Economics and Measurement:

New measures to model decision making*

Ingvild Almås†, Orazio Attanasio‡, Pamela Jervis§

February 1, 2023
Abstract
Most empirical work in economics has considered only a narrow set of mea-
sures as meaningful and useful to characterize individual behavior, a restriction
justified by the difficulties in collecting a wider set. However, this approach of-
ten forces the use of strong assumptions to estimate the parameters that inform
individual behavior and identify causal links. In this paper, we argue that a more
flexible and broader approach to measurement could be extremely useful and allow
the estimation of richer and more realistic models that rest on weaker identifying
assumptions. We argue that the design of measurement tools should interact with,
and depend on, the models economists use. Measurement is not a substitute for
rigorous theory, it is an important complement to it, and should be developed in
parallel to it. We illustrate these arguments with a model of parental behavior
estimated on pilot data that combines conventional measures with novel ones.

* This paper is based on the Presidential Address Attanasio presented at the World Congress of the
Econometric Society in August 2020. At different stages, we received extremely valuable feedback from
many colleagues, including Joe Altonji, Manuel Arellano, Jere Behrman, Alberto Bisin, Richard Blun-
dell, Margheria Borella, Agar Brugiavini, Andrew Caplin, Rossella Calvi, Sarah Cattan, Flavio Cunha,
Maria Cristina De Nardi, Aureo de Paula, Armin Falk, Jim Heckman, Mike Keane, Sonya Krutikova,
Chuck Manski, Costas Meghir, Rohini Pande, Elena Pastorino, Larry Samuelson, Guglielmo Weber, Ken
Wolpin. Attanasio presented some of this material in seminars at the University of Chicago, Bonn Uni-
versity, the European University Institute, St Andrews. We thank Marianne Moreira and Diana Lopez
Perez for capable research assistance. We are grateful to the British Academy and the European Research
Center for the provision of grants that helped the collection of data in Tanzania, including Attanasio’s
ERC Advanced Gran 695300 (HKADeC). Jervis gratefully acknowledges financial support from the In-
stitute for Research in Market Imperfections and Public Policy MIPP (ICS13 002 ANID) and the Center
for Research in Inclusive Education, Chile (SCIA ANID CIE160009).
† Institute for International Economics, Stockholm University and Norwegian School of Economics,

e-mail: [email protected]
‡ Yale University, Institute for Fiscal Studies, FAIR@NHH, NBER; e-mail:
[email protected]
§ Universidad de Chile, Institute for Research in Market Imperfections and Public Policy, Institute for

Fiscal Studies and Center for Research in Inclusive Education, e-mail: [email protected]

1
1 Introduction
For many years, at least since Samuelson’s (1938; 1948) and Arrow’s (1959) contribu-
tions, most economists thought that good empirical work can only be based on a narrow
set of measures. The prevalent perception, with a few exceptions, was that we can only
meaningfully use measures of what people buy or do, their resources, prices, and (pos-
sibly) the markets they have access to. While many economists have also proficiently
used objective data, including biological and anthropometric data, the set of measures
deemed acceptable and interesting by a large fraction of the profession has been limited.
Data on stated preferences and their intensity, intentions to buy, stated actions in hypo-
thetical situations, subjective expectations, and attitudes and tastes – as well as data
on the possible drivers of individual choices, such as social norms, culture, or political
attitudes – were seen as not particularly useful and outside the realm of economics.
From the measures perceived as meaningful and acceptable one could, under cer-
tain assumptions, infer and characterize preferences and other structural parameters
that drive individual behavior, as well as some of the features of market structure and,
maybe, identify empirically the causal links among different variables. To perform such
exercises on a restricted set of measures, however, requires strong assumptions.
The reliance on restrictive and specific sets of measures, disregarding richer mea-
sures, such as data on rankings of different choices or the intensity of preference, data
on stated intentions, or answers to questions about choices in hypothetical scenarios has
been widespread in different economic fields and types of models, both static and dy-
namic. Such an approach has forced the use of restrictive models in empirical analysis.
In most static models, for instance, empirical research has used mostly choice-based
data in combination with objectively observed variables (such as prices or other indi-
vidual and environmental data).1 This approach has led researchers to focus on models
that either imposed homogeneity assumptions or a very specific set of heterogeneous
preferences. And even in models with heterogeneous preferences, such as the Random
Utility Model, the use of choice-based data poses key identification problems. Such
problems are particularly salient when individual choices are determined not only by
preferences and resources but also by other factors that are typically deemed unobserv-
1 Berry et al. (2004) is an exception, where rankings or ‘second hypothetical choices’ are used in the

identification of demand models for cars.

1
able, such as individual beliefs.2 In the case of dynamic models, where uncertain future
variables play a role, researchers used assumptions such as adaptive, myopic or, in re-
cent decades, rational expectations, which allowed an internally consistent solution of
the model under study and its empirical applications.
The kind of measures researchers consider as meaningful and usable implicitly de-
fines the models one can bring to data and, therefore, the domain of economics as a
social science with empirical content. In this paper, we discuss the important role new
measures can play in the study of economic behavior. The use of innovative measures,
such as answers to hypothetical questions used in conjunction with choice data, can
help the identification of causal links with less restrictive assumptions. It also limits the
need to identify exogenous sources of variation in observational data.
We discuss what should be measured and how these measures relate to models of
economic behavior. While we stress that rigorous and coherent models of human be-
havior and human interactions are key to understand economic reality, we do not think
such an approach implies that empirical studies should exclusively use measures de-
rived from choice data and easily and objectively measurable quantities (such as prices
or anthropometric data). Instead, we argue that economic theory can and should inform
the design of new measures that capture the factors salient for the models at hand. These
measures can lead to the use of more flexible and richer models of individual behavior.
One of the main reasons a large part of the profession shied away from certain
types of measures has been the challenge in designing appropriate tools to gather them.
Therefore, a substantial amount of research effort should be devoted to the design and,
importantly, the validation of new measurement tools, to ensure that they capture the
phenomena they are meant to. New measures, designed and devised by researchers,
should be piloted, validated and shown to be correlated with individual choices. Appro-
priate techniques in psychology and survey design can (and should) be used to develop
tools that could capture latent variables relevant for models of individual behavior.
Measurement research and progress are not new. An example is the development
of the unfolding bracket technique (Juster et al., 2006), to measure variables, such as
household wealth, which had been seen as impossible to assess accurately. New mea-
2A very recent paper by Dardanoni et al. (2022) discusses the restrictions that the use of choice data
requires for models with preference and other types of heterogeneity.

2
sures continue to emerge: Bloom and Van Reenen (2007), for instance, started a new
research agenda to construct innovative measures of management skills, seen as inputs
of a production function. Caplin (2021) has been proposing the design of new measures
thorough data engineering with a logic close to what we argue for in this paper.
A focus on measurement does not remove the need for theoretical models. Indeed,
the use and design of new measures should be driven by theory and the need to identify
key parameters of the theoretical models under study, using direct measures instead of
restrictive (structural) assumptions. Further, an approach to measurement and empirical
strategies that goes beyond standard measures does not imply rejecting the modeling
of individual behavior based on some sort of constrained optimization that satisfies a
set of axioms. As we argue below, the use of new and innovative measures allows
researchers to give empirical content to flexible and powerful models, which can replace
restrictive models constrained by the lack of available measures. The development of
new measures and new more flexible models should go hand in hand.
The rest of the paper is organized as follows. In Section 2, we provide some back-
ground material about measurement, its role and its interactions with economic theory.
In Section 3, we discuss various aspects of the relationship between theory and mea-
surement, and present the case for going beyond standard measures. We discuss what
measures are useful for different models, how the parameters of measurement systems
can be identified and how they can be relevant for defining the metric of the relevant
latent factors and how they can help the identification of causal links between the var-
ious latent factors. Having discussed these conceptual issues, we sketch, in Section 4,
a model of household behavior where some of the issues on the relationship between
theory and measurement are fleshed out and made explicit in terms of the latent fac-
tors and drivers that populate the model. In Section 5, we provide some examples of
measurement work and its use, using a novel data set that was collected to pilot new
measures of the abstract concepts that are used in the model presented in Section 4.
In Section 6, we use the new measures and suggest a theory-funded empirical model
of parental investment that shows how these can be used in combination with standard
ones. Section 7 concludes the paper.

3
2 The context
A stark statement of a restrictive approach to the study of preferences in economics is
in Stigler and Becker (1977), entitled De Gustibus Non Est Disputandum:

“... tastes neither change capriciously nor differ importantly between peo-
ple. [...] one does not argue over tastes for the same reason that one does
not argue over the Rocky Mountains - both are there, will be there next
year, too, and are the same to all men.” (emphasis added).

It is interesting to note the explicit reference to the stationarity of preference and its
cross-sectional homogeneity. Although empirical studies based on choice data and ob-
jective measures do not necessarily require cross-sectional homogeneity – and indeed
much empirical and econometric research has studied models which allow for hetero-
geneity – this statement implicitly asserts what are acceptable measures for economists
and imposes important restrictions on the models researchers can eventually study.
Imposing the use of a limited set of measures implies that identification (that is,
the possibility of retrieving empirically the fundamental features of the model under
study) may only be achieved with strong assumptions on tastes, beliefs, expectations,
and information individuals have access to. This is the price that is paid to assume that
such variables and factors cannot be observed or measured. However, the profession
has shown much skepticism towards novel measures that could provide information on
these types of variables, such as questions that pose hypothetical situations and evidence
from stated rather than actual choices.
These issues have been analyzed for a long time. An example of the arguments
about what and how to measure can be found in the discussions of stated preferences and
conjoint analysis, for instance, in Luce (1956, 1959); Luce and Tukey (1964), and Luce
and Suppes (1965) Likewise, the discussion of the Random Utility Model by Block
and Marschak (1960) states that the way of defining the class of basic observations and
testable conditions is to some extent arbitrary and dependent on the range of possible
experiments and observations. They further argue that it may be beneficial to follow
the practice in psychology of accepting subjects’ ranking of objects and intensity of
preferences, even if observed through a verbal statement (see also Caplin, 2021, for a
discussion of these issues).

4
 There are good reasons for the profession’s skepticism about certain measures. Mea-
suring hypothetical choices, preferences, and attitudes is fraught with difficulties. Fram-
ing effects, for instance, seem to be pervasive and introduce a number of potentially
severe biases. Several studies, such as List and Gallet (2001) and Murphy et al. (2005),
discuss common biases in answers to questions about hypothetical situations. An inter-
esting debate in this respect regards the use of contingent valuation. While this type of
measure is widely used in other disciplines,such as marketing,3 its use in economics has
received a considerable amount of resistance. Hausman (1994) expressed doubts about
its usefulness and Hausman (2012) labels the enterprise as hopeless. This skepticism is
partly due to measurement difficulties, though it is also probably due to the ambiguities
around what one is measuring when asking questions about hypothetical choices.
Other interesting early discussions of what could and should be measured was the
lively exchanges between Tobin (1959) and Katona (1959) on the usefulness of data on
buying intentions; Tobin strongly criticized the usefulness of such data on the basis that
they were not a good predictor of actual consumer choices. The reliability and predic-
tive power of buying intentions and purchasing probabilities data were later discussed
in Juster (1964, 1966) and then again in Manski (1990), who noticed that intention
questions are not necessarily useless if formulated properly. Manski argues the issue is
not what is being measured, but the specific tools and questionnaires being used.4
While these issues were being debated, some researchers used stated preferences
and elicitation of hypothetical choices to estimate key parameters of economic models.
Juster and Shay (1964), for instance, used the elicitation of stated choices in hypotheti-
cal situations to estimate the elasticity of the demand for loans to interest rates and loan
maturities; the interest rate and maturity of these hypothetical loans were exogenously
varied across respondents to the survey. Cross-sectional differences in loan demand
elasticities were then used to discuss the importance of liquidity constraints. More re-
cently, Lancaster and Chesher (1983) used, in conjunction with a model of employment
search behavior, “the answers to two simple questions” which could be interpreted as
providing information about the distribution of offer wages and reservation wages to
“...deduce structural parameters rather than estimate them” (p. 1661). What could and
3 Good references are Louviere et al. (2000) and Carson (2012).
4 Manski also remarked that intention data, while not been used much by economists, are instead
widely used in other disciplines, including marketing. Curtin (2016) provides a nice survey.

5
should be measured and its relation to black theory was discussed in Haavelmo’s (1958)
Presidential Address to the Econometric Society. Haavelmo perceived measurement
questions to be central to the development of economic theory:

“I think most of us feel that if we could use explicitly such variables as, e.g., what
people think prices or incomes are going to be, or variables expressing what peo-
ple think the effects of their actions are going to be, we would be able to establish
relations that could be more accurate and have more explanatory value. But be-
cause the statistics on such variables are not very far developed, we do not take
the formulation of theories in terms of these variables seriously enough. It is my
belief that if we can develop more explicit and a priori convincing economic mod-
els in terms of these variables, which are realities in the minds of people even if
they are not in the current statistical yearbooks, then ways and means can and will
eventually be found to obtain actual measurements of such data.” (emph. added).

Haavelmo’s address argues new measurement tools should be developed in reac-

tion to theoretical models that are both rich enough to capture complex phenomena and
require novel measurement tools. Such new measurement tools would allow the identi-
fication and characterization of richer models, where strong assumptions are relaxed. In
what follows, we similarly argue that new measures for variables that play an important
role in several theoretical models should be developed and validated.
These early efforts and discussions were not without consequence. For instance,
Katona’s work (see e.g., Katona, 1974) led to the establishment of the Michigan survey
of consumer sentiment, which is still running (and used) today. In recent decades,
however, the design and use of such surveys has become rare. From both a theoretical
and measurement point of view, the consensus within most of the economic profession
went towards an almost exclusive use of choice based and objectively observable data.
An exception is the field of experimental economics, where researchers construct
controlled – and often artificial – settings in which subjects make choices and, in so do-
ing, reveal their preferences, beliefs, and other key drivers of behavior (Plott and Smith,
2008). This approach allows researchers to control the situations presented to subjects,
typically in a laboratory setting, so to control away many potential confounding fac-
tors to reveal features of human decision-making. A considerable amount of such work
has been done on mechanisms and protocols to elicit primitives of individual behavior.

6
However, lab experiments often require participants to behave abstracting from their
present circumstances, imposing a separation between experimental and actual behav-
ior. Indeed, background information on experiment participants is rarely collected and
experimental data are rarely used in conjunction with observational data.
The expanded use in the field of techniques and protocols developed in labs and
the simultaneous collection of experimental and standard observational data are a sign
that the economic profession has been changing its approach to what can and should
be measured. Evidence on measures of preference and attitudes towards redistribution,
attitudes towards migrants, bargaining and social preferences, reciprocity in conflict
areas, and willingness to compete through experiments combined with observational
data are reported in Almås et al. (2020b), Alesina et al. (2018b), Almås et al. (2018);
Buser et al. (2014); Cavatorta and Groom (2020). At the same time, techniques used in
empirical studies on standard data have been reproduced in the laboratory, in particular,
for the analysis of auction models.5
Several other measurement novelties have proliferated in recent years. One impor-
tant contribution, which has enlarged the set of variables economists consider measur-
able, is the study of subjective expectations (e.g., about income, inflation, or rates of
return) promoted in an important contribution by Manski (2004).
Subjective expectations data were already contained in the early NLS data, some-
times known as the Parnes data and discussed in Parnes (1975).6 Another example of
early collection of subjective expectations is Visco (1984). References relevant for the
design of effective expectations measures include Dominitz and Manski (1997), Do-
minitz and Manski (1996), Potter et al. (2017).
The collection of expectations data has evolved considerably and these types of data
are now routinely collected. the Bank of Italy has collected subjective expectations data
for a number of years and the Federal Reserve Bank of New York has been and similar
data are also collected by the Bank of Spain and the European Central Bank. An impor-
tant example of high quality subjective expectations data are those collected systemat-
ically since 2013 by the Federal Reserve Bank of New York: the Survey of Consumer
Expectations (https://www.newyorkfed.org/microeconomics/sce/background.html).
5 See Ertaç et al. (2011); Salz and Vespa (2020). We thank Aureo de Paula for pointing this out.
6 We are grateful to Ken Wolpin for pointing us to these data.

7
 Research economists have also started to use subjective expectations data within
structural models of economic behavior. An early use is Wolpin and Gonul (1985),
while a non-exhaustive list of more recent studies include Jappelli and Pistaferri (2000),
Pistaferri (2001), Pistaferri (2003), Van der Klaauw and Wolpin (2008), Kaufmann and
Pistaferri (2009), Attanasio and Augsburg (2016), Paiella and Pistaferri (2017), Attana-
sio et al. (2018), and Giustinelli et al. (2019).
The availability of subjective expectations data allows researchers to avoid strong
assumptions, such as rational expectations. Moreover, such data allow the identification
of genuine measures of uncertainty, which, using data on actual realizations of the vari-
able of interest is not easily disentangled from individual heterogeneity or variability
that is known and deterministic to individuals. Moreover, the absence of expectations
data implies strong limitations to the identification of certain parameters. Chen et al.
(2020) show that, without assuming rational expectations (or data on subjective expec-
tations), only set (rather than point) identification can be achieved.
A related topic is the measurement of beliefs about the return to specific invest-
ments, such as different types of investment in human capital and education. Rather
than assuming that individuals have rational expectations about the returns to certain
investments, researchers have started eliciting data on beliefs about returns. Several
studies have started collecting and analyzing data on parental beliefs, following a prac-
tice that has been used for some time in psychology and child development, as surveyed,
for instance, by Miller (1988). Examples of such studies include Cunha et al. (2013),
Boneva and Rauh (2018), Attanasio et al. (2019b), Attanasio et al. (2019a), and Biroli
et al. (2022). Likewise, Dominitz and Manski (1996), Wiswall and Zafar (2015), and
Delavande and Zafar (2019) have studied beliefs about the returns to college and col-
lege enrollment choices, Bobba and Frisancho (2020) assessed how college application
choices are affected by students’ perceptions of their own ability and how these can
be changed by additional information, while Dizon-Ross (2019) studied how parental
beliefs about their children’s abilities affect their choices.
Another example of new measurement tools being developed and used by economists
is the study of stated preferences and answers to hypothetical questions. A good exam-
ple of such a practice is Ameriks et al. (2020), which uses a combination of stated
preferences and actual choice to identify complex structural models. Bernheim et al.

8
(2021) discuss how to use data on hypothetical choices to identify causal links in eco-
nomic models. These approaches are analogous to questions about intentions (e.g., how
respondents would allocate hypothetical resources among different potential uses), to
elicit information about individual tastes and preferences.7
While resources, preferences and tastes are obvious drivers of individual choices,
other factors can also be important drivers of behavior. In certain contexts, for example,
the quantity and quality of information available to individuals is important. Individu-
als’ access to markets or networks can influence the allocation of resources across time
and space. Additional factors, such as learning, risk sharing arrangements, preferences
about different policy options, attitudes and social norms, might also affect individual
choices (and preferences). When studying household-level choices, for instance, who
controls resources and bargaining power within the household might be important.
A variety of studies have used new and innovative measures to analyze many of
these phenomena. While the scope of this paper is not to provide an exhaustive survey
of the relevant literature, we mention Attanasio and Krutikova (2020) on measuring the
quality of information in networks and the role that it plays in providing informal insur-
ance, Almås et al. (2018) and Jayachandran et al. (2021) on measuring bargaining power
within couples, and Alesina and Angeletos (2005), Almås et al. (2020b), Alesina and
La Ferrara (2005), Kuziemko et al. (2015), and Alesina et al. (2018a) on measuring atti-
tudes towards and perceptions of immigration, social mobility, redistribution, and other
policy factors. These studies and others use novel measures of attitudes, sometimes
in combination with standard survey data, to quantify the effects of such variables.8
Kaiser and Oswald (2022) uses longitudinal data from three countries to document the
predictive value of ‘happiness scales’ on individuals’ important decisions.9
These studies illustrate the active and ongoing discussions in economics about mea-
surement and its relation to theory. Recent developments indicate that the profession
is moving towards using choice data and directly observable variables in combination
with stated preferences and answers to hypothetical questions. An interesting and re-
7 Studies of stated preferences and answers to hypothetical questions include Blass et al. (2010),

Kesternich et al. (2013), Ben-Akiva et al. (2019), Harris and Keane (1998), and Erdem et al. (2005).
8 These issues are discussed extensively by La Ferrara (2019).
9 This study stresses that comparability across different contexts implies the need to establish a car-

dinal metric for measures that are often obtained as ordinal indexes, an issue we discuss below.

9
cent take on measurement and its relation to theory is discussed in Stantcheva (2022):
“Surveys are not merely a research tool. They are also not only a way of collecting
data. Instead, they involve creating the process that will generate the data.” We share
this view. In what follows, we develop it to include the design and validation of new
measures to be used in combination with traditional ones for the empirical study and
characterization of economic models.

3 Measurement and Theory

The economic models economists work with often use abstract constructs that are not di-
rectly observable. The measures relating to these constructs that economists collect and
use are (and should be) driven by the theoretical framework that organizes our think-
ing. Examples abound in several fields of economics including the work by Keynes,
Stone, and others that led to the development of National Accounts and the analysis
of consumer demand, which is behind most methods to construct price indices. The
measurement of growth, consumption, and price indices has used theoretical models as
a foundation for the construction of new measures. Analogously, the creative use of
micro-data has informed the development and calibration of theoretical models.10
We discuss the relation between theory and measurement in Section 3.1, and, in
Section 3.2, we argue that new measurement tools, beyond standard measures, are de-
sirable and useful. In Section 3.3, we discuss some issues relevant for measurement
systems.

3.1 Economic Models, Latent Factors and Measurement Systems

Many theoretical models in economics can be represented by:

F(θ ; φ ) = 0 (1)
10 Classicstudies that led to the development of Modern National Accounts include Keynes (1936),
Kuznets et al. (1937), Kuznets (1941), Gilbert et al. (1949), and Stone (1984). Examples on the effects of
new products and quality on the measurement of inflation include Bils and Klenow (2001), Bils (2009),
and Crawford and Neary (2021). For price indices, see Stone (1954), Christensen et al. (1975), Deaton
and Muellbauer (1980), and more recently Nordhaus (1998). On the importance of creative destruction to
measure growth Aghion et al. (2019); and Neary (2004) and Almås (2012) on international comparisons.

10
where θ is a matrix of variables of interest or factors, some of which are latent, in that
they are not necessarily observed. The parameter vector φ characterizes the function
F, which represents individual behavior and interactions (such as markets), that is, the
relevant economic model. F typically defines what the variables of interest are.
Within this framework it is easy to introduce a number of details about the features
of the economy under study. For instance, one could include in model (1) uncertainty
and imperfect information, and consider additional factors relating to the information
available to the model’s individuals. The dimension of the model’s latent factors de-
pends on the specific issues under study. Richer and more realistic F functions require
a richer set of factors and, to be characterized empirically, a richer set of measures.
The factors that populate a theoretical model are often well-defined but unobserved
variables, such as prices of very finely defined goods or the quality of family environ-
ment and schools. In practice, what is often available are markers corresponding to
(some of) these theoretical constructs. To bring the theoretical models to data and to
identify and estimate the parameters φ that define the causal links one is interested in,
it is necessary to be explicit both about the theoretical model and about the relation-
ship between the relevant latent factors and the available measures. In other words, to
give empirical content to the function F in model (1), one needs a measurement system
that relates the latent factors in F to the available measures. A possible mapping is the
following:
m = g(θ , ε) (2)
where m are available measures related to the (potentially unobservable) factors θ
through the function g. The vector ε is measurement error that, together with the possi-
bility t the function g is not injective, prevents the direct observability of (some of the)
θ . The model F defines the factors of interest, and, in turn, guides what measures to
look for. The available measures and the measurement system in (2) define what latent
factors one can study empirically and which models can be taken to data.
Such mapping between latent factors and constructs of interest for economic theory
and a set of available measures resonates with Goldberger’s (1972) description of the
interplay of theory and measurement in his Fischer-Schulz lecture (emphasis added):

“By structural equation models, I refer to stochastic models in which each equation
represents a causal link, rather than a mere empirical association. The models arise

11
 in non-experimental situations and are characterized by simultaneity and/or errors
in the variables. The errors in the variables may be due to measurement error in
the narrow sense, or to the fact that measurable quantities are not the same as the
relevant theoretical quantities. Generally speaking the structural parameters do
not coincide with coefficients of regressions among observable variables, but the
model does impose constraints on those regression coefficients”.

Goldberger (1972) uses the Permanent Income model as an example, where perma-
nent income is the interesting construct and the empirical measures potentially related
to it are current income and consumption. In a similar vein, Griliches (1974) discusses
the relationship between earnings, schooling, and ability. More recently, Cunha et al.
(2010) use a relatively flexible version of measurement system (2) to estimate the pro-
duction function of human capital. In a different context, the estimation of a production
function with endogenous inputs could be viewed in a similar fashion.11 The early work
on Multiple Indicators Multiple Causes (MIMIC) models and, more generally, studies
on factor models in economics, psychology, sociology, and geneticsare relevant and
important.12
As, economists’ theoretical models are often populated with abstract constructs,
recognizing it explicitly is useful. It clarifies the research objectives and may moti-
vate attempts to measure additional relevant variables, which, in turn, can motivate
researchers to use more realistic models that are subject to less stringent assumptions.

Which latent factors to measure. The measurement system and the measurement
tools used in empirical studies – as well as what should be measured (and possibly how)
– should be informed by the specific questions researchers ask and by the theoretical
models being used. Expanding the set of objects one measures allows the consideration
of more flexible models and avoids strong and sometimes misleading assumptions.
The latent factors of interest depend on the complexity of the theoretical model be-
ing studied, which, in turn, might depend on the phenomenon being interpreted. In this
regard, an explicit discussion of the restrictions imposed on the theory by data availabil-
11 See, for instance, Olley and Pakes (1992); Levinsohn and Petrin (2003); Ackerberg et al. (2015);
Gandhi et al. (2020); Doraszelski and Jaumandreu (2013), and Doraszelski and Jaumandreu (2018).
12 See Wright (1934), Duncan (1966), Goldberger (1971), Goldberger (1972), Griliches (1974),

Jöreskog and Goldberger (1975), and Chamberlain and Griliches (1975).

12
ity and measurement challenges is useful. In certain contexts, it might be apparent that
these restrictions have negligible impacts for what is being studied; in other contexts,
however, restrictive definitions of the relevant latent factors might substantially limit
the ability of a model to explain observed phenomena.
When working with demand systems, researchers typically aggregate different com-
modities in coarse categories. However, available data (even when very detailed) might
miss important components of the commodities considered, such as quality of certain
commodities or the market structure faced by consumers.13 When studying intrahouse-
hold allocation, measures of individual-level consumption (in addition household-level
expenditure) are often unavailable, forcing strong assumptions.
Consider, for instance, studies of production functions where output is the result of
combining different inputs, such as human and physical capital. Until the late 1990s,
studies of labor market inequality used a basic model where production is performed
via a production function that uses two types of labor (skilled and unskilled), fitted well
to a set of labor market facts.14 That model, however, could not explain what happened
in the first part of the 21st century. As a result, new models that disentangle skills from
tasks have been developed, such as in Acemoglu and Autor (2011) and Deming (2017).
The empirical needs of these models require new types of data, such as the O*Net or
DOT data sets used, for instance, by Autor and Dorn (2013) and Acemoglu and Re-
strepo (2019). To study empirically more complex production functions with flexible
roles of different skills, such as sociability, drive, and motivation, in addition to cogni-
tion, it is key to measure these skills, how they differ across individuals, how different
occupations might require different combinations, and how they are remunerated in the
labor market. Likewise, the measurement of these latent factors, and the comparability
of measurement tools, often used in different contexts, then becomes particularly chal-
lenging and key to the results one obtains. Similarly, in studying firms heterogeneity,
Bloom and Van Reenen (2007), Bloom et al. (2019), and Scur et al. (2021) use an in-
13 Researchers typically use price indices for the aggregate measures of commodities that are used in
analysis. However, in some contexts prices may not be linear and change with the quantity purchased,
as in many models of price discrimination (e.g., Maskin and Riley (1984), Jullien (2000), and Attanasio
and Pastorino (2020)). Recent progress has been made with the analysis of scanner data (e.g., Einav et al.
(2008), Griffith and O’Connell (2009), and Dubois et al. (2020)).
14 As discussed, for instance, by Katz and Murphy (1992), which developed Tinbergen’s original

explanation of labor market inequality as the effect of the relative demand and supply of different skills.

13
novative survey, now deployed in several countries, to measure managerial skills as an
input in the production function.
In summary, the questions a researcher is addressing, the theoretical models they
use, and their empirical performance define the key latent factors of interest. This, in
turn, defines which measures are needed to give empirical content and bite to the theo-
retical framework at hand. While in some cases standard measures, possibly anchored
by choice data, are sufficient, in many other cases they are not.

3.2 Beyond standard measures

The revealed preference approach has been very useful for structuring our thinking
around agents’ decision-making and relating rigorous theory to data with relatively few
assumptions. With observations about individual choices, the conditions under which
these choices are made (such as the resources individuals control and the prices they
face), and objectively measurable variables, preferences consistent with a set of axioms
which defines a theory of economic behavior may be revealed. Identification of struc-
tural parameters can then be achieved without explicit measures of tastes, beliefs, or
attitudes, and without questions about behavior in hypothetical circumstances.
However, this approach often rests on strong assumptions. Many interesting eco-
nomic models routinely deviate from such assumptions and constructs, posing difficult
questions if their empirical analysis relies exclusively on choice data and variables
objectively measured. Often these challenges are tackled with strong theoretical or
empirical assumptions. For instance, to study dynamic problems where future and un-
certain variables are key, researchers frequently have used the assumption of rational
expectations and of an efficient use of all the available information.
On the empirical side, to avoid endogeneity issues, researchers try to isolate or
create sources of exogenous variation through ‘natural’ experiments, a strategy that
often requires ad-hoc and un-testable assumptions. Even genuinely exogenous variation
induced by randomization can often identify a narrowly defined set of parameters.
Expanding the set of measures one uses might make it possible to study more realis-
tic yet still rigorous models. With new and more comprehensive measures, it is possible
to maintain consistency with economic theory while at the same time relax some strong
assumptions. For instance, questions about behavior in hypothetical scenarios can in-

14
troduce exogenous variation in a much richer way than natural experiments. While
important contributions on new measures have been made in the literature – e.g., the
aforementioned literature on subjective expectations – there is still a need to develop
and validate new measures that can be used in the analysis of agents’ decision-making.
The choices between alternative theoretical approaches should not be forced and
limited by the availability (or lack thereof) of data and appropriate measures. For in-
stance, when trying to explain the presence of what has been called present bias, an in-
teresting debate is whether one should model time consistent decision makers or allow
for the simultaneous presence of present and future selves. While some experimental
measures could also be (and have been) devised to measure the presence of present
bias, the choice between different models is a theoretical one. Whether present bias is
introduced via temptation preferences, as suggested by Gul and Pesendorfer (2011) –
which satisfies time consistency and the axiomatic approach, where the decision unit
is the same over time – or whether one allows the presence of multiple selves for an
individual, it is a theoretical choice. However, the availability of appropriate measures
could crucially help researchers discriminate between these alternative models.

The need for and use of new measures: a few examples. In order to clarify the
need for new measures that go beyond data on choices and to better understand the
framework in which such new measures can be used, this section provides a few ex-
amples, some of which relate directly to the applications we present in Sections 4 and
5. These examples are relevant for: (i) the definition of decision units in models of
choice; (ii) the separate identification and characterisation of preferences; and (iii) the
characterization of the economic environment.

Decision units. A first step when modeling individual behavior is to define the deci-
sion unit. In standard consumer theory, the household has most often been considered as
the relevant decision unit. In such a unitary model, the household as a whole is consid-
ered the relevant decision unit with well-defined preferences. In recent decades, how-
ever, researchers have focused on how resources are controlled and allocated within the
household, developing alternative models where multiple decision makers, each with
distinct preferences, interact within the household to arrive at household-level choices.

15
 The collective model, first proposed in Chiappori (1988), is one such attractive al-
ternative. Its key assumption is that choices, while resulting from interaction between
decision-makers with potentially distinct preferences, are efficient. While a number
of important theoretical results, have been derived,15 characterizing the parameters of
the models used and testing their validity exclusively with choice data on household
consumption and expenditure is challenging.
Much more can be learned by using additional information on private consump-
tion. A number of researchers have used information at the individual level within
households, such as Dercon and Krishnan (2000), Dunbar et al. (2013); and more re-
cently Lechene et al. (2022). However, even when individual level data are available,
identifying the determinants of individual and eventually household behavior can only
be achieved with strong assumptions. For instance, it is difficult to allow for caring-
preferences, i.e., when one partner cares about the consumption of the other partner.
Instead, nonstandard measures that do not rely on the exclusive use of choice-based
data can generate important insights about the process of intrahousehold allocation of
resources. For instance, hypothetical choice scenarios elicited separately from house-
hold members can generate direct information on individual tastes. Likewise, it may
be possible to derive information on the relative bargaining power within the house-
hold, a measure that goes beyond standard choice-based data. We discuss some of these
measures in Section 5.1.

Disentangling beliefs and tastes. Agents in most models in economics make deci-
sions to maximize an objective function, given the resources available to them. These
decisions then depend on their preferences and on their perception of the process that
links actions to outcomes. Often, the characterization of such a process is of key inter-
est to researchers. Identifying the causal links that define it requires understanding how
individual choices are made. It is often assumed that individual decision-makers know
the process that determines the outcomes they care about, given their actions and other
variables. However, in many situations, this assumption is a strong one.
15 See, for instance, Browning and Chiappori (1998); Bourguignon et al. (2009); and Cherchye et al.
(2011). Attanasio and Lechene (2014) present tests of the collective model using the variability induced
in a household demand curve by two different distribution factors, i.e. variables that affect Pareto weights
but not utility.

16
 The challenges related to disentangling individual perceptions or beliefs and tastes
have been extensively discussed in several different settings (see e.g., Caplin, 2021).
Possible approaches are the direct elicitation of beliefs (and retrieving preferences from
choice data); and the elicitation of preferences through experimental approaches and
hypothetical choice scenarios, holding beliefs constant by giving surveys’ respondents
full information about the context.
Beliefs elicitation. Direct elicitation of beliefs has been used in a model of child
development and parental investment in Cunha et al. (2013) and by Attanasio et al.
(2019b), eliciting the perceived productivity of parental investment on child develop-
ment. We follow a similar approach in Section 5. In a different context, Mueller and
Spinnewijn (2021), studying on search behavior among unemployed, suggest using di-
rect measures of beliefs while retrieving tastes as a residual.
Preferences elicitation. Another approach is to elicit preferences directly holding
beliefs constant in controlled situations with full information about the actual setting.
This can be done either with experimental methods, revealing preferences through real
choices, or posing different hypothetical scenarios to respondents and eliciting their
(stated) preferences. An early example of such a strategy is the aforementioned Juster
and Shay (1964). More recently, Ameriks et al. (2020) and Bernheim et al. (2021)
have used hypothetical questions to estimate parameters that characterize individual
preferences.
Elicitation of preference and beliefs. Instead of eliciting beliefs (preferences) and
inferring preferences (beliefs) as a residual by combining the elicitation with choice
data, one may directly elicit both beliefs and preferences in particular choice situations.
For instance, in a recent paper Adams and Andrew (2019) who use survey experiments
to elicit average beliefs and preferences in the Indian (specifically Rajasthani) marriage
market for young brides.16 To allow for heterogeneity in both beliefs and preferences
it may be beneficial to combine the strength of economic experiments or hypothetical
choices to elicit preferences, and the direct elicitation of beliefs through surveys.
16 Other recent studies include List et al. (2021) who have elicited beliefs from parents in Chicago;

Bobba and Frisancho (2020) collect and use data on self-perceptions of academic achievement among
high school students in Mexico; Miller et al. (2020) study beliefs about contraception effectiveness.

17
The environment. To better understand individual (economic) behavior, it is useful to
measure how wider constructs, such as institutions, communities, and society at large
affect individual behavior via social norms, attitudes, or culture, and model their evolu-
tion (see e.g., Bisin and Verdier (2000) for a discussion of how culture evolves).
Social norms and attitudes affect individuals’ objective functions in significant ways.
In a recent paper, for example, Field et al. (2021) study the effect of an intervention
aimed at increasing female’s control of resources and find that its impact resulted in an
increase in female labor supply, contradicting the implications of a standard collective
model with individual utility depending on consumption and leisure. The paper’s au-
thors argue that, in reality, social norms play an important role in determining choices
and this type of intervention might have led to a shift in such norms. The challenge then
is to determine appropriate and validated measures of such norms.
Along similar lines, some interesting measures are those designed to capture what is
sometimes defined as social capital, i.e. a set of norms that inform individual behavior
and affect the ability of a society or community to provide public goods and internalize
externalities, or other social attitudes that might affect individual interactions. A variety
of measures, ranging from participation in certain activities (from blood donation to
church attendance, see, for example, Guiso et al. (2004) and Guiso et al. (2006)) to data
derived from field experiments (see Attanasio et al., 2012) to the effect of deterrence
on preferences (see Cavatorta and Groom, 2020), have tried to capture social norms,
attitudes, beliefs, and the role of culture.
In characterizing empirically certain markets (such as credit or insurance) and deter-
mining the model that best describes them, quantitative measures of specific frictions,
jointly with choice data, can be very useful. In models of insurance with imperfect
information, it may be useful to devise measures of the quality of information in a risk
sharing group, as done, for instance, by Attanasio and Krutikova (2020).

What measures for what theories. We have discussed a few examples of theoretical
models whose empirical analysis might need additional measures. One example is the
identification of individual beliefs and preferences without strong assumptions.
In modeling parental behavior, for instance, it has been often assumed parents are
fully aware of the nature of the process of child development. While such an approach

18
can ease the analysis, a less restrictive model, embedding a more complex structure F in
model (1), allowing for distortions in parental beliefs, may be more realistic and avoid
misleading conclusions. In Section 4, to illustrate how additional and somewhat uncon-
ventional measures could (and should) be used in conjunction with traditional ones, we
sketch a model of parental investment, which we analyze empirically in Section 6.
In some cases, the additional measures that allow the empirical characterization of
more general models are just finer and more detailed versions of existing ones (as in
the case of individual rather than the household level of consumption). In others, the
new measures try to capture new concepts that are specific to the model being ana-
lyzed, such as (distorted) beliefs about child development or bargaining power within
the household.
The lack of measures of key variables in the theoretical framework considered can
force strong assumptions, and yield misleading results. These considerations are rel-
evant, we believe, for the debate between Gul and Pesendorfer (2011) and Camerer
(2011). While both papers make some interesting and important points, they take what
we think is an over-restrictive approach. Gul and Pesendorfer (2011) insist that eco-
nomic models describe the behavior and interactions of agents that are assumed to max-
imize a given objective function and that these models’ features should be consistent
with a set of axioms that help to frame them. The insistence on a framework consis-
tent with theory is important, as it gives discipline and empirical bite to the theoretical
models considered and, importantly, makes a well-defined welfare analysis possible.
However, Gul and Pesendorfer want to refrain from using data and measures different
from data on actual choice to characterize empirically these models. While the premise
that the empirical models economists yaw should be theory consistent is a sound one,
we believe that complex models are often better analyzed and characterized using data
beyond those derived from observed choices. More importantly, such data could allow
the use of richer models. Finally, while the measures of certain key factors might be
affected by multiple biases, the problem, as clearly discussed by Manski (1990), is not
with the measures per se but with the tools used to collect them.
Camerer (2011), on the other hand, points out that, again correctly in our opin-
ion, new measures (such as the neurological data and biological markers he discusses)
can be useful to better characterize individual behavior. While it is not obvious to us

19
whether data from a Functional Magnetic Resonance Imaging (fMRI) could ever be re-
lated to specific aspects of individual behavior, such as discount factors or risk aversion,
using measures different from choices, either bio-markers or others, to describe better
individual behavior can be very useful, at the very least to improve the precision of our
estimates and the efficiency of our tests. However, Camerer (2011) seems to want to
characterize individual behavior in a way that abstracts from a set of theoretical axioms
and to describe directly the relationship between biological mechanisms and behavior.
Apart from the difficulty in pursuing such a strategy, such an approach goes beyond
the realm of economics. Finally, the rejection of a specific model in one context is not
a good reason to throw away the whole approach and work with models that are not
consistent with a set of axioms.17
Interestingly, Benhabib and Bisin (2011) argue that traditional decision theory, as
advocated by Gul and Pesendorfer (2011), focuses only on the need to model choices,
while other approaches, somewhat misleadingly labeled as behavioral economics, want
to understand the processes that lead to a specific set of choices. To better understand
processes, additional measures, such as biological and anthropometric ones but also
measures of a wide variety of latent factors, can be useful, if such measures are inte-
grated in well-defined decision models of individual choices.
As economists, we need models that focus on economic ideas. Such models should
not aim to describe completely psychological processes or define what happiness is.
Well-constructed economic models, whose main aim is to describe and understand in-
dividual choices and interactions, should be based on a set of axioms and be consistent
with them. To give empirical content to such models additional measures that comple-
ment data on choices (and prices and resources), whose design and features should be
driven by the needs of the theory, can be useful.

3.3 Measurement systems

System (2) defines a measurement system in a fairly general way. Having discussed its
relation to theory, we now move to a number of issues relevant for the specification and
17 The discussion by Gul and Pesendorfer (2011) of present biases and the different approaches based
on hyperbolic discounting (which leads to time inconsistent preferences) and a temptation model (which
yields at the same time consistent preferences and present bias) is particularly illuminating in this respect.

20
estimation of such measurement systems. To make the discussion concrete, we use a
specific characterization of the measurement system (2), similar to that used by Cunha
j
et al. (2010). We denote with θit the j − th element of the vector θ for individual i at
jk
time (or age) t. Let mi,t j be measure k j of the K j available and relevant for factor j. We
assume that factors and continuous measures are related by the following system.

jk jk j jk j jk
mi,t j = αt + βt j θit + εit j , j = 1, ...J; k j = 1, ..., K j . (3)
For discrete measures, we assume an Item Response Theory (IRT) model, extensively
used in the psychometric literature and that we discuss at length in Appendix A1. For
binary variables we have:
( jk jk jk
j
jk j 1 i f αt j + βt j θit + εit j > 0
mit = (4)
0 otherwise
jk j
where εit in both equations are additive measurement errors independent from each
jk j
other and from the latent factors. The binary case makes evident the role of the αt
jk
and βt j parameters representing the discriminating and salience properties of different
measures. An item with very low or high values of α will have relatively low variability,
and therefore will not be able to distinguish individuals with different values of θ , the
factor of interest. Under some assumptions, the parameters of systems (3) and (4) (and
of the distribution of the latent factors) can be estimated so to obtain estimates of the
unobserved factors from the available measures.

Specification, identification and estimation of the measurement system. The rela-

tion between factors and measures as written in systems (3) and (4) is, in many ways,
restrictive. However, it is useful to discuss some of the issues with measurement sys-
tems, starting with what is needed to identify empirically such a model.
The identification of the parameters of the model above or, more generally, that
in system (2), requires the availability of several measures linked to each latent factor
with measurement errors that are uncorrelated among at least two of them. An extensive
discussion of the identification issues of these models is in Cunha et al. (2010), referring
to the contributions of, e.g., Schennach (2004) and Hu and Schennach (2008).18
18 In some cases, non-parametric identification requires at least 3 variables. Identification here refers
to the parameters of the function g in system (2).

21
 We will not repeat the formal arguments here, but we do stress that the need of
independent measurement errors should inform the way surveys are designed and data
collected. It would be desirable, for instance, to randomly assign different interviewers
to collect different measures targeted at the same variables or collect them on different
(randomly allocated) days. A similar argument applies to attrition (an extreme form
of measurement error). One could, for instance, allocate resources spent on minimiz-
ing attrition randomly across observations. These considerations about data collection
make clear the existing trade-offs: certain strategies might maximize data quality, while
others might provide information that could be used to deal with measurement error.
Considering systems (3) and (4), it is clear that, even with specific assumptions
about the distribution of the latent factors θ j , not all parameters of these systems (that
jk jk
is, the αt j ’s, βt j ’s, and the moments of the θ j ’s), can be identified. It will be necessary
to define a metric for the unobserved latent factors through appropriate normalization.
One possibility is to normalize the mean of the factors to 0 and the variance-covariance
matrix to the unit one, an option often used in standard software packages, together with
jk j
that of normality of the latent factors. Alternatively, one could normalize the αt and
jk j
βt of a specific measure to 0 and 1, respectively, effectively using that measure as
the relevant metric for the unobserved factor. Both approaches are valid and effectively
equivalent, with some important caveats.
Regardless of whether one normalizes the means and variance-covariance matrix
of the factors or some of the parameters of the measurement system, when one has
longitudinal data, it is necessary to establish whether one imposes these normalizations
for the first t or for every t. Depending on the context, different approaches might be
more useful. If one is interested in how the latent factors change over time,19 it might be
more useful to normalize only at one point in the observation sample. Such an approach,
however, imposes the assumption that the relationship between the various measures
and the latent factors is unchanged (i.e. measurement invariance). The imposition of
these normalizations has implications for the interpretation of the evolution of different
factors and their growth, as discussed in Agostinelli and Wiswall (2017).20
An issue, related to the normalization of the different measures and the identification
19 Forchild development, for instance, one might be interested in measuring children growth.
20 Arelated issue, relevant for longitudinal data, is when the measures that are appropriate for a factor
change with time. These issues are discussed in the case of child development in Attanasio et al. (2020a).

22
of a metric for the unobserved latent factors, is that they often enter economic models as
cardinal variables. In some cases, the relevant cardinal metric can be easily identified,
in others that is not the case. This statement applies not only to models that consider,
say, consumption or income but also, for instance, to the models of child development
we consider below or in studies that consider the value added provided by schools.
The issue is particularly difficult to deal with when the available measures are of an
ordinal nature, using, for instance, Likert scales. The specification of the measurement
system, in most contexts, should strive to obtain cardinal measures, which can be used
in combination with ordinal measures but that provide the necessary anchor and metric
that allows to obtain cardinal measures of the relevant factors.
As written in systems (3) and (4), the measurement system assumes that each factor
affects only one measure; that is, it is a dedicated measurement system. This assump-
tion can be relaxed and have several factors affecting a single measure. However, as
mentioned by Cunha et al. (2010), to achieve identification it will be necessary to have
at least one measure per factor that is affected only by that factor.
In many contexts, researchers estimate unobserved latent factors using a variety of
tests that have become standard practice in the academic community and beyond. In the
context of child development, for instance, much work exists in psychometrics which
has influenced and shaped the development of a number of tests designed to measure
different dimensions of child development or its drivers. Most of them are made of a
large number of individual items that are routinely aggregated using scoring algorithms
that deliver estimated developmental scores or indexes of parental investment. These
algorithms were typically developed using factor models similar to those we discussed
on samples of children on which the original items were tested. In other cases, the
scoring mechanism is very simple, like the sum of correct answers to a number of
questions.21 While the original algorithms were eventually validated in a number of
samples, it is not necessarily the case that, several decades later and in contexts that
might be different from those in which the scoring algorithm was constructed, the same
scoring algorithm is necessarily the best way to aggregate the information from the
individual items. A feasible and more effective use of the data collected would be to
21 An example in child development is the MacArthur Language Inventory test, a list of a few hundred
words, with the child’s caregiver informing whether the child understands or can say each of them. There
is no reason to take the unweighted sum of such words as the estimated latent factor.

23
re-estimate the scoring algorithms (that is, the measurement system that relates latent
factors to the available measures) in the new context where these data are collected. As
we discuss in Section 5, this can also lead to the collection of different tests that could
more easily be deployed in a given context. Along the same lines, new items could be
piloted to complement the existing standard tests.
These issues are particularly relevant in developing countries, which are contexts
very different from those where the tests were developed and the scoring algorithms
designed, typically in developed countries. Many items might present flooring or ceil-
ing effects, so that the specific tool is not able to capture any variability in the study
sample. The estimation of a measurement system to construct context specific scor-
ing algorithms is an effective and simple way to summarize efficiently the available
measures.

Identifying causal links between latent factors. Another important set of related
considerations is about the identification of causal links between latent factors. While
the issue here is in terms of the theoretical structure that links the variables of inter-
est, whether such links can be identified empirically might depend on the nature of the
data that are available and collected. Indeed, the need to identify specific causal links
should and often does drive the design and collection of specific surveys and data sets.
A good example is the design of a Randomised Controlled Trial (RCT) where subjects
are randomly exposed to a treatment or an intervention. In this particular example, the
investigator creates exogenous variation (exposure to a treatment) on which informa-
tion is collected to establish (if the experiment is performed correctly) a specific causal
link. However, when the object of interest is more complex than the average impact of
the intervention in a given context, the data collection strategy should be informed by
the specific questions that researchers (or policymakers) might have. If one wants to
extrapolate the results obtained in a given context to a different environment, or change
details of an intervention, it will be useful and necessary to collect information on the
drivers of individual behavior that might affect the outcome of interest.
In Attanasio et al. (2020b), for instance, the objective is to establish what mecha-
nism generates the observed impact of a stimulation intervention on child development,
with particular emphasis on the hypothesis that parental investment (in time and ma-

24
terials) played an important role. To find an answer to this question it is necessary to
identify the causal link from parental investment to child development, in addition to
the overall impact of the intervention on child development. One cannot use the treat-
ment as a valid instrument to identify such a causal link, as the hypothesis of interest is
whether the intervention has (or does not have) an impact on child development (directly
or through other mechanisms). Attanasio et al. (2020b) use variation across different
towns in prices of toys and other items as well as exposure to violence of mothers at
the time when they were adolescents as determinants of parental investment that are
assumed not to affect child development directly. This assumption (and the availability
of the relevant measures), allow then to identify the relevant causal links and perform a
mediation analysis that takes into account the fact that some of candidate mediators are
variables determined by choice.

The general point we want to make is that the design of surveys should be informed
by the specific needs and research questions that are being addressed. Survey instru-
ments should use a variety of measurement tools and should not be restricted to col-
lect information about individual choices. Identification of structural models could be
achieved with much weaker assumptions by combining standard measures with, for
instance, the elicitation of respondents’ responses in hypothetical situations.22 In our
application in Section 6, we discuss how the empirical study of the model we present
in Section 4 is made easier and more interesting by the construction of a different set of
variables, some of which require the development of new measurement tools.

4 Modeling parental investment

In this section, we consider a specific application of the ideas we have discussed so
far, sketching a theoretical model of household decision-making where some of its key
components are often not directly observable. The main aim is to show how a number
of variables that are not typically measured in standard surveys can be useful to identify
the determinants of parental investment.
The framework we present is based on the collective model (Chiappori, 1988),
which suggests that the household utility function is a weighted sum of the mothers
22 The example of Juster and Shay (1964) is again a good one here.

25
utility function, U m (·), and the father’s utility function, U f (·):

U = µ(p,Y ; z)U m (c, H(x)) + (1 − µ(p,Y ; z))U f (c, H(x)), (5)

where c is a vector of consumption of both private and public goods, p is the corre-
sponding vector of prices, and H is child human capital, dependent on investment in
child development, x. µ(·), often referred to as Pareto weights, represents the relative
importance of the problem in equation (5) given to the mother’s utility. The defining
feature of this model is that, household choices, resulting from the interactions of dif-
ferent decision makers with possibly different preferences, are efficient, in that they
maximize the function in (5), given certain constraints and a set of Pareto weights.
The model is silent about determining the Pareto weights. They are allowed to
depend both on variables that enter the problem through the budget constraint, such as
prices and income, and, crucially, other variables z, often labelled distribution factors.
These are variables that matter for the sway in household decision-making but do not
influence the budget constraint or utility functions directly (Browning et al., 2013).
The household faces two constraints: a standard (static) budget constraint and a
production function of child development. In particular we assume that children human
capital depends on initial conditions, H0 , parental investments, the financial investment
in education, x, and other (unobserved) factors ε:

H = f (H0 , x, ε). (6)

If the ‘household’ maximizes the utility function in equation (5), subject to the con-
straints we mention, parental investment will depend on household resources,on indi-
vidual preferences as aggregated into household preferences by the Pareto weights µ(·),
and the properties of the assumed production function. We notice, however, that for the
determination of parental investment, the properties of the production function of child
development are not necessarily key, unless they coincide with those of the production
function as perceived by the parents. In the presence of potentially distorted beliefs
about the process of child development, what matters is the perceived productivity of
parental investment, which can be different for husband and wives.23
23 Note also that this framework can easily be extended to child development being a function of both
material investment, x, and time investment, e.g., reading or talking to the child (see e.g., Attanasio et al.

26
 A simple parametric example can make this framework clear and help us to relate
it to the empirical exercises in Section 6. We assume that there are q private goods for
husband and wife and that the only public good is child human capital. We also assume
that both the individual utility functions and the production function of human capital
are Cobb-Douglas (CD). While the commodities we consider are private, we allow the
consumption of each spouse to affect the utility of the other spouse. Therefore, we have:

q
i
α ijm lnC jm + α ij f lnC j f + αki ln H k , i = {m, f };


lnU = ∑ (7)
j=1

where C jm is the consumption of commodity j consumed by the wife, and C j f is the

consumption of commodity j consumed by the husband. α ijm and α ij f are the param-
eters of the CD function corresponding to the consumption of commodity j by each
of the two household members in the utility function of member i; i = m, f . Analo-
gously, αki determines the utility derived by member i from the child’s human capital.
We assume that the utility function for the collective household is:

U = µU m + (1 − µ)U f . (8)

The perceived production function for human capital is given by:

ln H k = γ0i ln H0 + γ i ln X, i = {m, f }; (9)

where γ0i and γ i , i = m, f are the parameters of the production function as perceived by
the husband and wife. Finally, the budget constraint, with prices normalized to 1 for
notational simplicity, is:
q
X + ∑ (C jm +C j f +C jk ) = Y. (10)
j=1

It is easy to show that the household parental investment function is given by:

f
µU m αkm γ m + (1 − µ)U f αk γ f
X= Y, (11)
A
where the denominator is given by:
(2020b,c); Cunha and Heckman (2008); Heckman et al. (2013, 2020); and Todd and Wolpin (2003)).

27
 ! !
m f f f
A = µU ∑ (α m m m m
jm + α j f ) + αk γ + (1 − µ)U f
∑ (α jm + α j f ) + αk γ f . (12)
j j

While the expression for parental investment in equation (11) is particularly simple
because of the CD assumption (which, for instance, implies unit elasticity and constant
shares for all commodities), the expression illustrates the role played by the preference
parameters (the α iji0 , i, i0 = m, f ), the Pareto weights, and the ‘perceived’ production
function γ i ’s. We note that mother’s and father’s individual preferences for each private
commodity and child human capital are mediated by the Pareto weights given by µ.

The Engel curve of parental investment. While the CD assumption makes deriva-
tions very simple, the implied homotheticity is obviously not a plausible assumption.
One possibility is to use a generalization of Deaton and Muellbauer (1980)’s AIDS sys-
tem for Engel curves, as done, for instance, by Browning and Chiappori (1998). In
this case, household i’s budget shares for investment in human capital (si = Xi /Yi ) is a
function of parental preference parameters, parental distribution factors, prices and total
household expenditure, allowing the expenditure elasticity to be different from 1:

si = G(τi , µi , γi , p) + β (lnYi − a(p)) + u, (13)

where G is a function that depends on the main determinants of the parental problem
previously presented: parents’ tastes for child human capital relative to other com-
modities, represented by two latent factors τi , parental beliefs about the productivity
of parental investment in the production of human capital (again represented by two
latent factors γi ), the relative bargaining power within the couple µi , the vector of rela-
tive prices, and, as in the AIDS system, log real income lnYi − a(p) and an unobserved
shock u. It is also possible to make both the slope coefficient, β , and the price index,
a(p), functions of preference, τi , beliefs, γi , and bargaining power, µi , and to extend
this model to a Quadratic Almost Ideal Demand System (QUAIDS), as in Banks et al.
(1997).
We will get back to these issues when, in Section 6, we estimate a simple version
of equation (13) for investment in child human capital using measures on parental al-
location preferences, parental decision-making power, and beliefs, in conjunction with

28
actual choice data we collected in Tanzania. Some implications of this model, how-
ever, are clear. The share of expenditure on child development will be larger the higher
the perceived returns to investment, and the more utility parents derive from child hu-
man capital over private consumption. If wives have a stronger preference for child
development over private consumption than husbands, the investment share should be
increasing in the weight wives have in decision-making. The opposite would hold true
if husbands have a stronger preference for human capital than their wives.

5 New and old measures: using them jointly

In this section, we briefly describe a novel survey that was collected explicitly to pilot a
number of new measures. In Section 6, we use (some of) these measures to get an em-
pirical characterization of the model of parental investment we have sketched in Section
4. Here we provide a description of the different measures and how they were obtained.
While the sample we use is relatively small, this data set is indicative of the possible
use of a combination of new measures and choice-based data to estimate models of in-
dividual behavior. The description we provide in this section is also indicative of the
difficulties that come with collecting new measures and the importance of a direct link
between theory and the abstract constructs used in researchers’ theoretical frameworks.

5.1 A pilot survey in Tanzania

The data we use in this and the next section were collected in two districts of the largely
rural Kagera region in Tanzania. The overall goal of this data collection was to improve
the measurement of child development and its drivers and design new measures. The
survey included three samples: in the first, the respondents were mothers, in the sec-
ond, fathers and, in the third, couples. The different samples are useful to characterize
differences across mothers and fathers and how joint (couple) decisions are made. We
describe the data and their collection in detail in Appendix A2. Here we focus on the
novel measures. Unfortunately, the sampling strategy made the three samples some-
what unbalanced, in particular with respect to age. However we still find it useful to
compare them here for illustration purposes.

29
Bargaining Power. The power that women have within households has received much
attention, especially when analyzing models of intrahousehold allocations. In our Tan-
zania sample, a first measure related to the bargaining power within the couple repli-
cated the approach used by Almås et al. (2018), who conducted a controlled (“labo-
ratory”) experiment in a sample collected within an RCT in North Macedonia. The
measure was designed to capture the potential impact of targeting women rather than
men with a Conditional Cash Transfer, which was given to women in some villages
and to men in others. To capture the bargaining power latent factor, after the initial
data collection, the wives were called to an office to run an incentivized experiment.
They were told: “Here are 100 Denars that we will give to your husband. How much
are you willing to pay to have them paid to you?” This amount, which is completely
independent of the government-administered cash transfer, was actually paid out (to the
husband or the wife, depending on her choice).24 A second hypothetical question was
asked considering much larger amounts. The idea is that women willing to sacrifice a
higher proportion of the amount offered are less powerful within the couple.
Almås et al. (2018) show that such a measure of bargaining power, which we label
Willingness To Pay (WTP), correlates with a number of observables in a predictable
manner. Moreover, in villages where the government grant was targeted at wives, the
WTP decreased significantly; in these villages, women were willing to pay less to get
control of any additional transfers. The incentivized and the hypothetical exercises
in Macedonia yield similar results, indicating that hypothetical formulations of these
questions could also be used in surveys.
In the Tanzanian sample, a hypothetical version of the WTP elicitation was con-
ducted. Furthermore, unlike in North Macedonia where the question was asked only
to wives, in Tanzania it was also asked to husbands in the fathers sample. In Table 1,
we report the average share of the 6600 TSH that the respondents (i.e., the wives in
the mothers and couples samples and the husbands in the fathers sample) were willing
to forfeit so that the payment would be made to them rather than their spouse. As in
Almås et al. (2018), we interpret this share as being inversely related to the control of
resources the respondent has on the household resources, which in turn can be related
to the Pareto weights of the collective model discussed in Section 4.
24 100 Denars corresponds, for this sample, to two days of paid work.

30
 Table 1: Willingness to pay (out of 6600 TSH)

Wives Husbands
Mothers and Couples samples Fathers sample

Average share 0.320 0.100

Median share 0.061 0.008
Std. Dev. 0.394 0.258
Observations 215 98
Note: The table displays the descriptive statistics for the Willingness To
Pay (WTP) variable. The WTP is measured as the share of the 6600 TSH
that the respondents (i.e., the wives in the mothers and couples samples
and the husbands in the fathers sample) were willing to forfeit so that the
payment would be made to them rather than their spouse. In the mothers
sample, the statistics are computed over the households where the father is
present. Source: Tz Pilot.

The distribution of WTP is very skewed, with a few observations with very high
values. This feature of the distribution is reflected in large differences between median
and mean of the distribution. We observe a considerable difference between husbands’
and wives’ WTP, with wives willing to sacrifice on average 32% of the transfer to get
control over it. The median wife, both in the mothers and couples samples, is willing to
pay just over 6% of the transfer. Husbands, on the other hand, are willing to pay only
10% on average, with the median being 0. As we mention above, the three samples we
are considering (in particular the fathers sample) are not strictly comparable, because
of the age differences induced by the sampling design. However, the difference in the
WTP is very marked and probably reflects different bargaining positions within the
marriage, indicating that men have more control over resources than women.
Translating the WTP measure into a measure of bargaining power within the couple
is not simple. To perform such an exercise, it would be necessary to identify a mapping
from the theoretical constructs of the collective model, such as the Pareto weights, and
individual preferences with the possibility of altruism and public goods or some of its
intermediate outcomes, such as sharing rules, to the WTP measure.
Existing surveys provide alternative information on resource control within the house-
hold through a number of questions that explicitly ask who in the households is respon-
sible for decisions on expenditure on various commodities and other decisions, with

31
the possible answers being ‘the husband’ ‘the wife’ or ‘both’. In the Tanzania survey
we are using, a number of questions, taken from the Tanzania Demographic and Health
Survey (DHS), were posed to the respondents. The questions were about major house-
hold expenditures, children’s education, health expenditures, what food to cook, and
whether the wife can go out.
In particular, the respondents were asked who is mainly responsible for a number of
decisions, including major household expenditures, children’s education, health expen-
ditures, what food to cook, and whether the wife can go out. To avoid the often-observed
bunching around the ‘both’ answer, when this option was chosen the respondent was
asked who had the final say.

Table 2: Measurement system for decision making

Decision making
Wives Husbands
Mothers and Couples sample Fathers sample
β α β α
Own health 1.000 0.000 1.000 0.000
- - - -
Children’s health 1.283 -0.171 1.213 -1.140
(0.272) (0.170) (0.484) (0.249)
Children’s schooling 1.167 -1.261’ 1.021 -0.680
(0.381) (0.181) (0.419) (0.194)
Household purchases 0.662 -0.988 1.506 -0.386
(0.219) (0.132) (0.638) (0.242)
Cooking 0.278 0.904 0.563 -1.039
(0.120) (0.117) (0.296) (0.167)
Visiting 0.639 -0.932 0.582 0.235
(0.198) (0.129) (0.302) (0.138)
Factor mean 0.468 -0.217
Factor variance 0.882 0.402
Note: This table shows the loading factors (β ) and the intercept (α) of an IRT estimated
from questions about who is mainly responsible for a number of decisions within the house-
hold, including: major household expenditures, children’s education, health expenditures,
what food to cook, and whether the wife can go out. Columns (1) and (2) are for the moth-
ers and couples sample, and Columns (3) and (4) for the fathers sample. Standard errors in
parentheses. Source: Tz Pilot.

Using the answers to these questions in the mothers and couples samples (where the
wife is the respondent) and the fathers sample (where the husband is the respondent)
we estimate a measurement system and, from that, the latent factor of interest, reflect-
ing women’s decision power in the couple. The estimates of the measurement system

32
parameters are reported in Table 2.
We can now relate the WTP measure to the ‘decision making’ factor extracted from
more traditional measures of control of resources. In particular, in Table 3, we report
the results of a regression of the WTP on the decision making latent factor estimated
first in the sample of mothers and couples and then on the sample of fathers.
We limit the sample of mothers to the households where the husband is present and
allow, in the first regression, an intercept shift for the couples sample. We find that in the
mothers and couples samples, the two variables are significantly and negatively related
(the higher the share respondents are willing to forfeit to get control of the payment,
the less the mothers’ decision making power within the household), and the couples
sample shift is not significant. The R-squared of this regression, however, is very low,
indicating that there is a considerable amount of variation in WTP which is not related
to the traditional measures of control. A similar result applies to the sample of fathers,
although the coefficient is not significant.

Table 3: Willingness to pay and decision making factors

1
Willingness to pay
Wives Husbands
Mothers and Couples samples Fathers sample
Decision making factor -0.079∗∗ 0.071
(0.033) (0.054)
Couple -0.041
(0.054)
R-squared 0.030 0.018
Observations 215 98
Note: The table displays regressions of the willingness to pay on the decision making
variable and a dummy for the couple treatment arm of the experiment. Standard errors in
parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Source: Tz Pilot.

Jayachandran et al. (2021) have used an approach similar to the one we have used to
derive the WTP described above in India. They also use a machine learning algorithm
to identify questions that accurately reflect women’s decision power in the couple. They
conclude that, in the Indian context, the latter approach seems to work better.

33
Beliefs on returns to parental investment. Another important driver of individual
choices that we consider in our application is parental perception about the process of
child development. As is clear from the model presented in Section 4, parental invest-
ment is driven by parental perception of the return on child development. While much
of the existing literature assumes that parents know the process of child development
and how it depends on the child’s current development, parental investment, and possi-
bly other factors, it is increasingly clear that these perceptions might be distorted.
Several studies, which we mention in Section 3.2, have started eliciting beliefs about
the relationship between parental behavior and child development. We use an approach
similar to Attanasio et al. (2019b), who measure mothers’ beliefs about the process of
child development within a survey of an RCT evaluating a parenting intervention.
The approach consists of presenting mothers with scenarios in terms of initial con-
ditions and investment and asking them to map these scenarios into child development
outcomes. The implicit assumption is that mothers use the same mapping between la-
tent factors and observable markers, so the scenarios proposed in the questionnaires
have a relation to the latent factors that researchers want to capture. This approach
allows researchers to estimate perceived rates of return to parental investments under
different initial conditions.
In the Tanzania sample, we use this approach to elicit beliefs – for both fathers
and mothers – about different aspects of the developmental process and the importance
of certain parental inputs. We use the answers to the beliefs questions to estimate a
measurement system and extract a ‘beliefs factor’, which we can use to estimate the
perceived return to parental investment. The estimates of the measurement systems are
reported in Table A2 in Appendix A3. The returns we consider for high and low levels
of initial development are measured as the difference in the expected outcomes between
high and low levels of investment for the two levels of initial conditions.
In Figure 1, we plot the distribution of expected returns under low and high initial
conditions for the whole sample as well as for the mothers and fathers sub-samples. Fig-
ure 1 also reports a test for the difference of the distribution means and a Kolmogorov-
Smirnov (KS) test for the difference between the two distributions. For beliefs about
language development, returns to parental investment are perceived to be higher for low
than for high initial conditions in the whole sample and for mothers (in the mothers and

34
 couples samples): the difference in means for the whole sample is equal to 0.140 (p-
value=0.000) and to 0.200 (p-value=0.000) for mothers’. For the fathers sample there
is no significant difference between the returns with high or low initial conditions: the
point estimate of the difference is 0.030 (p-value=0.570).

Figure 1: Beliefs on language development: Returns to parental investment

Whole Sample Mothers Fathers
1.5

1.5

1.5
Difference in means p-value = 0.000 Difference in means p-value = 0.000 Difference in means p-value = 0.570
kmirnov p-value = 0.000 kmirnov p-value = 0.000 kmirnov p-value = 0.200
Density (kernel=epanechnikov)

Density (kernel=epanechnikov)

Density (kernel=epanechnikov)
1

1
.5

.5

.5
0

0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Differences H-L Investment for Average Words Differences H-L Investment for Average Words Differences H-L Investment for Average Words

Low Initial Condition High Initial Condition Low Initial Condition High Initial Condition Low Initial Condition High Initial Condition

The first two facts are qualitatively consistent with the evidence from Colombia re-
ported in Attanasio et al. (2019b): poor parents seem to think that parental investment
is more productive and effective at low levels of initial development. More generally,
the entire distribution seems to be different, with the low initial condition returns pre-
senting more dispersion and shifted to the right. Mothers have a higher expected return
to investment for low initial condition children than fathers, as the difference is equal to
0.175 (p-value=0.000). There is no significant difference between fathers and mothers
on expected returns for high initial condition children.
In addition to the beliefs about cognitive development, we perform a similar exercise
to measure beliefs about the effect of parenting on socio-emotional development, a type
of belief that has not been measured before. In Figure 2, we report the distributions
of expected returns from different initial values in the whole sample and, as for beliefs
about language development, for the mothers and fathers samples separately. As in
Figure 1, we report KS tests for the difference between the two distributions and a test
for the difference between the low and high initial conditions distributions.
The results are similar to those on the beliefs about the effect of parenting on cog-
nitive development and language. In the whole sample, returns to parental investment
on socio-emotional development are perceived to be higher for low than for high ini-
tial conditions: the difference in means for the whole sample is equal to 0.044 (p-

35
 Figure 2: Beliefs on socio-emotional: Returns to parental investment
Whole Sample Mothers Fathers
Difference in means p-value = 0.150 Difference in means p-value = 0.010 Difference in means p-value = 0.240
2

2
kmirnov p-value = 0.000 kmirnov p-value = 0.000 kmirnov p-value = 0.260
Density (kernel=epanechnikov)

Density (kernel=epanechnikov)

Density (kernel=epanechnikov)
1.5

1.5

1.5
1

1
.5

.5

.5
0

0
-3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3 -3 -2 -1 0 1 2 3
Differences H-L Investment for Average Socioemotional Differences H-L Investment for Average Socioemotional Differences H-L Investment for Average Socioemotional

Low Initial Condition High Initial Condition Low Initial Condition High Initial Condition Low Initial Condition High Initial Condition

value=0.150). This effect is coming from mothers, where the difference is equal to
0.097 (p-value=0.010). We see no such effect in the fathers sample, where the differ-
ence in means is equal to -0.063 (p-value=0.240).

Measuring preferences. As discussed above, an important determinant of individ-

ual choices is preferences. While, under a number of assumptions, the structural pa-
rameters that characterize them can be estimated from choice data, in some contexts,
answers to hypothetical questions can be particularly useful. In a recent paper, Ameriks
et al. (2020) use stated hypothetical choices elicited with Strategic Survey Questions
(SSQs) to estimate some preference parameters in a fully specified structural model.
The authors note that to develop direct measures of preferences, it is desirable to de-
velop survey instruments that allow respondents to provide information that identifies
preference parameters in a language they are comfortable with as well as in a format
that allows for a precise mapping to the structural parameters of interest.
In studying the model of parental investment in Section 4, in which preferences
of different decision makers interact to determine household choices, it is important
to elicit not only preferences on the allocations of scarce resources among different
commodities but also between different individuals within the family. Cherchye et al.
(2021) use a similar approach in Kenya to elicit tastes for couples’ allocation between
child investment and own consumption.
In the Tanzania sample, before the standard survey, a key respondent or respon-
dents (the wife, the husband, or the couple jointly) were asked to allocate a hypothetical
amount, represented by a pile of beans, between different expenditure categories and

36
between three individuals (the husband, the wife, or the child in the household).25 The
six possible expenditure categories considered were: clothing, food, learning materials
(such as books, notebooks, and pens), health expenditures, transportation, and school
expenditures. These categories were chosen to be able to match the information col-
lected on actual expenditure. While the question was not explicit about this issue, we
interpret the answers to the hypothetical questions as referring to the allocation of addi-
tional resources that the household would normally have access to.

Table 4: Average share of expenditure allocated to household members

Mother Father diff Couple diff

decision decision decision
(s.e.) (s.e.) (p-value) (s.e.) (p-value)
To self 0.268 0.257 -0.01 0.250 -0.02
(0.008) (0.007) (0.348) (0.009) (0.148)
To spouse 0.175 0.219 0.04 0.232 0.06
(0.009) (0.007) (0.000) (0.009) (0.000)
To child 0.558 0.524 -0.03 0.518 -0.04
(0.011) (0.013) (0.043) (0.013) (0.019)
Note: This table shows the average share of expenditure to household
members for the different samples. Standard errors are in parentheses.
The p-values refer to the test of the difference between the mothers and
fathers samples and the mothers and couples samples. Source: Tz Pilot.

In Table 4, we report the share of the total additional resources allocated to each
individual in the family (spouse, child, and self) for each of the three samples. A number
of interesting results emerge from this exercise. First, despite the good considered being
private, both mothers and fathers allocate some resources to their spouse, indicating
that the participants care about their spouse’s consumption. Second, mothers allocate
more than fathers to children, but allocate the same share to themselves, which implies
mothers allocate less to their spouses than fathers. Third, the couple’s decisions seem
much closer to those of fathers than those of mothers. The similarity between the fathers
and the couple allocations might indicate large differences in decision-making power
25 The question was posed as: “We would now like to understand how you would prefer to spend 300k
Shillings, if we were to give this money to you. Use these 60 beans, each representing 5k Shillings,
and this cardboard card with 3 different expenditure options (mother, father, and your child); for each
question distribute the beans according to your preferences. Imagine that your child is 5 for this exercise.”
See Almås et al. (2020a) for a full description of the protocol followed.

37
between spouses as fathers, consistent with the evidence on the WTP, hold considerably
more decision power within the household.
Next, we look at the allocation among different commodities, and in particular,
for the resources allocated to the child. In Table 5, we report the shares allocated to
the child split into the six different commodities. There are some differences between
mothers and the other samples, particularly in clothing and health (where the mothers’
shares are significantly higher) and in learning materials (where the fathers’ share is
marginally higher). Once again, the couple’s decisions are more similar to those of
fathers than those of mothers.

Table 5: Allocation to the child

Mother Father diff Couple diff

decision decision decision
(s.e.) (s.e.) (p-value) (s.e.) (p-value)
Clothing 6.628 5.559 -1.07 5.493 -1.13
(0.225) (0.311) (0.005) (0.195) (0.000)
Food 6.062 5.338 -0.72 4.401 -1.66
(0.302) (0.269) (0.076) (0.251) (0.000)
School exp. 7.434 7.529 0.09 7.282 -0.15
(0.353) (0.573) (0.886) (0.456) (0.791)
Learning mat. 5.503 5.213 -0.29 5.697 0.19
(0.247) (0.285) (0.441) (0.317) (0.629)
Health exp. 5.159 5.213 0.05 5.761 0.60
(0.207) (0.252) (0.866) (0.232) (0.054)
Transportation 2.683 2.603 -0.08 2.430 -0.25
(0.182) (0.202) (0.769) (0.199) (0.349)
Notes: This table shows the descriptive statistics of allocation of expendi-
ture on children. Standard errors are in parentheses. The p-values refer to
the test of the difference between the mothers and fathers samples and the
mothers and couples samples. Source: Tz Pilot.

In the model we discussed in Section 4, it was clear that the relative parental taste
for child human capital and alternative allocations of resources is a key determinant
of parental investment. While we do not estimate a structural version of that model,
as it would be implied, for instance, by the AIDS version of Engel curves in equation
(13), we use the answers about stated preferences in the couple samples to derive some
information about individual couples’ tastes and relate them to parental investment.
As is evident from equation (11), which was derived under homothetic preferences,

38
parental investment can depend on the shares of resources allocated to adult goods
relative to the resources allocated to child consumption and parental investment.
To estimate a factor representing the taste for child human capital, we use the an-
swers to the allocation questions to construct eight variables as the ratios of the re-
sources allocated to each spouse for four adult commodities (food, clothing, health, and
transportation) to the resources allocated to the child and estimate a factor model to
extract a latent factor we label relative taste for child human capital from these eight
variables. We perform this analysis in each of the samples and report the loading factors
for each of the variables, as well as the intercept for the equations of the measurement
system corresponding to each observable variable, in Table A3 in Appendix A3. The
factor estimated from this analysis is what we use in Section 6 to model parental in-
vestment. From Table A3, several variables seem to be important markers of the taste
for child development. Furthermore, we find important differences in the three sam-
ples, with the preferences in the couples sample seeming more similar to the fathers’
rather than the mothers’. In the latter, measures of the mother’s and father’s health
expenditures and father’s food expenditures are not particularly important.
Having presented some evidence on the new measures that were collected in Tan-
zania, one important issue and challenge is their validation. In particular, we check if
these measures co-vary in a sensible way with choice data or, more generally, with stan-
dard measures. This would be a first step towards a systematic use of these measures
within models of individual behavior. We turn to this in the next section.

5.2 Using available measures efficiently

Even when the latent factors that one wants to analyze are reasonably standard, it may
not be easy to obtain meaningful measures from the available raw data that can be used
in empirical analysis. In many contexts, certain algorithms that aggregate available
measures become the accepted standard and are widely used, even when alternative
approaches may be more efficient and meaningful. As discussed in Section 3.3, the
explicit specification of a measurement system can be interpreted as the construction of
a scoring mechanism that efficiently uses the available data to estimate the latent factors
of interest. The study of child development and its drivers, such as parental investment
or school quality, provides a particularly salient example that has received substantial

39
attention in recent years.
It is recognized that measuring child development is difficult, especially in the early
years and when one wants to assess different dimensions of development, including so-
cioemotional skills. Analogous considerations apply to measures of the drivers of child
development, such as parental investment and school quality. These difficulties are even
more serious in developing country contexts, both because the administration of some
of the most frequently used tests is often difficult and requires specialized testers and
because many of these tests were developed and validated in what has been termed
Western, Educated, Industrialised, Rich, and Democratic (WEIRD) samples (Henrich
et al., 2010).26 While a number of sophisticated tests have been developed and validated
in developed countries, the relevance and effectiveness of such tests in completely dif-
ferent contexts might be limited.
Because of these considerations, a number of efforts to develop a new generation of
child development tests are under development.27 Within the Tanzania project we have
been discussing, a new test was also developed, described in Attanasio et al. (2022),
which combined elements from different well-established tests (such as the Bayley
Scales of Infant and Toddler Development, Third Edition - Bayley-III, the Caregiver
Reported Early Development Instruments (CREDI), and others) to construct an efficient
and easy to administer test with a limited number of items. The approach consisted of
estimating a measurement system such as in systems (2) and (3), relating the dimen-
sions of interest to the various elements that make up these tests. This procedure, now
widely used (see, for instance, Cunha et al. (2010), Agostinelli and Wiswall (2017),
Heckman et al. (2020) among others) can be seen as an effective alternative to the use
of standard algorithms that typically come with these tests. The construction of a new
scoring algorithm through the estimation of a measurement system obtained in a given
context is an effective way to adapt the existing tests to new realities and make different
26 For concepts such as parental investment and school quality, the application of tests developed in

different contexts can easily result in flooring and ceiling effects.

27 The UN has facilitated the development of the Multiple Indicator Cluster Survey (MICS), which

has recently been discussed in Bornstein et al. (2021). The Gates Foundation has funded a large effort
to pull together a number of new indicators that could be comparable across contexts and countries and
would be relatively easy to administer. This has given rise to the Global Scale for Early Development
(GSED) initiative, discussed in GSED (2021) and Black et al. (2019). McCoy et al. (2021) describe the
construction and use of the CREDI questionnaire.

40
contexts comparable.
Attanasio et al. (2022) use the estimates of the measurement system to identify the
most informative items to measure different dimensions of child development. Attana-
sio et al. (2022) show that the use of the most informative items yields estimates of child
development that contain virtually the same information as the complete tests and are
much cheaper and quicker to collect, especially in developing countries.28

Table 6: A measurement system for parental investment

Couples sample
β α
Parental Activities 1.000 0.000
- -
Play Material 0.362 -0.201
(0.136) (0.086)
Material Investment 1.007 0.201
(0.377) (0.176)
Expenditure on 0.351 -0.080
children share (0.128) (0.070)
Social scale 0.657 0.181
(0.329) (0.173)
Didactic scale 0.415 0.031
(0.235) (0.136)
Factor mean (variance) -0.376 (0.436)
Note: This table shows the loading factors (β ) and
the intercept (α) for markers of parental investment.
Columns (1) and (2) are for the couples sample. Stan-
dard errors in parentheses. Source: Tz Pilot.

A similar set of issues is relevant for measures of latent factors that are key drivers
of child development, such as parental investment. It is not always clear how many
dimensions of investment to consider and how to adapt available measures to differ-
ent contexts.29 A number of standardized measures exist and have been widely used,
28 As we do not use these data in the exercise in Section 6, we do not report here the details of these
results and refer the reader to Attanasio et al. (2022).
29 Researchers use different strategies to measure parental investment. Attanasio et al. (2020b), for in-

stance, consider time and material investment separately and show that they might have different impacts
on child development. Others, such as Cunha et al. (2010), consider a single dimension.

41
such as the Home Observation Measurement of the Environment (HOME) index or the
Family Care Indicators (FCI). However, given the set of items that make up these tests,
it is not clear that the same scoring algorithm should be used in different contexts, as
different items might be differently salient and relevant depending on the context.
In the Tanzania context considered here, we follow an approach to measuring child
development similar to that taken by Attanasio et al. (2022) and estimate a factor model
that identifies a single latent factor. We use this factor as our measure of parental invest-
ment in Section 6. In particular, we use a number of items as markers of parental in-
vestment, including: (i) time spent in activities with children; (ii) play materials present
in the house; (iii) material investments for the child (including food, clothing, footwear,
confectioneries, among others); (iv) share of expenditure on children items over the
total expenditures of the household; (v) items from the social scale from the Parental
Style Questionnaire (PSQ; Bornstein (1996)); and (vi) items from the didactic scale
from the PSQ. In Appendix A3.3 we define (i)-(vi) and Table A4 reports descriptive
statistics on each components of parental investment.
In Table 6, we report some of the estimates of the parameters of systems (3) and
(4) for the sample of couples. We notice that several markers are relevant to the factor
we are considering. Different measures of parenting skills, for instance, do not seem
particularly relevant for the parenting investment factor, while material investment plays
an important role.

6 A model of parental investment

In the model we presented in Section 4, the Engel curve for parental investment can be
seen as an approximation similar to the one used by Browning and Chiappori (1998). In
this section, we estimate a version of equation (13) for parental investment using both
choice data on investment and total expenditure, as well as the measures of preferences,
beliefs, and bargaining power that we collected and discussed in Section 5. Rather
than estimating a detailed version of the model that incorporates tastes for a variety of
different commodities, we focus on parental investment and summarize the information
on couples’ preferences for child development relative to other uses of resources. We
want to investigate how our novel measures co-vary with actual behavior in terms of

42
parental investment in a way which is consistent with the model we presented.
We presume that in the sample where the non-conventional questions are directed
to the couple, the answers reflect the ‘aggregated’ couple preferences and beliefs about
child development. Therefore, for this exercise, we use only the couple samples, where
questions are directed to the couple, as it would be hard to model observed investment
by the couple on the basis of the preferences and beliefs of only one partner.
We estimate different measurement systems to extract from the available measures
information about factors that enter the model in equation (13). In Section 5, we discuss
the systems we use to extract the latent factors representing couples’ relative tastes
for child development, their beliefs about the productivity of parental investment, and
bargaining power within the couple as well as actual parental investment. The estimates
of these measurement systems are reported in Appendix A2.
In equation (13), the share of parental investment in total expenditure depends on
(the log of) total expenditure and the latent factors representing both spouses’ tastes and
beliefs as well as a bargaining power factor, which aggregates the individual factors. As
we interpret the taste and beliefs factors elicited in the couples sample as reflecting the
couples’ preferences and beliefs, one can modify equation (13) as:

si = G(τi (µi ), γi (µi ), p) + β (lnYi − a(p)) + ui (14)

where τi and γi are now unidimensional factors that aggregate (through the bargaining
power factor µi ) the tastes and beliefs of the two spouses. If we interpret the responses
from the couples sample as reflecting these aggregate factors, the bargaining power
should not enter the parental investment equation once we control for the aggregated
factors. However, it is possible that the linear specification that approximates equation
(13) is too restrictive so that the bargaining power factor could enter it significantly.
In Table 7, we report the results of regressing the parental investment factor on: log
total expenditure, the beliefs and taste factors (and their interactions), and our measure
of bargaining power. As parental investment is a factor with an arbitrary scale, the size
of the coefficients in Table 7 is difficult to interpret. However, we notice that (log) total
expenditure attracts a positive and significant coefficient in all specifications.
In column 1, we only use choice data, in that we regress the investment factor (which
is determined by investment in time spent with children and expenditure in commodi-

43
ties targeted to children) on the log total expenditure only. In column 2, we introduce
our estimates of taste and beliefs questions to capture the determinants of parental in-
vestment as function G() in equation (14). We observe that the relative taste for other
commodities is strongly significant and with the expected negative sign. This result can
be partly interpreted as a validation of the measure of taste that we use, which is derived
from an experiment on hypothetical allocations; the specific experiment does not refer
to actual investment choices at all. Analogously, the factor measuring beliefs about the
productivity of investment is also strongly significant with the expected positive sign in
the equation for parental investment.

Table 7: Modeling expenditure: Couples sample

(1) (2) (3) (4) (5)
Exp share Exp share Exp share Exp share Exp share
Log of total expenditure -0.097∗∗∗ -0.082∗∗∗ -0.080∗∗∗ -0.080∗∗∗ -0.079∗∗∗
(0.030) (0.027) (0.027) (0.027) (0.027)
Bargaining Power (BP) -0.008 -0.009 -0.009 0.054
(0.062) (0.063) (0.064) (0.105)
Relative taste for child human capital (RC) -0.137 -0.143 0.146
(0.179) (0.331) (0.421)
Beliefs 0.035 0.032 0.040
(0.121) (0.198) (0.122)
Beliefs*RC 0.023
(1.118)
BP*RC -0.362
(0.488)
R-squared 0.142 0.202 0.208 0.208 0.211
Observations 142 126 126 126 126
Note: The table displays regressions using the factor of expenditures share as a left hand side variable and the following right
hand side variables: i) Log of total expenditure is the logarithm of total household expenditure from the household survey;
ii) Bargaining power is a measure of female bargaining power, and more precisely, it is 1 minus the share that the woman is
willing to pay to gain control over a fixed amount of money from our experiment; iii) Relative taste for child human capital is
a factor of ratios of the resources allocated to each spouse for four adult commodities to the resources allocated to the child in
our allocation experiment; and iv) Beliefs is a factor of the returns to parental investment on Language and Socio-emotional
skills elicited in our survey. All models are controlled by child and mothers characteristics: child’s age and gender, number
of siblings in the household, dummy for mother’s secondary education and mother’s cognition skills measured by the Raven’s
test. Standard errors in parentheses. ∗ p < 0.10, ∗∗ p < 0.05, ∗∗∗ p < 0.01. Source: Tz Pilot.

In column 3, we add to the variables in column 2 the bargaining power factor. As

mentioned above, if we interpret the answers given by the couple’s questions on their
tastes and beliefs as representing the aggregation of the individual tastes and beliefs,
(and if the equation is linear), such a variable should not be a significant determinant
of investment. We find that, while the other coefficients do not change much with the
introduction of this variable, representing the wife’s power within the couple, takes a
negative coefficient which is significantly different from zero. Without a further detailed

44
analysis of the various factors at play, it is difficult to interpret this coefficient.
As the significance of the bargaining power factor might be signal the presence
of nonlinearities in the function G() in equation (14), in columns 4 and 5 of Table
7, we introduce interactions of the taste factor with beliefs (in column 4) and with
bargaining power (in column 5). In neither of the two columns do we find any significant
interactions between taste and the two variables considered.
This exercise is a first step in utilizing novel measures in a unifying framework that
combines elicited beliefs, preferences, and decision-making power with observational
data. This research’s next step is to introduce these measures into a structural model of
parental behavior and map them more directly onto its parameters. To achieve such a
goal, it may be necessary to develop finer measurement tools than those used here.

7 Conclusions
In this paper, we have analyzed the role that measurement does and should play in
economics and its relation to economic theory. We argued that measurement issues –
including what should be measured, how to construct effective measures of the latent
factors that populate economic models, and how to use such measures – should be in-
formed by economic theory. Economic models are attempts to describe certain aspects
of human behavior in specific contexts in a coherent fashion that allows generaliza-
tions, extrapolation, and ultimately the identification of causal links between different
variables. Depending on what is being modeled or ‘explained,’ bringing the relevant
latent factors to data might require the measurement of different variables.
Academic economists have, for a long time, shied away from measuring certain vari-
ables. With some exceptions, they have relied almost exclusively on data on choices,
prices, and resources, or, more generally, objectively observable variables. They re-
frained from using measures of attitudes, intentions, stated preferences, beliefs, subjec-
tive expectations, social norms, etc. There are good reasons to treat such measures with
caution and even skepticism as they might be difficult to collect and can be affected
by different types of bias. However, empirical work that relies exclusively on choice
data, and supposedly objective measures, imposes strong restrictions to the economic
theories and models that can be brought to data, which typically take the form of strong

45
assumptions on the structure of the models one works with. Many important models,
some of which we discussed, could be analyzed with much more substance and empir-
ical bite were they to use a wider set of measures. Such measures, while not widely
used by economists, have been extensively used in other disciplines, from marketing to
psychology and child development.
Obviously, given the difficulties in collecting the innovative measures we are advo-
cating, they should be validated properly. Moreover, as we have argued, these measures
are particularly useful when utilized in combination with choice data and other standard
measures. This combination could be used both to validate the new measures, possibly
via simple correlations and to estimate and test richer models.
In the second part of the paper, we have put in practice some of these ideas, us-
ing a set of new measures collected in rural Tanzania to estimate a model of parental
investment, which we relate to measures of parental preferences, elicited from stated
preferences, bargaining power within couples, and parental beliefs about the process of
child development. While we do not estimate a fully structural model, we show how
these data can be used in combination with standard measures to quantify the impor-
tance of different factors affecting parental behavior. In a next step, these data can be
used to identify important causal links. Eliciting from respondents’ information about
choice under different counterfactual and hypothetical scenarios gathering information
about preferences, one can, in some contexts, solve through measurement some of the
endogeneity and identification issues that make empirical work challenging.
Every day, new data are being created and used, most noticeably administrative
data from a variety of sources and contexts. This is obviously a positive development.
However, we believe that well-designed and innovative survey measures of variables
and constructs that are theoretically relevant can be just as useful. Indeed, many new
measures are being developed in this direction, including those we have cited. Future
research should devote substantial efforts to develop, design, implement, and validate
new measurement tools that can provide useful evidence for economic theory and, ulti-
mately, public policy.

46
References
Acemoglu, D. and D. H. Autor (2011). “Skills, tasks and technologies: Implications for
employment and earnings.” In “Handbook of Labor Economics Volume 4,” (edited by
Ashenfelter, O. and D. E. Card). Amsterdam: Elsevier.

Acemoglu, D. and P. Restrepo (2019). “Automation and new tasks: How technology
displaces and reinstates labor.” Journal of Economic Perspectives 33(2), 3–30.

Ackerberg, D. A., K. Caves, and G. Frazer (2015). “Identification properties of recent

production function estimators.” Econometrica 83(6), 2411–2451.

Adams, A. and A. Andrew (2019). “Preferences and beliefs in the marriage market for
young brides.” Tech. rep., IFS Working Papers.

Aghion, P., A. Bergeaud, T. Boppart, P. J. Klenow, and H. Li (2019). “Missing growth

from creative destruction.” American Economic Review 109(8), 2795–2822.

Agostinelli, F. and M. Wiswall (2017). “Identification of dynamic latent factor models:

The implications of re-normalization in a model of child development.” Tech. rep.,
NBER, WP No 22441.

Alesina, A. and G.-M. Angeletos (2005). “Fairness and redistribution.” American eco-
nomic review 95(4), 960–980.

Alesina, A. and E. La Ferrara (2005). “Preferences for redistribution in the land of

opportunities.” Journal of Public Economics 89(5), 897–931.

Alesina, A., A. Miano, and S. Stantcheva (2018a). “Immigration and redistribution.”

Tech. rep., NBER, WP 24733.

Alesina, A., S. Stantcheva, and E. Teso (2018b). “Intergenerational mobility and prefer-
ences for redistribution.” American Economic Review 108(2), 521–54.

Almås, I. (2012). “International income inequality: Measuring ppp bias by estimating

engel curves for food.” American Economic Review 102(2), 1093–1117.

Almås, I., A. Armand, O. Attanasio, and P. Carneiro (2018). “Measuring and chang-
ing control: Women’s empowerment and targeted transfers.” The Economic Journal
128(612), F609–F639.

Almås, I., O. Attanasio, P. Jervis, and C. Ringdal (2020a). “Targeted cash transfers and
children’s welfare: Should women be targeted?” Tech. rep., NHH working paper.

Almås, I., A. W. Cappelen, and B. Tungodden (2020b). “Cutthroat capitalism versus

cuddly socialism: Are americans more meritocratic and efficiency-seeking than scan-
dinavians?” Journal of Political Economy 128(5), 1753–1788.

47
Almås, I. and Å. A. Johnsen (2018). “The cost of a growth miracle–reassessing price
and poverty trends in china.” Review of Economic Dynamics 30, 239–264.

Almås, I. (2012). “International income inequality: Measuring ppp bias by estimating

engel curves for food.” American Economic Review 102(2), 1093–1117.

Ameriks, J., J. Briggs, A. Caplin, M. D. Shapiro, and C. Tonetti (2020). “Long-term-care

utility and late-in-life saving.” Journal of Political Economy 128(6), 2375–2451.

Arrow, K. J. (1959). “Rational choice functions and orderings.” Economica 26(102),

121–127.

Attanasio, O. and B. Augsburg (2016). “Subjective expectations and income processes

in rural india.” Economica 83(331), 416–442.

Attanasio, O., B. Augsburg, and R. De Haas (2018). “Microcredit Contracts, Risk Diver-
sification and Loan Take-Up.” Journal of the European Economic Association 17(6),
1797–1842.

Attanasio, O., A. Barr, J. C. Cardenas, G. Genicot, and C. Meghir (2012). “Risk pool-
ing, risk preferences, and social networks.” American Economic Journal: Applied
Economics 4(2), 134–67.

Attanasio, O., R. Bernal, M. Giannola, and M. Nores (2020a). “Child development in

the early years.” Tech. rep., NBER, WP No 27812.

Attanasio, O., T. Boneva, and C. Rauh (2019a). “Parental beliefs about returns to differ-
ent types of investments in school children.” NBER Working Papers 25513, National
Bureau of Economic Research, Inc.

Attanasio, O., S. Cattan, E. Fitzsimons, C. Meghir, and M. Rubio-Codina (2020b). “Es-

timating the production function for human capital.” American Economic Review
110(1), 48–85.

Attanasio, O., F. Cunha, and P. Jervis (2019b). “Subjective parental beliefs. their mea-
surement and role.” Tech. rep., National Bureau of Economic Research.

Attanasio, O. and S. Krutikova (2020). “JEEA-FBBVA LECTURE 2019: Consump-

tion Insurance in Networks with Asymmetric Information: Evidence from Tanzania.”
Journal of the European Economic Association 18(4), 1589–1618.

Attanasio, O. and V. Lechene (2014). “Efficient responses to targeted cash transfers.”

Journal of Political Economy 122(1), 178 – 222.

Attanasio, O., C. Meghir, and E. Nix (2020c). “Human capital development and parental
investment in india.” The Review of Economic Studies 87(6), 2511–2541.

48
Attanasio, O. and E. Pastorino (2020). “Nonlinear pricing in village economies.” Econo-
metrica 88(1), 207–263.

Attanasio, O. P., M. Bornstein, and P. Jervis (2022). “Constructing effective measures of

child development: ages 6 to 36 months.” Tech. rep., In progress.

Autor, D. H. and D. Dorn (2013). “The growth of low-skill service jobs and the polar-
ization of the us labor market.” American Economic Review 103(5), 1553–97.

Banks, J., R. Blundell, and A. Lewbel (1997). “Quadratic engel curves and consumer
demand.” Review of Economics and statistics 79(4), 527–539.

Ben-Akiva, M. E., D. McFadden, K. Train et al. (2019). Foundations of stated preference

elicitation: Consumer behavior and choice-based conjoint analysis. Now.

Benhabib, J. and A. Bisin (2011). “Choice and process: Theory ahead of measurement.”
In “The Foundations of Positive and Normative Economics: A Hand Book,” (edited by
Caplin, A. and A. Schott). Oxford University Press.

Bernheim, B. D., D. Bjorkegren, J. Naecker, and M. Pollmann (2021). “Causal inference

from hypothetical evaluations.” Tech. rep., NBER, WP No 29616.

Berry, S., J. Levinsohn, and A. Pakes (2004). “Differentiated products demand sys-
tems from a combination of micro and macro data: The new car market.” Journal of
Political Economy 112(1), 68–105.

Bils, M. (2009). “Do Higher Prices for New Goods Reflect Quality Growth or Infla-
tion?*.” The Quarterly Journal of Economics 124(2), 637–675.

Bils, M. and P. J. Klenow (2001). “Quantifying quality growth.” American Economic

Review 91(4), 1006–1030.

Biroli, P., T. Boneva, A. Raja, and C. Rauh (2022). “Parental beliefs about returns to
child health investments.” Journal of Econometrics 231(1), 33–57. Annals Issue:
Subjective Expectations & Probabilities in Economics.

Bisin, A. and T. Verdier (2000). “Beyond the melting pot: cultural transmission, mar-
riage, and the evolution of ethnic and religious traits.” Quarterly Journal of Economics
115(3), 955–988.

Black, M. M., K. R. Bromley, V. Cavallera, J. Cuartas, T. Dua, I. Eekhout, G. Fink, M. J.

Gladstone, K. L. Hepworth, M. Janus, P. K. Kariger, G. A. Lancaster, D. C. McCoy,
G. P. J. McCray, A. Raikes, M. Rubio-Codina, S. van Buuren, M. Waldman, S. P.
Walker, and A. M. Weber (2019). “The global scale for early development -gsed.”
BMC Public Health .

49
Blass, A. A., S. Lach, and C. F. Manski (2010). “Using elicited choice probabilities to
estimate random utility models: Preferences for electricity reliability.” International
Economic Review 51(2), 421–440.

Block, H. D. and J. Marschak (1960). “Random orderings and stochastic theories

of response.” In “Contributions to Probability and Statistics,” (edited by Olkin, I.,
S. Ghurye, W. Hoeffding, W. Madow, and H. Mann). Stanford University Press.

Bloom, N., E. Brynjolfsson, L. Foster, R. Jarmin, M. Patnaik, I. Saporta-Eksten, and

J. Van Reenen (2019). “What drives differences in management practices?” American
Economic Review 109(5), 1648–83.

Bloom, N. and J. Van Reenen (2007). “Measuring and Explaining Management Practices
Across Firms and Countries*.” Quarterly Journal of Economics 122(4), 1351–1408.

Bobba, M. and V. Frisancho (2020). “Self-perceptions about academic achievement:

Evidence from mexico city.” Journal of Econometrics .

Boneva, T. and C. Rauh (2018). “Parental beliefs about returns to educational invest-
ments—the later the better?” Journal of the European Economic Association 16(6),
1669–1711.

Bornstein, M. H. (1996). “Ideas about parenting in argentina, france, and the united
states.” International Journal of Behavioral Development 19(2), 347–368.

Bornstein, M. H., W. A. Rothenberg, J. E. Lansford, R. H. Bradley, K. Deater-Deckard,

A. Bizzego, and G. Esposito (2021). “Child Development in Low- and Middle-Income
Countries.” Pediatrics 148(5). E2021053180.

Bourguignon, F., M. Browning, and P. A. Chiappori (2009). “Efficient intra-household

allocations and distribution factors: Implications and identification.” The Review of
Economic Studies 76(2), 503–528.

Browning, M. and P. A. Chiappori (1998). “Efficient intra-household allocations: A

general characterization and empirical tests.” Econometrica 66(6), 1241–1278.

Browning, M., P.-A. Chiappori, and A. Lewbel (2013). “Estimating consumption

economies of scale, adult equivalence scales and household bargaining power.” Re-
view of Economic Studies 80, 1267–1303.

Buser, T., M. Niederle, and H. Oosterbeek (2014). “Gender, competitiveness, and career
choices.” The quarterly journal of economics 129(3), 1409–1447.

Camerer, C. (2011). “The case for mindful economics.” In “The Foundations of Positive
and Normative Economics,” (edited by Caplin, A. and A. Schott). Oxford Univ. Press.

50
Caplin, A. (2021). “Economic data engineering.” Working Paper 29378, National Bu-
reau of Economic Research.

Carson, R. T. (2012). “Contingent valuation: A practical alternative when prices aren’t

available.” Journal of Economic Perspectives 26(4), 27–42.

Cavatorta, E. and B. Groom (2020). “Does deterrence change preferences? evidence

from a natural experiment.” European Economic Review 103456.

Chamberlain, G. and Z. Griliches (1975). “Unobservables with a variance-components

structure: Ability, schooling, and the economic success of brothers.” International
Economic Review 422–449.

Chen, X., L. P. Hansen, and P. G. Hansen (2020). “Robust identification of investor

beliefs.” Tech. rep., National Bureau of Economic Research.

Cherchye, L., P.-A. Chiappori, B. De Rock, C. Ringdal, and F. Vermeulen (2021). “Feed
the children.” Tech. rep., CEPR Discussion Paper No. DP16482.

Cherchye, L., B. D. Rock, and F. Vermeulen (2011). “The revealed preference approach
to collective consumption behaviour: Testing and sharing rule recovery.” The Review
of Economic Studies 78(1), 176–198.

Chiappori, P. A. (1988). “Rational household labor supply.” Econometrica 56(1), 63–90.

Christensen, L. R., D. W. Jorgenson, and L. J. Lau (1975). “Transcendental logarithmic

utility functions.” The American Economic Review 65(3), 367–383.

Costa, D. (2000). “American living standards, 1888-1994: evidence from consumer

expenditures.”

Crawford, I. and J. P. Neary (2021). “New Characteristics and Hedonic Price Index
Numbers.” The Review of Economics and Statistics 1–49.

Cunha, F., I. Elo, and J. Culhane (2013). “Eliciting maternal beliefs about the technology
of skill formation.” NBER Working Paper 19144.

Cunha, F. and J. J. Heckman (2008). “Formulating, identifying and estimating the tech-
nology of cognitive and noncognitive skill formation.” Journal of human resources
43(4), 738–782.

Cunha, F., J. J. Heckman, and S. M. Schennach (2010). “Estimating the technology of

cognitive and noncognitive skill formation.” Econometrica 78(3), 883–931.

Curtin, R. (2016). “Katona: a founder of behavioral economics.” In “Handbook of

Behavioral Economics,” 30–47. Routledge.

51
Dardanoni, V., P. Manzini, M. Mariotti, H. Petri, and C. Tyson (2022). “Mixture choice
data: revealing preferences and cognition.” Journal of Political Economy forth.

Deaton, A. and J. Muellbauer (1980). “An almost ideal demand system.” The American
economic review 70(3), 312–326.

Delavande, A. and B. Zafar (2019). “University choice: The role of expected earn-
ings, nonpecuniary outcomes, and financial constraints.” Journal of Political Economy
127(5), 2343–2393.

Deming, D. J. (2017). “The Growing Importance of Social Skills in the Labor Market*.”
The Quarterly Journal of Economics 132(4), 1593–1640.

Dercon, S. and P. Krishnan (2000). “In sickness and in health: Risk sharing within
households in rural ethiopia.” Journal of Political Economy 108(4), 688–727.

Dizon-Ross, R. (2019). “Parents’ beliefs about their children’s academic ability: Impli-
cations for educational investments.” American Economic Review 109(8), 2728–65.

Dominitz, J. and C. F. Manski (1996). “Eliciting student expectations of the returns to

schooling.” The Journal of Human Resources 31(1), 1–26.

Dominitz, J. and C. F. Manski (1997). “Using expectations data to study subjective

income expectations.” Journal of the American Statistical Association 92(439), 855–
867.

Doraszelski, U. and J. Jaumandreu (2013). “R&d and productivity: Estimating endoge-

nous productivity.” Review of Economic Studies 80(4), 1338–1383.

Doraszelski, U. and J. Jaumandreu (2018). “Measuring the bias of technological

change.” Journal of Political Economy 126(3), 1027–1084.

Dubois, P., R. Griffith, and M. O’Connell (2020). “How well targeted are soda taxes?”
American Economic Review 110(11), 3661–3704.

Dunbar, G. R., A. Lewbel, and K. Pendakur (2013). “Children’s resources in collective

households.” American Economic Review 103(1), 438–71.

Duncan, O. D. (1966). “Path analysis: Sociological examples.” American journal of

Sociology 72(1), 1–16.

Einav, L., E. S. Leibtag, and A. Nevo (2008). “On the Accuracy of Nielsen Homescan
Data.” Economic Research Report 56490, US Dept. of Agriculture.

Erdem, T., M. Keane, and T. Öncü (2005). “Learning about computers: An analysis of
information search and technology choice.” Quantitative Market and Econonomics
3(3), 207–247.

52
Ertaç, S., A. Hortaçsu, and J. W. Roberts (2011). “Entry into auctions: An experimental
analysis.” International Journal of Industrial Organization 29(2), 168–178.

Field, E., R. Pande, N. Rigol, S. Schaner, and C. Troyer Moore (2021). “On her own ac-
count: How strengthening women’s financial control impacts labor supply and gender
norms.” American Economic Review 111(7), 2342–75.

Gandhi, A., S. Navarro, and D. A. Rivers (2020). “On the identification of gross output
production functions.” Journal of Political Economy 128(8), 000–000.

Gilbert, M., C. Clark, J. Stone, F. Perroux, D. Lieu, Evelpides, F. Divisia, Tinbergen,

Kuznets, Smithies et al. (1949). “The measurement of national wealth: discussion.”
Econometrica: Journal of the Econometric Society 255–272.

Giustinelli, P., C. F. Manski, and F. Molinari (2019). “Precise or imprecise probabilities?

evidence from survey response on late-onset dementia.” Tech. rep., NBER.

Goldberger, A. S. (1971). “Econometrics and psychometrics: A survey of communali-

ties.” Psychometrika 36(2), 83–107.

Goldberger, A. S. (1972). “Structural equation methods in the social sciences.” Econo-

metrica 40(6), 979–1001.

Griffith, R. and M. O’Connell (2009). “The use of scanner data for research into nutri-
tion*.” Fiscal Studies 30(3-4), 339–365.

Griliches, Z. (1974). “Errors in variables and other unobservables.” Econometrica:

42(6), 971–998.

GSED (2021). “The global scale for early development (gsed).” Tech. rep., Early Child-
hood Matters.

Guiso, L., P. Sapienza, and L. Zingales (2004). “The role of social capital in financial
development.” American Economic Review 94(3), 526–556.

Guiso, L., P. Sapienza, and L. Zingales (2006). “Does culture affect economic out-
comes?” Journal of Economic Perspectives 20(2), 23–48.

Gul, F. and W. Pesendorfer (2011). “The case for mindless economics.” In “The Foun-
dations of Positive and Normative Economics,” (edited by Caplin, A. and A. Schott).
Oxford Univ. Press.

Haavelmo, T. (1958). “The role of the econometrician in the advancement of economic

theory.” Econometrica 26(3), 351–357.

Hamilton, B. W. (2001). “Using engel’s law to estimate cpi bias.” American Economic
Review 91(3), 619–630.

53
Harris, K. M. and M. P. Keane (1998). “A model of health plan choice:: Inferring
preferences and perceptions from a combination of revealed preference and attitudinal
data.” Journal of Econometrics 89(1-2), 131–157.

Hausman, D. M. (1994). The philosophy of economics: An anthology. Cambridge

University Press.

Hausman, J. (2012). “Contingent valuation: from dubious to hopeless.” Journal of

Economic Perspectives 26(4), 43–56.

Heckman, J., R. Pinto, and P. Savelyev (2013). “Understanding the mechanisms through
which an influential early childhood program boosted adult outcomes.” American
Economic Review 103(6), 2052–86.

Heckman, J. J., B. Liu, M. Lu, and J. Zhou (2020). “Treatment effects and the measure-
ment of skills in a prototypical home visiting program.” Tech. rep., NBER.

Henrich, J., S. J. Heine, and A. Norenzayan (2010). “The weirdest people in the world?”
Behavioral and Brain Sciences 33(2-3), 61–135.

Hu, Y. and S. M. Schennach (2008). “Instrumental variable treatment of nonclassical

measurement error models.” Econometrica 76(1), 195–216.

Jappelli, T. and L. Pistaferri (2000). “Using subjective income expectations to test for
excess sensitivity of consumption to predicted income growth.” European Economic
Review 44(2), 337–358.

Jayachandran, S., M. Biradavolou, and J. Cooper (2021). “Using machine learning and
qualitative interviews to design a five-question survey module for women’s agency.”
Tech. rep., Northwestern University Working Paper.

Jöreskog, K. G. and A. Goldberger (1975). “Estimation of a model with multiple in-

dicators and multiple causes of a single latent variable.” Journal of the American
Statistical Association 631–639.

Jullien, B. (2000). “Participation constraints in adverse selection models.” Journal of

Economic Theory 93(1), 1–47.

Juster, F., H. Cao, M. Perry, and M. Couper (2006). “The effect of unfolding brackets
on the quality of wealth data in hrs.” SSRN Electronic Journal .

Juster, F. T. (1964). Anticipations and Purchases: An Analysis of Consumer Behavior.

Princeton University Press.

Juster, F. T. (1966). “Consumer buying intentions and purchase probability: An experi-

ment in survey design.” J. of the American Statistical Association 61(315), 658–696.

54
Juster, F. T. and R. P. Shay (1964). Consumer Sensitivity to Finance Rates: An Empirical
and Analytical Investigation. NBER.

Kaiser, C. and A. J. Oswald (2022). “The scientific value of numerical measures of

human feelings.” PNAS 119(42), e2210412119.

Katona, G. (1959). “On the predictive value of consumer intentions and attitudes: A
comment.” The Review of Economics and Statistics 317–317.

Katona, G. (1974). “Understanding consumer attitudes.” Surveys of Consumers 1976,

203–219.

Katz, L. F. and K. M. Murphy (1992). “Changes in Relative Wages, 1963–1987: Supply

and Demand Factors*.” The Quarterly Journal of Economics 107(1), 35–78.

Kaufmann, K. and L. Pistaferri (2009). “Disentangling insurance and information in

intertemporal consumption choices.” American Economic Review 99(2), 387–92.

Kesternich, I., F. Heiss, D. McFadden, and J. Winter (2013). “Suit the action to the word,
the word to the action: Hypothetical choices and real decisions in medicare part d.”
Journal of Health Economics 32(6), 1313–1324.

Keynes, J. M. (1936). “The general theory of interest, employment and money.”

Kuziemko, I., M. I. Norton, E. Saez, and S. Stantcheva (2015). “How elastic are prefer-
ences for redistribution?” American Economic Review 105(4), 1478–1508.

Kuznets, S. (1941). “Statistics and economic history.” The Journal of Economic History
1(1), 26–41.

Kuznets, S. et al. (1937). “National income and capital formation, 1919-1935.” NBER
Books .

La Ferrara, E. (2019). “Presidential Address: Aspirations, Social Norms, and Develop-

ment.” Journal of the European Economic Association 17(6), 1687–1722.

Lancaster, T. and A. Chesher (1983). “An econometric analysis of reservation wages.”

Econometrica 51(6), 1661–76.

Lechene, V., K. Pendakur, and A. Wolf (2022). “Ols estimation of the intra-household
distribution of expenditure.” Journal of Political Economy 130, forthcoming.

Lee, J., R. Wedow, A. Okbay, E. Kong, O. Maghzian, M. Zacher, T. Nguyen-Viet,

P. Bowers, J. Sidorenko, R. Linnér, M. Fontana, T. Kundu, C. Lee, H. Li, R. Li,
R. Royer, P. Timshel, R. Walters, E. Willoughby, and D. Cesarini (2018). “Gene dis-
covery and polygenic prediction from a genome-wide association study of educational
attainment in 1.1 million individuals.” Nature Genetics 50.

55
Levinsohn, J. and A. Petrin (2003). “Estimating production functions using inputs to
control for unobservables.” The review of economic studies 70(2), 317–341.

List, J. A. and C. A. Gallet (2001). “What experimental protocol influence disparities

between actual and hypothetical stated values?” Environmental and Resource Eco-
nomics 20(3), 241–254.

List, J. A., J. Pernaudet, and D. Suskind (2021). “It all starts with beliefs: Addressing
the roots of educational inequities by shifting parental beliefs.” WP 29394, NBER.

Louviere, J., D. Hensher, and J. Swait (2000). Stated Choice Models: Analysis and
Application. Cambridge University Press.

Luce, R. and P. Suppes (1965). “Preference, utility, and subjective utility.” Handbook of
Mathematical Psychology, III, New York: Wiley 249–409.

Luce, R. D. (1956). “Semiorders and a theory of utility discrimination.” Econometrica,

Journal of the Econometric Society 178–191.

Luce, R. D. (1959). Choice Behavior. A Theoretical Analysis. New York: Wiley.

Luce, R. D. and J. W. Tukey (1964). “Simultaneous conjoint measurement: A new type

of fundamental measurement.” Journal of mathematical psychology 1(1), 1–27.

Manski, C. F. (1990). “The use of intentions data to predict behavior: A best-case

analysis.” Journal of the American Statistical Association 85(412), 934–940.

Manski, C. F. (2004). “Measuring expectations.” Econometrica 72(5), 1329–1376.

Maskin, E. and J. Riley (1984). “Monopoly with incomplete information.” RAND Jour-
nal of Economics 15(2), 171–196.

McCoy, D. C., J. Seiden, M. Waldman, and G. Fink (2021). “Measuring early childhood
development.” Annals of the New York Academy of Sciences 1492(1), 3–10.

Miller, G., Á. de Paula, and C. Valente (2020). “Subjective expectations and demand for
contraception.” Working Paper 27271, National Bureau of Economic Research.

Miller, S. A. (1988). “Parents’ beliefs about children’s cognitive development.” Child

Development 59(2), 259–285.

Mueller, A. and J. Spinnewijn (2021). “Expectations data, labor market and job search.”
Handbook Chapter (Draft) .

Murphy, P. G., James J.and Allen, T. H. Stevens, and D. Weatherhead (2005). “A meta-
analysis of hypothetical bias in stated preference valuation.” Environmental and Re-
source Economics 30(3), 313–325.

56
Neary, J. P. (2004). “Rationalizing the penn world table: True multilateral indices for
international comparisons of real income.” American Economic Review 94(5), 1411–
1428.

Nordhaus, W. D. (1998). “Quality change in price indexes.” Journal of Economic Per-

spectives 12(1), 59–68.

Olley, G. S. and A. Pakes (1992). “The dynamics of productivity in the telecommunica-

tions equipment industry.” Tech. rep., National Bureau of Economic Research.

Paiella, M. and L. Pistaferri (2017). “Decomposing the Wealth Effect on Consumption.”

The Review of Economics and Statistics 99(4), 710–721.

Parnes, H. S. (1975). “The national longitudinal surveys: New vistas for labor market
research.” The American Economic Review 65(2), 244–249.

Pistaferri, L. (2001). “Superior Information, Income Shocks, and the Permanent Income
Hypothesis.” The Review of Economics and Statistics 83(3), 465–476.

Pistaferri, L. (2003). “Anticipated and unanticipated wage changes, wage risk, and in-
tertemporal labor supply.” Journal of Labor Economics 21(3), 729–754.

Plott, C. R. and V. L. Smith (2008). Handbook of experimental economics results, vol. 1.

Elsevier.

Potter, S., M. Del Negro, G. Topa, and W. Van der Klaauw (2017). “The advantages of
probabilistic survey questions.” Review of Economic Analysis 9(1), 1–32.

Salz, T. and E. Vespa (2020). “Estimating dynamic games of oligopolistic competition.”

The RAND Journal of Economics 51(2), 447–469.

Samuelson, P. A. (1938). “A note on the pure theory of consumer’s behaviour.” Eco-

nomica 5(17), 61–71.

Samuelson, P. A. (1948). “Consumption theory in terms of revealed preference.” Eco-

nomica 15(60), 243–253.

Schennach, S. M. (2004). “Estimation of nonlinear models with measurement error.”

Econometrica 72(1), 33–75.

Scur, D., R. Sadun, J. Van Reenen, R. Lemos, and N. Bloom (2021). “The world man-
agement survey at 18.” Oxford Review of Economic Policy 37(2), 231–258.

Stantcheva, S. (2022). “How to run surveys: A guide to creating your own identifying
variation and revealing the invisible.” Working Paper 30527, NBER.

57
Stigler, G. J. and G. S. Becker (1977). “De gustibus non est disputandum.” The American
Economic Review 67(2), 76–90.

Stone, R. (1954). “Linear expenditure systems and demand analysis: an application to

the pattern of british demand.” The Economic Journal 64(255), 511–527.

Stone, R. (1984). “Richard stone-prize lecture: The accounts of society.” Nobelprize.

org. Nobel Media AB .

Tobin, J. (1959). “On the predictive value of consumer intentions and attitudes.” The
review of economics and statistics 1–11.

Todd, P. E. and K. I. Wolpin (2003). “On the specification and estimation of the produc-
tion function for cognitive achievement.” The Economic Journal 113(485), F3–F33.

Van der Klaauw, W. and K. I. Wolpin (2008). “Social security and the retirement and
savings behavior of low-income households.” J. of Econometrics 145(1-2), 21–42.

Visco, I. (1984). Price Expectations in Rising Inflation. Contributions to economic

analysis. North-Holland.

Wiswall, M. and B. Zafar (2015). “Determinants of college major choice: Identification

using an information experiment.” The Review of Economic Studies 82(2), 791–824.

Wolpin, K. I. and F. Gonul (1985). “On the use of expectations data in micro surveys:
The case of retirement.” Tech. rep., Ohio State University.

Wright, S. (1934). “The method of path coefficients.” The annals of mathematical

statistics 5(3), 161–215.

58
Appendices

A1 Specification of Measurement Systems

This Appendix provides more details on the general issues in the specification and esti-
mation of measurement systems that were discussed in Section 5.1 of the main text.
One assumption that can be relaxed in the context of the measurement system, and
that we mention in the main text, is that of linearity, especially when the available
measures are binary or discrete. Here, we explain the logic. In such a situation, a
common approach model is the Item Response Theory (IRT). In the case of binary
measures, an IRT model relates a factor to a latent measure, which in turn determines
the observed outcome. Omitting the superscript for the latent factor θ , we have:


k
1 i f αtk + βtk θit + εitk > 0
mit = (A1)
0 otherwise

The specific IRT model is then determined by assumptions about the distribution of
the measurement error term εitk ; assuming normality, one gets a Probit type of model,
while assuming a logistic distribution one obtains a logistic relation, or what is usually
referred to as a Rasch model. What we are considering is often defined as a 2-parameter
Rasch model. Restricting βtk to 1, one obtains a 1-parameter Rasch model. In the liter-
ature, a 3-parameter Rasch model considers the possibility of random ‘correct’ answers
through an additional parameter.
Another more recent example of IRT use, whose use has become more common in
economics, is that of polygenic scores, which aggregate data from many different sites
of the human genome and are based on correlations from a wide population data set.
Lee et al. (2018), for instance, present estimates of a polygenic score which is associated
with individual educational attainment in a specific population. Interestingly, several
authors recently noticed that several estimates of the same polygenic score might be
available, where the weights to aggregate information from individual loci are based on
different samples and/or slightly different methodologies. These alternative estimates
can then potentially be used to deal with measurement error problems, as in the model

A1
discussed above.

A2 The Tanzania data, new measures and results

The Bukoba region is located in the northwest corner of Tanzania on the western shore
of Lake Victoria. The population is around 3 million people which is about 5 percent
of the national population. It has an intercensal growth rate 2002/2012 of 3.2 (2.7 aver-
age for Tanzania) and the double percentage of children age 6-59 months classified as
having malaria, according to the Rapid Diagnostic Test in 2017 in terms of country per-
centage (15.4% vs 7.3%). The Kagera region is predominantly rural, and agricultural-
dependent, producing mainly bananas and coffee in the North and rain-fed annual crops
in the South.
The study was conducted in several stages (see Figure A1). The study sample was
randomly chosen to be representative of the total population. The villages were selected
from a random sample in Kagera region. In the selected villages, all households were
listed, and from those satisfying the selection criteria, a random sample was drawn.
During the pilot period a number of well-known and tested measures, were adapted
to the local context using multiple techniques (focus group discussions, ethnography
study, and pre-pilot in small samples). We used a multi-method approach for the
early childhood development (ECD) instruments, combining three main paths: par-
ent/observer report, direct assessment, and observation (through filming and coding).
After the pilot period, the first stage was implemented in June 2018. Approximately
450 households were recruited across 10 villages. The sample households were ran-
domly selected among those with a child aged between 6 and 36 months and a mother
aged between 15 and 25. In this sample, child development was measured using several
tools. In addition to standard demographic information about the households involved,
the survey also included measures of parental investment.
Information collected during a partial census was used to randomly sample pairs
of primary caregivers and children of eligible households from each of the following
groups: 150 primary caregivers aged 15-25 years with a child aged 6-11 months, 150
primary caregivers aged 15-25 years with a child aged 12-23 months, and 150 primary
caregivers aged 15-25 years with a child aged 24-36 months.

A2
 Figure A1: Study Design

Second stage
First stage Validate tools
Pilot stage Adapted tools +
+ Main respondent
Main respondent mother, father, and
mother of selected child couple of selected child
+
Experimental design

In a second stage, which was implemented in August 2018, an additional 450 house-
holds were recruited according from the same villages the following procedure. First
a random sample of 150 households was selected from a list of 5200 households with
children aged 6 to 36 months and mothers aged 15 to 25 years in 8 villages. This is
what we will refer to as the mothers sample. In addition, a second list of around 2000
households with children aged 6 to 36 months and the fathers present was identified
in 5 different villages.30 From this list, which also included mothers older than 25, an
additional sample of households with fathers present and mothers aged between 15 and
25 years was selected and formed what we label as the couples sample. From the re-
maining households in the second list, a third sample of 150 households was selected
that we label as the fathers sample. Because of the way the second list was formed, the
fathers sample comprises couples who are considerably older. Indeed, only 6.6% of the
mothers in the sample were younger than 25.31
The data collected in the second stage is what we mostly use in this appendix and
in Sections 5 and 6. The main goal of the second stage data collection was to validate
some of the child development measures constructed with the factor analysis of first
stage data. However, we also collected a number of new variables aimed at measuring
individual tastes, beliefs, and bargaining power within the couple. Each sample’s label
corresponds to the respondents in this new set of questions. In the mothers sample,
for instance, the questions about tastes, subjective beliefs about parental investment
returns, and bargaining power were answered by the mother in private; in the couples
30 The two sets of villages belonged to different wards, the Kashai and the Bakoba. Both wards
belonged to the Bukoba municipality.
31 In the mothers and couples samples, there are a few mothers older than 25.

A3
sample, the taste questions were answered jointly by the couple while the others by
the mother in private; in the fathers sample, fathers answered the bargaining power and
taste questions in private. All other (standard) modules were answered by the mother.
All data collection followed a rigorous process of tools, instruments, and survey
development.32 Of the 450 pairs recruited for the last stage of the survey, we obtained
usable information from 423 households, comprising 145 in the mothers sample, 136
in fathers sample, and 142 in the couples sample.
The fact that the questions that elicit information on tastes, subjective beliefs about
parental investment returns, and bargaining power within the couple are answered by
different respondents in the three subsamples makes the survey particularly interesting,
as it allows to test the hypothesis that individuals within the family are characterized
by differences in these variables. Unfortunately, the recruitment process followed in
the field (described above) makes comparisons across different sub-samples difficult
to interpret. This is particularly true for the fathers sample, which includes consider-
ably older individuals, as we show below. While the fact that the mothers and couples
samples were drawn from different villages is not particularly worrying, given the sim-
ilarities and the proximity of the villages, the systematic difference in the age structure
of the study sample makes comparisons across samples problematic.
In Table A1, we report descriptive statistics on the main features of the three samples
discussed above. The table also contains p − values for tests (adjusted for multiple
hypothesis testing) of the difference between the mothers and the fathers samples (in
column 5) and the mothers and the couples samples (in column 8). In addition to the
new measures (discussed below), information on a standard set of variables, including
demographics, education, and wealth markers, was also collected in the three samples.
Consistent with the sampling scheme, the mothers and couples samples have rel-
atively younger mothers, with an average age of 22, while in the fathers sample, the
average mother’s age is about 7 years higher. This difference is also reflected in the
average father’s age in the fathers sample, (between 5 and 7 years higher than in the
mothers and couples samples). About 20% of the mothers sample is made of single
32 All instruments were translated into the local language, Swahili. Translations were carried out by
the field staff following a rigorous back-translation procedure. The household surveys were administered
in the child’s home by enumerators. The adapted version of the assessments of child development were
conducted by trained staff in the presence of the primary caregiver.

A4
 Table A1: Household characteristics

Mothers sample Couples sample Fathers sample

Mean Observations Mean Observations Adj. p-value Mean Observations Adj. p-value
Demographics
Age, M 23.12 145 22.82 142 0.941 30.42 136 0.010
Age, F 28.98 105 28.83 139 1.000 37.05 134 0.010
Secondary school or more, M 0.35 144 0.37 142 1.000 0.23 136 0.079
Secondary school or more, F 0.50 105 0.46 139 0.990 0.34 134 0.059
Food Share 0.69 145 0.61 142 0.030 0.63 136 0.079
Wealth index -0.17 144 -0.04 142 0.931 0.22 136 0.010
Child characteristics
Child is male 0.57 145 0.54 142 0.990 0.45 136 0.079
Age in months 18.44 145 18.25 142 1.000 18.36 136 0.901
# siblings 0.57 145 0.62 142 0.990 2.16 136 0.010
Birthrank 2.00 145 2.04 142 1.000 3.53 136 0.010
Note: The table displays the descriptive statistics for the households.“M” indicates values for mothers and “F” indicates values for fathers. Column (5) compares
the mothers and couples samples while Column (8) compares the mothers and fathers samples. Columns (5) and (8) reports p-values adjusted for multiple
hypothesis testing using the Romano-Wolf approach. Source: Tz Pilot.

mothers (i.e., the child’s father is not present). The average child’s age is uniform
across the three samples and the share of male children is lowest in the fathers sample.
The only significant and somewhat surprising difference between the mothers and
couples samples is the share of total expenditure spent on food, which could be consid-
ered a useful indicator of economic well-being33 This variable is higher in the mothers
sample, at 0.69,indicating a poorer sample, and lowest in the couples sample, at 0.61.
We also observe a number of wealth indicators, which we use to estimate a wealth
index, normalized to have zero mean in the whole sample. Consistent with the evidence
on the food share, the wealth index is lowest in the mothers sample. These differences
in permanent income and wealth, however, are not consistent with the information on
education: both mothers and fathers in the fathers sample are the least educated, while
the mothers sample is the most educated. These differences might reflect cohort effects.

A3 Additional Tables
In this section, we present additional tables that are referred to in the main text of the
paper.
33 Theintuition for this goes back to the established economic regularity that the food share falls with
income Food shares have been used to identify differences in real income and well-being (Hamilton,
2001; Costa, 2000; Almås, 2012; Almås and Johnsen, 2018).

A5
A3.1 Parental beliefs
In this section we present the estimates of the measurement system for parental be-
liefs. In estimating the perceived effectiveness of parental investment we consider both
cognitive, language and socioemotional outcomes. A possible alternative which we
have explored is to separate the perceived productivity of parental investment on socio-
emotional development from that on cognition and language.
Table A2: Measurement system for Beliefs
A Measurement System for Beliefs
Couples sample Mothers sample Fathers sample
β α β α β α
Language hard 1.000 0.000 1.000 0.000 1.000 0.000
High Dev. - - - - - -
Language hard 2.098 -0.023 1.220 0.024 1.314 -0.042
Low Dev. (0.514) (0.201) (0.257) (0.124) (0.243) (0.114)
Language medium 0.868 0.040 0.573 0.095 1.015 -0.090
High Dev. (0.175) (0.074) (0.136) (0.066) (0.138) (0.065)
Language medium 1.628 0.009 0.556 0.255 1.039 -0.066
Low Dev. (0.430) (0.166) (0.172) (0.082) (0.173) (0.081)
Language easy 0.746 0.065 0.057 0.296 0.877 -0.008
High Dev. (0.180) (0.075) (0.091) (0.043) (0.125) (0.058)
Language easy 1.082 0.167 -0.181 0.556 1.062 -0.075
Low Dev. (0.307) (0.122) (0.137) (0.065) (0.179) (0.083)
Socio-emotional Nine 0.332 0.156 0.566 0.216 0.965 -0.081
High Dev. (0.161) (0.067) (0.128) (0.060) (0.169) (0.076)
Socio-emotional Nine 0.397 0.259 1.420 0.072 1.168 -0.206
Low Dev. (0.317) (0.133) (0.238) (0.106) (0.199) (0.090)
Socio-emotional Five 0.362 0.128 0.379 0.190 0.819 -0.068
High Dev. (0.131) (0.055) (0.088) (0.041) (0.135) (0.060)
Socio-emotional Five 0.024 0.352 0.964 0.118 0.928 -0.182
Low Dev. (0.240) (0.100) (0.172) (0.078) (0.170) (0.077)
Socio-emotional Three 0.425 0.075 0.118 0.191 0.589 0.003
High Dev. (0.145) (0.061) (0.065) (0.031) (0.118) (0.053)
Socio-emotional Three -0.180 0.374 0.368 0.168 0.702 -0.098
Low Dev. (0.202) (0.084) (0.117) (0.055) (0.170) (0.078)
Factor mean 0.346 0.303 0.348
Factor variance 0.069 0.151 0.109
Note: This table shows the loading factors (β ) and the intercept (α) for each returns to parental investment
on Language and Socio-emotional skills elicited in our survey which are estimated through a measurement
system model. Columns (1) and (2) are for the couples sample, Columns (3) and (4) for the mothers sample,
and Columns (5) and (6) for the fathers sample. Standard errors in parentheses. Source: Tz Pilot.

A3.2 Relative preferences for children human capital

As explained in the main text, we compute the ratio of (stated) allocation of 4 different
commodities allocated to the mother and father to the total expenditure on allocated to
children and use these 8 variables as markers of the ‘taste for children human capital’.
In the Table we report estimates of the measurement system so obtained.

A6
 Table A3: Measurement system for relative taste for child human capital
Relative taste for child human capital
Couples sample Mothers sample Fathers sample
β α β α β α
Mother’s clothing ratio 1.000 0.000 1.000 0.000 1.000 0.000
- - - - - -
Father’s clothing ratio 1.009 -0.021 0.936 -0.039 0.729 0.007
(0.058) (0.013) (0.190) (0.033) (0.165) (0.023)
Mother’s food ratio 0.850 0.043 4.294 -0.527 1.472 -0.031
(0.172) (0.038) (0.649) (0.111) (0.295) (0.041)
Father’s food ratio 0.850 0.001 -0.011 0.111 1.229 -0.023
(0.150) (0.032) (0.179) (0.031) (0.252) (0.036)
Mother’s health ratio 0.540 0.043 -0.028 0.115 1.410 -0.057
(0.065) (0.014) (0.106) (0.018) (0.294) (0.041)
Father’s health ratio 0.530 0.030 -0.005 0.078 0.795 -0.006
(0.060) (0.012) (0.116) (0.020) (0.189) (0.027)
Mother’s transportation ratio 0.540 0.043 -0.028 0.115 1.410 -0.057
(0.065) (0.014) (0.106) (0.018) (0.294) (0.041)
Father’s transportation ratio 0.530 0.030 -0.005 0.078 0.795 -0.006
(0.060) (0.012) (0.116) (0.020) (0.189) (0.027)
Factor mean (variance) 0.175 (0.020) 0.162 (0.004) 0.135 (0.002)
Note: This table shows the loading factors (β ) and the intercept (α) for each ratio of the resources allocated to
each spouse for four adult commodities (food, clothing, health, and transportation) from the allocation ques-
tions which are estimated through a measurement system model. Standard errors in parentheses. Columns
(1) and (2) are for the couples sample, Columns (3) and (4) for the mothers sample, and Columns (5) and (6)
for the fathers sample. Source: Tz Pilot.

A3.3 Parental investment

In terms of the measure of parental investment, we use multiple sources of information:

Expenditures for children: Reported spending on children’s health, clothing, shoes,

toys, education, books and bedding.

Material investments: Reported frequency of acquiring the following products for

their child: food, clothing, footwear, confectioneries medicine and oral re-hydration
salts, mosquito net, books, toys, and learning materials (notebooks, pens, and pencils).

Play material: The number of toys the child has, made at home or bought, music
instruments, books, and drawing equipment.

Adult activities with children: Reading books, singing, playing, and cooking with
the child.

Didactic scale: From the Parental Style Questionnaire (PSQ): whether the primary
caregiver (i) spends time playing with the child, (ii) provides the child with indepen-

A7
dent time to explore and learn on his/her own, (iii) provides the child with diverse
social and interactive experience with same-age peers through play groups or informal
get-together, (iv) provides the child with a structured organized, and predictable en-
vironment, (v) provides language learning opportunities for the child by labeling and
describing qualities of objects, events or activities, reading books etc., (vi) provides the
child with a variety of toys and objects for play and exploration, (vii) is patient with the
child’s misbehavior, and (viii) is flexible about behaviors expected from the child.

Social scale: The social scale from the PSQ accounts for whether the primary care-
giver (i) promptly and appropriately respond to the child’s expressed distress or discom-
fort, (ii) spend time talking to or conversing with child, (iii) provide child with quick
and positive feedback to his/her bid for attention, (iv) provide child with affectionate
displays of warmth and attention, and (iv) is aware of what child wants or feels.
Table A4 reports descriptive statistics on the components of parental investment.

Table A4: Parental Investment Components

Couples sample
Mean Observations

Parental Investment
Raw Activity Score (/26) 14.89 142
Raw Material Investment Score (/12) 5.23 142
Raw Play Material Score (/8) 1.51 142
Social scale (/5) 4.28 142
Didactic scale (/8) 4.94 142
Expenditure on children share 0.32 142
Total Parental Investment (factor) -0.48 142
Note: Mean and number of observations for markers of parental
investment for the couples sample. In parenthesis the maximum
value of each marker of parental investment. Source: Tz Pilot.

A8