.
[10] Berge, E., D. Kambewa, A. Munthali and H. Wiig. "Lineage and Land Reforms in Malawi:
Do Matrilineal and Patrilineal Landholding Systems Represent a Problem for Land Reforms
in Malawi?" Land Use Policy 41 (2014): pp. 61-69.
[11] Besley, T. "Property Rights and Investment Incentives: Theory and Evidence from Ghana."
Journal of Political Economy 103.5 (1995): pp. 903-937.
[12] Bobonis, G. J. "Is the Allocation of Resources within the Household E¢ cient? New Evidence
from a Randomised Experiment." Journal of Political Economy 117.3 (2009): pp. 453-503.
[13] Browning, M., F. Bourguignon, P.-A. Chiappori and V. Lechène. “Incomes and Outcomes: A
Structural Model and Some Evidence from French Data.”Journal of Political Economy 102.6
(1994): pp.1067-96.
38
[14] Browning, M. and P.-A. Chiappori. "E¢ cient Intra-Household Allocation: A General Characterisation and Empirical Tests." Econometrica 66.6 (1998): pp. 1241-1278.
[15] Brune, L., X. Giné, J. Goldberg and D. Yang. "Facilitating Savings For Agriculture: Field Experimental Evidence from Malawi." Economic Development and Cultural Change 64.2 (2016):
pp. 187-220.
[16] Browning, M., P.-A. Chiappori and Y. Weiss. "Economics of the Family." Cambridge: Cambridge University Press (2014).
[17] Chen, Z. and F. Woolley. "A Cournot-Nash Model of Family Decision Making." Economic
Journal 111 (2001): pp. 722-748.
[18] Cherchye, L., S. Cosaert, T. Demuynck and B. De Rock. "Noncooperative Household Consumption with Caring." Mimeo (2016).
[19] Chiappori, P.-A. "Rational Household Labour Supply." Econometrica 56.1 (1988): pp. 63-89.
[20] Chiappori, P.-A. "Collective Labour Supply and Welfare." Journal of Political Economy 100
(1992): pp. 437-467.
[21] Chiappori, P.-A. "Introducing Household Production in Collective Models of Labor Supply."
Journal of Political Economy 105.1 (1997): pp. 191-209.
[22] Chiappori, P.-A., B. Fortin and G. Lacroix. "Marriage Market, Divorce Legislation, and Household Labor Supply." Journal of Political Economy 110.1 (2002): pp. 37-72.
[23] Davison, J. Gender, Lineage and Ethnicity in Southern Africa. Oxford: Westview Press, 1997.
[24] Deaton, A. The Analysis of Household Surveys: A Microeconometric Approach to Development
Policy. Baltimore, Maryland: Johns Hopkins University Press, 1997.
[25] Deaton, A. and S. Zaidi. "Guidelines for Constructing Consumption Aggregates for Welfare
Analysis." LSMS Working Paper 135 (2002).
[26] Dercon, S. and P. Krishnan. "In Sickness and in Health? Risk Sharing Within Households in
Ethiopia." Journal of Political Economy 108.4 (2000): pp. 688-727.
[27] Doss, C. R. "Is Risk Fully Pooled within the Household? Evidence from Ghana." Economic
Development and Cultural Change 50.1 (2001): pp. 101-130.
[28] Dubois, P. and E. Ligon. "Nutrition and Risk Sharing within the Household." Mimeo (2011).
[29] Du‡o, E. and C. Udry. "Intrahousehold Resource Allocation in Cote d’Ivoire: Social Norms,
Separate Accounts and Consumption Choices." NBER Working Paper 10498 (2004).
39
[30] Food and Agriculture Organization of the United Nations. "Gender Inequalities in Rural Employment in Malawi: An Overview." Mimeo (2011).
[31] Goldberg, J. "Kwacha Gonna Do? Experimental Evidence about Labour Supply in Rural
Malawi." American Economic Journal: Applied Economics 8.1 (2016): pp. 129-149.
[32] Heckman, J. "Sample Selection Bias as a Speci…cation Error." Econometrica 47 (1979). pp.
153-61.
[33] Hirschmann, D. and M. Vaughan. "Food Production and Income Generation in a Matrilineal
Society: Rural Women in Zomba, Malawi." Journal of Southern African Studies 10.1 (1983):
pp. 86-99.
[34] Hoddinott, J. and L. Haddad. "Does Female Income Share In‡uence Household Expenditures?
Evidence from Côte D’Ivoire." Oxford Bulletin of Economics and Statistics 57.1 (1995): pp.
77-96.
[35] Iyigun, M. and R. P. Walsh. "Endogenous Gender Power, Household Labor Supply and the
Demographic Transition." Journal of Development Economics 82 (2007): pp. 138-155.
[36] Jakiela, P. and O. Ozier. "Does Africa need a Rotten Kid Theorem? Experimental Evidence
from Village Economies." Review of Economic Studies 83.1 (2016): pp. 231-268.
[37] Johnson, M. M. Strong Mothers, Weak Wives: The Search for Gender Equality. Berkeley:
University of California Press, 1988.
[38] Jones, C. "The Mobilization of Women’s Labor for Cash Crop Production: A Game Theoretic
Approach." American Journal of Agricultural Economics 65.5 (1983): pp. 1049-1054.
[39] Kerr, R. B.. "Food Security in Northern Malawi: Gender, Kinship Relations and Entitlements
in Historical Context." Journal of Southern African Studies 31.1 (2005): pp. 53-74.
[40] Kishindo, P. "The Marital Immigrant. Land and Agriculture: A Malawian Case Study." African
Sociological Review 14.2 (2010): pp. 89-97.
[41] La Ferrara, E. "Descent Rules and Strategic Transfers. Evidence from Matrilineal Groups in
Ghana." Journal of Development Economics 83.2 (2007): pp. 280-301.
[42] Lamphere, L. "Strategies, Cooperation and Con‡ict Among Women in Domestic Groups"
in Woman, Culture and Society, eds. M. Z. Rosaldo and L. Lamphere. Stanford, California:
Stanford University Press (1974): pp. 97-112.
[43] Lechene, V. and I. Preston. "Noncooperative Household Demand." Journal of Economic Theory 146.2 (2011): pp. 504-527.
[44] Lovo, S. "Tenure insecurity and investment in soil conservation. Evidence from Malawi." World
Development 78 (2016): pp. 219-229.
40
[45] Lundberg, S. J. and R. A. Pollak. "Separate Spheres Bargaining and the Marriage Market."
Journal of Political Economy 101.6 (1993): pp. 988-1010.
[46] Lundberg, S.J. and R. A. Pollak. "E¢ ciency in Marriage." Review of Economics of the Household 1 (2003): pp. 153-167.
[47] Lundberg, S.J., R. A. Pollak and T. J. Wales. "Do Husbands and Wives Pool their Resources?
Evidence from the United Kingdom Child Bene…t." Journal of Human Resources 32.3 (1997):
pp. 463-480.
[48] Mazzocco, M. "Household Intertemporal Behavior: A Collective Characterization and a Test
of Commitment." Review of Economic Studies 74.3 (2007): pp. 857-895.
[49] McPeak, J. G. and C. R. Doss. "Are Household Production Decisions Cooperative? Evidence
on Pastoral Migration and Milk Sales from Northern Kenya" American Journal of Agricultural
Economics 88.3 (2006): pp. 525-541.
[50] Murdock, G. P. Ethnographic Atlas. Pittsburgh: University of Pittsburgh Press, 1967.
[51] Mwambene, L. "Divorce in Matrilineal Customary Law Marriage in Malawi: A Comparative Analysis with the Patrilineal Customary Law Marriage in South Africa." LLM Thesis,
University of the Western Cape (2005).
[52] Nunn, N. and L. Wantchekon. "The Slave Trade and the Origins of Mistrust in Africa." The
American Economic Review 101.7 (2011): pp. 3221-3252.
[53] Oster, E. "Unobservable Selection and Coe¢ cient Stability: Theory and Evidence." Mimeo
(2015).
[54] Peters, P. E. "Against the Odds. Matriliny, Land and Gender in the Shire Highlands of
Malawi." Critique of Anthropology 17 (1997): pp. 189-210.
[55] Peters, P. E. "Bewitching Land: The Role of Land Disputes in Converting Kin to Strangers
and in Class Formation in Malawi." Journal of Southern African Studies 28.1 (2002): pp.
155-178.
[56] Peters, P. E. "‘Our daughters inherit our land, but our sons use their wives’…elds’: matrilinealmatrilocal land tenure and the New Land Policy in Malawi." Journal of Eastern African Studies
4.1 (2010): pp. 179-199.
[57] Place, F. and K. Otsuka. "Tenure, Agricultural Investment, and Productivity in the Customary
Tenure Sector of Malawi." Economic Development and Cultural Change 50.1 (2001): pp. 77-99.
[58] Rainer, H. "Should We Write Prenuptial Contracts?" European Economic Review 51.2 (2007):
pp. 337-363.
41
[59] Rangel, M. A. and D. Thomas. "Gender, Production and Consumption: Allocative E¢ ciency
within Farm Households." Mimeo (2012).
[60] Reniers, G. "Divorce and Remarriage in Rural Malawi." Demographic Research Special Collection 1.6 (2003).
[61] Robinson, J. "Limited Insurance within the Household: Evidence from a Field Experiment in
Kenya." American Economic Journal: Applied Economics 4.4 (2012): pp. 140-64.
[62] Schatz, E. "Numbers and Narratives: Making Sense of Gender and Context in Rural Malawi."
University of Pennsylvania PhD Dissertation (2002).
[63] Spring, A. Agricultural Development and Gender Issues in Malawi. Maryland: University Press
of America, Inc., 1995.
[64] Telalagić Walther, S. "Moral Hazard in Marriage: The Use of Domestic Labour as an Incentive
Device." (2016). Mimeo.
[65] Thomas, D. "Intra-Household Resource Allocation: An Inferential Approach." Journal of Human Resources 25.4 (1990): pp. 635-664.
[66] Tyler, E. B. "On a Method of Investigating the Development of Institutions; Applied to Laws
of Marriage and Descent." The Journal of the Anthropological Institute of Great Britain and
Ireland 18 (1889): pp. 245-272.
[67] Udry, C. "Gender, Agricultural Production and the Theory of the Household." Journal of
Political Economy 104.5 (1996): pp. 1010-46.
[68] Udry, C. and M. Goldstein. "The Pro…ts of Power: Land Rights and Agricultural Investment
in Ghana." Journal of Political Economy 116.6 (2008): pp. 981-1022.
[69] Voena, A. "Yours, Mine and Ours: Do Divorce Laws A¤ect the Intertemporal Behavior of
Married Couples?" American Economic Review 105.8 (2015): pp. 2295-2332.
[70] World Bank Development Economics Research Group and Malawi National Statistical O¢ ce.
Malawi Living Standards Measurement Study: 2010-2011 IHS3 Survey (2012).
42
A
Additional results tables
Table 14: Sensitivity to controls: Labor supply
Estimated coe¢ cient on Patrilineal
(1)
(2)
1. Agricultural
0.480
0.863
(0.706)
(0.614)
-1.544
-1.476
(0.565)
(0.547)
2.024
2.339
(0.971)
(0.890)
-1.064
-0.612
(0.832)
(0.747)
Head controls
Y
Y
Household controls
N
Y
Community controls
N
N
Geographical controls
N
N
Number of observations
7203
7203
Number of community clusters
628
628
2. Wage
3. Wage - Agric.
4. Wage + Agric.
Standard errors are clustered by community and reported in parentheses.
notes signi…cance at 1% level,
at 5% level and
de-
at 10% level. This table reports
the value of the coe¢ cient on the variable Patrilineal in regressions where the dependent variables are husbands’ agricultural labour (row 1), wage labour (row 2),
the di¤erence between wage and agricultural labour (row 3) and the sum of wage
and agricultural labour (row 4). The head controls are the household head’s age,
age squared, education level, gender and log of the predicted wage from a Heckman
selection model. The individual controls are acres of land farmed, the number of
male children, female children, male adults, female adults, male elderly and female
elderly in the household, and dummy variables measuring the month and year of
the interview. The community controls are a dummy variable for the existence of a
women’s group, the presence of immigrants, the presence of wage or business labour
opportunities, the district-level divorce rate, the distance to the nearest road and
the proportion of households that farm maize, tobacco, groundnut, rice and mango.
The geographical controls are temperature, rainfall, greenness and soil quality, which
are listed in Appendix C, as well as dummy variables for the Southern and Central
regions.
43
Table 15: Sensitivity to controls: Consumption
(1)
Patrilineal
Head controls
Household controls
Community controls
Geographical controls
Number of observations
Number of community clusters
R2
(2)
Ln (real expenditure)
-0.145
-0.143
(0.034)
(0.032)
Y
Y
N
Y
N
N
N
N
7203
7203
628
628
0.145
0.274
Standard errors are clustered by community and reported in parentheses.
denotes signi…cance at 1% level,
at 5% level and
at 10% level. The head
controls are the household head’s age, age squared, education level and gender.
The individual controls are acres of land farmed, the number of male children,
female children, male adults, female adults, male elderly and female elderly in the
household, and dummy variables measuring the month and year of the interview.
The community controls are a dummy variable for the existence of a women’s group,
the presence of immigrants, the presence of wage or business labour opportunities,
the district-level divorce rate, the distance to the nearest road and the proportion of
households that farm maize, tobacco, groundnut, rice and mango. The geographical
controls are temperature, rainfall, greenness and soil quality, which are listed in
Appendix C, as well as dummy variables for the Southern and Central regions.
Table 16: Tobit model of labor supply
Patrilineal
Geographical controls
Individual controls
Community controls
Number of observations
Number of community clusters
(1)
Total
-0.052
(0.966)
Y
Y
Y
7203
628
(2)
Agricultural
1.529
(0.526)
Y
Y
Y
7203
628
(3)
Wage
-0.489
(0.477)
Y
Y
Y
7203
628
(4)
Agric. - Wage
1.656
(0.593)
Y
Y
Y
7203
628
(5)
Agric. + Wage
-0.076
(0.896)
Y
Y
Y
7203
628
This table reports the marginal e¤ects at means from a Tobit model. Standard errors are clustered by community
and reported in parentheses.
denotes signi…cance at 1% level,
at 5% level and
at 10% level. The geographical
controls are temperature, rainfall, greenness and soil quality, which are listed in Appendix C. The individual controls
are the natural log of the husband’s predicted wage from a Heckman selection model, acres of land farmed, the head’s
age, age squared, education level and gender, the number of male children, female children, male adults, female
adults, male elderly and female elderly in the household, dummy variables measuring the month and year of the
interview, and the household’s distance to the nearest road. The community controls are a dummy variable for the
existence of a women’s group, the presence of immigrants, the presence of wage or business labour opportunities,
and the proportion of households that farm maize, tobacco, groundnut, rice and mango. All regressions also include
dummy variables for the Southern and Central regions and the district-level divorce rate.
44
Table 17: Consumption in levels (OLS)
Patrilineal
Geographical controls
Individual controls
Community controls
Number of community clusters
Number of observations
R2
(1)
Exp.
(2)
Pc exp.
-24.657
(9.324)
Y
Y
Y
628
7203
0.318
-4.670
(1.893)
Y
Y
Y
628
7203
0.310
(3)
Equiv.
exp.
-4.923
(2.109)
Y
Y
Y
628
7203
0.304
Standard errors are clustered by community and reported in parentheses.
level and
(4)
Purchases
-26.918
(8.879)
Y
Y
Y
628
7203
0.286
(5)
Public
exp.
-9.779
(3.634)
Y
Y
Y
628
7203
0.239
denotes signi…cance at 1% level,
(6)
Private
exp.
-14.877
(6.529)
Y
Y
Y
628
7203
0.296
at 5%
at 10% level. The dependent variables are measured in local currency (’000s Malawi Kwacha). The geographical
controls are temperature, rainfall, greenness and soil quality, which are listed in Appendix C. The individual controls are the
acres of land farmed, the head’s age, age squared, education level and gender, the number of male children, female children,
male adults, female adults, male elderly and female elderly in the household, dummy variables measuring the month and
year of the interview, and the household’s distance to the nearest road. The community controls are a dummy variable for
the existence of a women’s group, the presence of immigrants, the presence of wage or business labour opportunities, and
the proportion of households that farm maize, tobacco, groundnut, rice and mango. All regressions also include dummy
variables for the Southern and Central regions and the district-level divorce rate.
Table 18: Consumption in levels (Tobit)
Patrilineal
Geographical controls
Individual controls
Community controls
Number of community clusters
Number of observations
(1)
Exp.
(2)
Pc exp.
-22.675
(8.435)
Y
Y
Y
628
7203
-4.353
(1.751)
Y
Y
Y
628
7203
(3)
Equiv.
exp.
-4.623
(1.970)
Y
Y
Y
628
7203
(4)
Purchases
-22.986
(7.356)
Y
Y
Y
628
7203
(5)
Public
exp.
-7.732
(2.780)
Y
Y
Y
628
7203
(6)
Private
exp.
-13.827
(6.039)
Y
Y
Y
628
7203
This table reports marginal e¤ects at means from a Tobit model. Standard errors are clustered by community and reported
in parentheses.
denotes signi…cance at 1% level,
at 5% level and
at 10% level. The dependent variables are measured
in local currency (’000s Malawi Kwacha). The geographical controls are temperature, rainfall, greenness and soil quality,
which are listed in Appendix C. The individual controls are the acres of land farmed, the head’s age, age squared, education
level and gender, the number of male children, female children, male adults, female adults, male elderly and female elderly in
the household, dummy variables measuring the month and year of the interview, and the household’s distance to the nearest
road. The community controls are a dummy variable for the existence of a women’s group, the presence of immigrants, the
presence of wage or business labour opportunities, and the proportion of households that farm maize, tobacco, groundnut,
rice and mango. All regressions also include dummy variables for the Southern and Central regions and the district-level
divorce rate.
45
Table 19: Income and husband’s earnings in levels
OLS
(1)
Income
Patrilineal
Geographical controls
Individual controls
Community controls
Number of community clusters
Number of observations
R2
(2)
Husband’s
earnings
-11.881
(4.038)
Y
Y
Y
628
7203
0.341
-31.002
(12.643)
Y
Y
Y
628
7203
0.146
Standard errors are clustered by community and reported in parentheses.
level and
Tobit model
(3)
Income
-18.313
(7.595)
Y
Y
Y
628
7203
N/A
(4)
Husband’s
earnings
-4.203
(2.394)
Y
Y
Y
628
7203
N/A
denotes signi…cance at 1% level,
at 5%
at 10% level. The dependent variables are measured in local currency (’000s Malawi Kwacha). Regressions
(1) and (2) estimate Ordinary Least Squares, while regressions (3) and (4) report the marginal e¤ects at means of
a Tobit model. The geographical controls are temperature, rainfall, greenness and soil quality, which are listed in
Appendix C. The individual controls are the acres of land farmed, the head’s age, age squared, education level and
gender, the number of male children, female children, male adults, female adults, male elderly and female elderly in
the household, dummy variables measuring the month and year of the interview, and the household’s distance to the
nearest road. The community controls are a dummy variable for the existence of a women’s group, the presence of
immigrants, the presence of wage or business labour opportunities, and the proportion of households that farm maize,
tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern and Central regions
and the district-level divorce rate.
46
Table 20: Controlling for observables: Labor supply
Estimated coe¢ cient on Patrilineal
1. Agricultural
(1)
1.176
(0.707)
(2)
1.796
(0.780)
(3)
1.727
(0.762)
(4)
1.724
(0.676)
2. Wage
-1.072
(0.808)
-1.385
(0.831)
-1.373
(0.832)
-1.452
(0.822)
3. Wage - Agric.
2.248
(1.159)
3.182
(1.242)
3.100
(1.225)
3.176
(1.118)
4. Wage + Agric.
0.104
(0.981)
Y
Y
Y
624
6781
0.411
(1.026)
Y
Y
Y
607
5946
0.353
(1.022)
Y
Y
Y
607
5946
0.272
(1.007)
Y
Y
Y
628
7203
Geographical controls
Community controls
Individual controls
Number of community clusters
Number of observations
Standard errors are clustered by community and reported in parentheses.
level and
denotes signi…cance at 1% level,
at 5%
at 10% level. This table reports the value of the coe¢ cient on the variable Patrilineal in regressions where
the dependent variables are husbands’agricultural labour (row 1), wage labour (row 2), the di¤erence between wage and
agricultural labour (row 3) and the sum of wage and agricultural labour (row 4). Regression (1) repeats the estimation
equations in Appendix C for the subsample with information on plot quality. Using dummy variables, regression (2)
controls for the major crop type (v29) and predominant form of animal husbandry (v40) from the Murdoch atlas.
Regression (3) controls for the class structure (v66) and marital structure (v9) from the Murdoch atlas. Regression
(4) controls for the predominant religion practised in the village. The geographical controls are temperature, rainfall,
greenness and soil quality, which are listed in Appendix 22. The individual controls are the log of the husband’s
predicted wage from a Heckman selection model, acres of land farmed, the head’s age, age squared, education level
and gender, the number of male children, female children, male adults, female adults, male elderly and female elderly
in the household, dummy variables measuring the month and year of the interview, and the household’s distance to
the nearest road. The community controls are a dummy variable for the existence of a women’s group, the presence of
immigrants, the presence of wage or business labour opportunities, and the proportion of households that farm maize,
tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern and Central regions
and the district-level divorce rate.
47
Table 21: Controlling for observables: Consumption
(1)
Patrilineal
Geographical controls
Individual controls
Community controls
Number of community clusters
Number of observations
R2
(2)
(3)
Ln (real expenditure)
-0.089
-0.090
(0.034)
(0.034)
Y
Y
Y
Y
Y
Y
607
607
5946
5946
0.355
0.355
-0.081
(0.030)
Y
Y
Y
624
6781
0.364
Standard errors are clustered by community and reported in parentheses.
at 5% level and
(4)
-0.107
(0.032)
Y
Y
Y
628
7203
0.372
denotes signi…cance at 1% level,
at 10% level. Regression (1) repeats the estimation equation (1) in Table 6 for the subsample
with information on plot quality. Using dummy variables, regression (2) controls for the major crop type (v29)
and predominant form of animal husbandry (v40) from the Murdoch atlas. Regression (3) controls for the class
structure (v66) and marital structure (v9) from the Murdoch atlas. Regression (4) controls for the predominant
religion practised in the village. The geographical controls are temperature, rainfall, greenness and soil quality, which
are listed in Appendix C. The individual controls are the acres of land farmed, the head’s age, age squared, education
level and gender, the number of male children, female children, male adults, female adults, male elderly and female
elderly in the household, dummy variables measuring the month and year of the interview, and the household’s
distance to the nearest road. The community controls are a dummy variable for the existence of a women’s group,
the presence of immigrants, the presence of wage or business labour opportunities, and the proportion of households
that farm maize, tobacco, groundnut, rice and mango. All regressions also include dummy variables for the Southern
and Central regions and the district-level divorce rate.
48
B
Figures of coe¢ cient bounds
8
Coefficient on Patrilineal
7
6
5
4
3
2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rmax
Figure 7: Estimated bounds on the coe¢ cient on patriliny in a regression of husband’s agricultural
- wage labor hours on controls.
49
8
Coefficient on Patrilineal
6
4
2
0
0.2
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
Rmax
Figure 8: Estimated bounds on the coe¢ cient on patriliny in a regression of husband’s agricultural
+ wage labor hours on controls.
C
Geographical variables
Table 22 lists the geographical variables included in the regressions.
50
Table 22: Geographical variables
Category
Temperature
Variable type
Continuous
Level
Community
Reference period
1960-1990
Temperature
Continuous
Community
1960-1990
Temperature
Continuous
Community
1960-1990
Temperature
Continuous
Community
1960-1990
Rainfall
Continuous
Community
Rainfall
Continuous
Community
Rainfall
Continuous
Community
2008-2009, 20092010
2008-2009, 20092010
2008-2009, 20092010
Greenness
Continuous
District
2008-2009, 20092010
Greenness
Continuous
District
Greenness
Continuous
District
Soil quality
Indicator
Household
2008-2009, 20092010
2008-2009, 20092010
N/A
Soil quality
Indicator
Household
N/A
Soil quality
Indicator
Household
N/A
51
Description
Average daily range: mean of
max. temp.- min. temp.
Temperature
seasonality:
standard deviation of monthly
climatology
Minimum temperature of coldest month
Average
temperature
of
wettest quarter
Average 12-month total rainfall
Average total rainfall in
wettest quarter
Average start of wettest quarter in dekads, from July onwards
Total change in greenness
within the primary growing
season
Onset of greenness increase in
day of year, starting July 1st
Onset of greenness decrease in
day of year, starting July 1st
Nutrient availability: 7 categories de…ning extent of constraint
Rooting conditions: 7 categories de…ning extent of constraint
Excess salts:
7 categories
de…ning extent of constraint
D
GPS Coordinates of pre-Independence Railway Stations
Railway station
Latitude
Longitude
Limbe
-15.8084
35.0574
Malabvi
-15.8403
35.1316
Nansadi
-15.8833
35.2
Makandi
-15.9333
35.2333
Luchenza
-16.0101
35.3082
Khonjeni
-16.1233
35.2423
Makapwa
-16.3333
35.2833
Sandama
-16.2073
35.2952
Chipho
-16.2893
35.2952
Thekerani
-16.3424
35.0994
Thukuta
-16.0686
34.8666
Sankhulani
-16.2325
35.2847
Osiyana
-16.4790
35.1885
Makhanga
-15.5408
35.7939
Blantyre
-15.7879
35.0159
South Lunzu
-15.6527
35.0195
Chilaweni
-15.2768
35.7189
Maleule
-15.6445
35.0567
Lirangwe
-15.5333
35.0167
Namatunu
-15.4
35.05
Gwaza
-14.8785
34.8114
Shire North
-15.2980
35.0799
Utale
-15.1667
35.05
Nkaya
-15.1282
35.0298
Bazale
-15.0167
34.9833
Rivirivi
-15.0167
34.9667
Balaka
-14.9870
34.9575
Faringdone
-14.8318
34.8670
Bilila
-14.8167
34.8333
Chinyama
-13.0167
33.6833
Lambulira
-15.4833
35.3
Mphonde
-16.55
34.9
These coordinates were obtained from the following sources: Geographic.org, Openstreetmap.org and Latlong.net. The GPS coordinates
for Thukuta station are the GPS coordinates of Mfera Health Centre,
which is the closest location with available GPS coordinates.
52