M A X P L A N C K S O C I E T Y
Preprints of the
Max Planck Institute for
Research on Collective Goods
Bonn 2010/07
Dictator Games:
A Meta Study
Christoph Engel
Preprints of the
Max Planck Institute
for Research on Collective Goods Bonn 2010/07
Dictator Games: A Meta Study
Christoph Engel
March 2010
revised January 2011
Max Planck Institute for Research on Collective Goods, Kurt-Schumacher-Str. 10, D-53113 Bonn
http://www.coll.mpg.de
1
Dictator Games: A Meta Study
Christoph Engel
Abstract
Over the last 25 years, more than a hundred dictator game experiments have been published.
This meta study summarizes the evidence. Exploiting the fact that most experiments had to fix
parameters they did not intend to test, in multiple regression the meta study is able to assess the
effect of single manipulations, controlling for a host of alternative explanatory factors. The re-
sulting rich dataset also provides a testbed for comparing alternative specifications of the statisti-
cal model for analysing dictator game data. It shows how Tobit models (assuming that dictators
would even want to take money) and hurdle models (assuming that the decision to give a posi-
tive amount is separate from the choice of amount, conditional on giving) provide additional in-
sights.
JEL: C24, C91, D03
2
1. Introduction
Some 25 years ago, Daniel Kahneman first had students in his classroom play an ultimatum
game (Güth, Schmittberger et al. 1982), which was then followed by the question:
“You will be matched at random with two other students, and you will get to share
some money with one or both of them. If the two people made different decisions
in the first stage (e.g. one of them took $10 and one took $18), then you must
make a decision about how to allocate the money. Call the person who took $10
and gave the other one $10 student E (for even). Call the person who took $18 and
gave the other one $2 student U (for uneven). Your choices are as follows: you
may allocate $5 to yourself, $5 to student E, and nothing to student U; or you may
allocate $6 to yourself, nothing to student E, and $6 to student U”.
74 % of their participants chose the first option although this cost them $1 (Kahneman,
Knetsch et al. 1986:S290 f.). This is how a literature started to which in the meantime 129
contributions have been published, testing a total of 616 different treatments.
1
The experi-
mental paradigm has proven so powerful precisely because it is so simple. Actually the suc-
cessors of Kahneman have further simplified the game (starting with Forsythe, Horowitz et al.
1994). The game now is typically stripped of the third party punishment component and
played in isolation. The dictator's action space is complete, so that she may distribute the pie
at her will between the recipient and herself.
Kahneman invented the game as part of his programme that turned textbook assumptions into
behavioural hypotheses. While normally a sizeable fraction of participants does indeed give
nothing, as predicted by the payoff maximisation hypothesis, only very rarely this has been
the majority choice. It by now is undisputed that human populations are systematically more
benevolent than homo oeconomicus. Later experiments have explored this predisposition in
two dimensions: situational and demographic. The former implicitly sticks to the claim that, at
least at the population level, behavioural dispositions are human universals. It engages in re-
fining the conditions under which benevolence is to be expected. The latter puts the research
question upside down and uses the extremely simple design as a tool for quantifying system-
atic behavioural differences between populations.
The dictator game has become popular among experimentalists. In the one year of 2008, 30
new papers with this game have been published. It therefore is time to take stock, and to make
the existing body of evidence accessible. Yet this paper aims beyond merely providing orien-
tation. Implicitly, through their design choices, experimenters have generated data on inde-
pendent variables they have not explicitly set out to test. They for instance have played a one-
shot game with students, asking dictators to divide a pie of $10 given to them between them-
selves and an anonymous recipient from the same subject pool. What looks like a perfectly
standard dictator game implicitly provides data on one shot versus repeated games; on games
with students versus other populations; on manna from heaven versus earned money; on
stakes; on a specified degree of social distance; on dictator-recipient anonymity versus dicta-
1 For detail, see the list of papers used for the meta study in Appendix 1.
3
tor identification. This evidence is untapped as yet. It is useful for two purposes. The basis for
testing the effect of isolated manipulations becomes much broader. More interestingly even, it
becomes possible to control for alternative explanations to a degree that by far transcends
what is feasible in individual experimental studies. That way one learns which effects are ro-
bust, and how big effects are once one controls for other factors that have been shown to mat-
ter for the willingness of dictators to give.
The remainder of this paper is organised as follows. Section 2 explains how the sample has
been collected and addresses analytic methodology. Section 3 treats all experiments as contri-
butions to one question: how much are dictators willing to give? Section 4 focuses on individ-
ual experimental manipulations. Section 5 simultaneously uses all these independent variables
to explain generosity with multiple regressions. In this section I also discuss different options
for specifying the statistical model. Section 6 concludes.
2. Data and Methodology
In disciplines like medicine or psychology, meta-analysis is standard.
2
In econometrics it is
also becoming increasingly frequent.
3
By contrast, the number of meta studies on economic
experiments is still fairly limited (Druckman 1994; Harless and Camerer 1994; Croson and
Marks 2000; Zelmer 2003; Huck, Normann et al. 2004; Oosterbeek, Sloof et al. 2004; Black-
well 2007; Engel 2007; Prante, Thacher et al. 2007; Jones 2008; Hopfensitz 2009; Percoco
and Nijkamp 2009; Weizsäcker 2010). A word on the potential and the limitations of the
method may therefore be in order.
Meta-analysis is best understood as quantitative literature review (Stanley 2001). It is much
more objective, and usually also more informative, than narrative review. It exploits the fact
that there is a whole body of studies that are sufficiently similar to make their joint analysis
meaningful; with lab experiments, this condition is much easier to fulfil than with field data,
given they are standardised by design (Thompson and Pocock 1991). Nonetheless, different
experiments have related, but different research questions. They may be differently well exe-
cuted. Non-results are difficult to publish, which may lead to publication bias. Heterogeneity
between studies can be pronounced, which may be due to unobserved explanatory factors (all
of this had already been discussed by Druckman 1994). These concerns can be mitigated by
the techniques presented in this section. Still results from meta analysis should be read with
caution. If a researcher remains sceptical about a finding from meta-analysis, she should de-
sign a new experiment that is specifically targeted to precisely this research question.
Two papers have engaged in a similar exercise. In 2003, Camerer has done a meta-study of 11
experiments (Camerer 2003:57 f.). In 2008 Cardenas and Carpenter have done the same for
2 The database PsychInfo reports 4,685 entries with the word „meta analysis” in the title of the paper.
3 The database EconLit reports 491 entries with the keyword „meta analysis”, most of which refer to work
in econometrics.
