CAN RISK‐TAKING PREFERENCES BE MODIFIED? SOME EXPERIMENTAL EVIDENCE Alison L. Booth Preliminary draft. Can Risk‐taking Preferences be Modified? 1 Some Experimental Evidence Alison L. Booth Australian National University and University of Essex June 2013 1 This address was prepared for the workshop entitled, ‘The Determinants of Gender Gaps: Institutional Design and Historical Factors ‘, at CESifo's 14th Venice Summer Institute, Venice, July 2013. It draws extensively on joint work with Patrick Nolen and Lina Cardona Sosa. (This paper’s file name is CESifo_VeniceJune2013.docx). For downloadable papers, see: http://ideas.repec.org/e/pbo47.html 0 1. Introduction Discussions of the origin of gender gaps in economic outcomes sometimes raise the issue of whether productivity‐enhancing characteristics are gender‐specific or are instead developed by cultural values within the community. For example, have men evolved to be innately more risk‐ taking than women or have they become that way in part through cultural pressures? In the nature versus nurture debate about male and female preferences and outcomes, where do we stand as economists? In my experimental work with co‐authors, we have tried to take a small step forward in this regard, and I will talk about some of this work in my lecture today. Women are increasingly found in employment in the market sector of the economy and they are increasingly enrolling in degree fields that have been largely male. Yet there are still many areas where women are under‐represented. Examples are found in the highest levels of mathematics, the physical sciences, and engineering (NAS 2006). This is in spite the fact that, in many countries, more women now attend university than men.2 Furthermore, in the US there is a gender gap in standardized mathematics tests scores.3 This is especially pronounced at the top of the distribution (Ellison and Swanson, 2010). It has been suggested that there might be gender differences in risk aversion, feedback preferences or in liking for competition, and that these might explain explain gender differences in observed educational and labour market outcomes. For example, obtaining promotion and pay raises often involves competition, and it may be that women do not like to compete but men do. A relatively recent and rapidly growing literature attempts to investigate – using either survey or experimental data ‐ if women and men differ systematically in some psychological characteristics that might explain the fact that women are under‐represented in high‐paying jobs and high‐level occupations. If so, then the suggestion is that ‐ once these 2 In the US in the 1960s there were 1.55 males for every female undergraduate but by 2003 there were 1.30 females for every male undergraduate (Goldin et. al. 2006). A similar ratio is found in Australia (Booth and Kee, 2011). 3 There is a gender gap in standardized mathematics tests that varies across countries (Else‐Quest et. al. 2010, Guiso et. al 2008) and some argue it may not be large enough to be of any practical importance (Hyde et. al. 2008). However Fryer and Levitt (2010) document a substantial gender gap in mathematics in the US. After six‐years of education, students have a 0.2 of a standard deviation gender gap in test scores. This gap is roughly half as large as the black‐white test score gap. 1 differences are controlled for ‐ perhaps the gender pay gap will disappear from empirical estimates. But is this the case? 2. Psychological factors and survey‐based evidence What does survey‐based evidence have to say about the role of psychological factors in affecting gender wage gaps? Clearly the use of contemporaneous survey‐based measures of risk‐aversion, self‐esteem and competitive or collaborative behavioural traits is dogged by potential endogeneity. However, personality variables are incorporated into the survey‐based analysis of, inter alia, Goldsmith et al. (1997), Bowles et al. (2001), and Mueller and Plug (2004). Using NLSY data, Goldsmith et al. (1997) show that personality variables as well as human capital are correlated with wages, but they do not investigate the gender dimension. However, Mueller and Plug (2006) do, using a Wisconsin‐based survey. While their psychological variables are contemporaneous with measured earnings, they nonetheless find interesting correlations between measures of personality and earnings. Manning and Swaffield (2008) avoid the endogeneity issue by using predetermined psychological information ‐ mainly measured at age 16 ‐ from the British Cohort Study to estimate gender wage gaps at age 30. They find that, on labour market entry, there is no gender wage gap for otherwise identical, fully “work‐committed” women and men (those with no children, no intention of having children, and with continuous full‐time work experience). However, by age 30 there remains a substantial unexplained gap: women who have continuous full‐time employment, have had no children and express no desire to have them, earn about 8 log points less than equivalent men after 10 years in the labour market. Manning and Swaffield then investigate the role of psychological variables in explaining this, focusing on risk attitudes, competitiveness, self‐ esteem, ‘other‐regarding’, and career‐orientation.4 The psychological variables are found to explain an ‘upper‐bound’ of 4.5 log points of the gender wage gap. Some of these survey‐based measures of psychological factors are rather indirect. Moreover it is not easy to find measures of psychological factors that are genuinely 4 Risk attitudes are proxied by wearing a seat‐belt, completion of a first‐aid course, smoking and drinking and the like, while competitiveness relates to sporting and game activities. The authors are disarmingly frank about the difficulties associated with using some of these proxies and the ‘tangential’ nature of some of the variables. 2 predetermined or that do not change over time. It is therefore of great interest to see if alternative methods of data collection can shed light on whether or not there are significant gender differences in psychological factors that could explain gender pay gaps and glass ceilings. 3. Experimental studies and personality differences 3.1 Overview Women and men may differ in their propensity to choose a risky outcome because of innate preferences or because pressure to conform to gender‐stereotypes encourages individuals to modify their innate preferences. In the remainder of this talk, I will outline two sets of experiments that I conducted with co‐authors to investigate the extent to which environmental factors or culture might determine gender differences in economic preferences.5 These experiments were, among other things, designed to elicit preferences for risk‐taking. In these experiments, we explored the role that culture might play in affecting economic preferences.6 Our first set of experiments used secondary school students as subjects, while the second set of experiments used first‐year university students. It is well‐known that the academic achievement of girls and boys responds differentially to co‐education, with boys typically performing better and girls worse than in single‐sex environments (Kessler et al., 1985; Brutsaert, 1999). Moreover, psychologists argue that the gendered aspect of individuals' behaviour is brought into play by the gender of others with whom they interact ( Maccoby, 1998). Our main conjecture in these two sets of experiments was that a same‐sex environment may modify preferences in an economically important way. Studies show that there may be 5 Recent laboratory‐based experiments show that, when given the choice of whether or not to enter tournaments, women do indeed `shy away from competition' while men might choose to compete too much (see inter alia Gneezy, Niederle and Rustichini (2003); Datta Gupta, Poulsen, and Villeval, 2005; Niederle and Vesterlund, 2007). Understanding why women seem less inclined than men to compete may provide insight into why a gender gap still exists in the workplace. 6 For a paper looking at the impact on competition of another cultural variable – whether a society is patriachal or matrilineal – see Gneezy, Leonard and List (2009). 3 more pressure for girls to maintain their gender identity in schools or colleges where boys are present than for boys when girls are present (Maccoby, 1990; Brutsaert, 1999). In a coeducational environment, girls are more explicitly confronted with adolescent subculture (such as personal attractiveness to members of the opposite sex) than they are in a single‐sex environment (Coleman, 1961). This may lead them to conform to boys' expectations of how girls should behave to avoid social rejection (American Association of University Women, 1992). If competitive behaviour or risk avoidance is viewed as being a part of female gender identity while risk‐seeking is a part of male gender identity, then being in a coeducational school or college environment might lead girls to make less competitive and risky choices than boys. 3.2 First experiment 3.2.1 Experimental design I noted above that women and men may differ in their propensity to choose a risky outcome because of innate preferences or because pressure to conform to gender‐stereotypes encourages them to modify their innate preferences. Single‐sex environments are likely to modify students’ risk‐taking preferences in economically important ways. To test this, in our first set of controlled experiments, Booth and Nolen designed an experiment in which subjects were given an opportunity to choose a risky outcome – a real‐stakes gamble with a higher expected monetary value than the alternative outcome with a certain payoff – and in which the sensitivity of observed risk choices to environmental factors could be explored. In Booth and Nolen (2012b), we investigated if individuals’ risk preferences are affected by (i) the gender composition of the group to which they were randomly assigned7, and (ii) the gender mix of the school they attended. The latter represented longer‐run nurturing experiences, while the former captured short‐run environmental effects.8 Our 260 subjects, 7 While this group effect has been explored in previous work by Gneezy et. al. (2003), Niederle and Yestrumskas (2007) and Datta Gupta et. al. (2005), those studies focused on competitive tasks. They did not investigate risk attitudes nor did they explore how risk preferences may change over time – the main focus of our investigation. 8 In a companion paper, Booth and Nolen (2012a) investigated how competitive behaviour (including the choice between piece‐rates and tournaments) is affected by single‐sex experimental peer‐groups and single‐sex schooling. In this talk I focus for ease of exposition only on that part of each of our experiments dealing with risk. 4 from eight publicly funded single‐sex and coeducational schools in the counties of Essex and Suffolk in the UK, were asked to choose between a real‐stakes lottery and a sure bet. Four of the schools were single‐sex. The students were from years 10 or 11, and their average age was just under 15 years. After being bused to the University of Essex, students from each school were randomly assigned into 65 groups of four. Groups were of three types: all‐girls, all‐boys or mixed. Mixed groups had at least one student of each gender and the modal group comprised two boys and two girls. The composition of each group – the appropriate mix of single‐sex schools, coeducational schools and gender – was determined beforehand. Thus only the assignment of the 260 girls and boys from a particular school to a group was random. The school mix was two coeducational schools from Suffolk (103 students), two coeducational schools from Essex (45 students), two all‐girl schools from Essex (66 students) and two all‐boy schools from Essex (46 students). The payments (both the show‐up fee of £5 plus any payment from performance in the randomly selected round) were in cash and were hand‐delivered in sealed envelopes (clearly labelled with each student’s name) to the schools a few days after the experiment. The average payment was £7. In addition, immediately after completing an Exit Questionnaire (eliciting demographic information), each student was given a bag containing a soft drink, packet of crisps and bar of chocolate. In the county of Suffolk, there are no single‐sex publicly funded schools. In the county of Essex, the old ‘grammar’ schools remain, owing to an accident of political history.9 These grammar schools are single‐sex and, like the coeducational schools, are publicly funded. It is highly unlikely that students themselves actively choose to go to the single‐sex schools. Instead Essex primary‐school teachers, with parental consent, choose the more able Essex children to 9 In the UK, schools are controlled by local area authorities but frequently ‘directed’ by central government. Following the 1944 Education Act, grammar schools became part of the central government’s tripartite system of grammar, secondary modern and technical schools (the latter never got off the ground). By the mid‐1960s, the central Labour government put pressure on local authorities to establish ‘comprehensive’ schools in their place. Across England and Wales, grammar schools survived in some areas (typically those with long‐standing Conservative boroughs) but were abolished in most others. In some counties, the grammar schools left the state system altogether and became independent schools; these are not part of our study. However, in parts of Essex, single‐sex grammar schools survive as publicly funded entities, whereas in Suffolk, they no longer exist. 5 sit for the Essex‐wide exam for entry into grammar schools.10 Parents must be resident in Essex for their children to be eligible to sit the entrance examination (the 11+). However, residential mobility across regions is very low in Britain (Boheim and Taylor, 2002). To attend a grammar school, a student must apply and then attain above a certain score, which varies from year to year. Therefore, students at the single‐sex schools are not a random subset of the students in Essex, since they are selected based on measurable ability at age 11. In that part of the experiment directed at risk, we had girls and boys chose between Option 1 (£5 for certain) and Option 2 (flip a coin and get £11 if the coin came up heads or £2 if the coin came up tails). Clearly the expected monetary value of the risky option, Option 2, exceeds the certain outcome in Option 1. The dependent variable in our analysis took the value one if the individual chose to enter the lottery and zero otherwise. The implied coefficient of relative risk version (CRRA) was 0.8. We imposed the coefficient of relative risk aversion because we had limited resources and a limited number of rounds. Table 1 below shows the marginal effects of those probit regressions. 3.2.2 Results The first column of Table 1 shows that, on average, girls choose to enter the lottery 16 percentage points less than boys. The sign and significance of this coefficient is consistent with other work looking at gender and risk aversion and suggests that, in our sample, female students are also more risk averse than male students. This provides evidence for the hypothesis that women are, on average, more risk averse than men. The reader might be interested to know that part of observed risk differences between men and women can be manipulated by framing operations, as shown in Booth and Nolen (2012c). However, that is not the focus of this lecture, where we are interested in environmental factors rather than in the way the question eliciting risk preferences is framed. 11 10 If a student achieves a high enough score on the exam, s/he can attend one of the 12 schools in the Consortium of Selective Schools in Essex (CSSE). The vast majority of these are single‐sex. The four single‐sex schools in our experiment are part of the CSSE. 11 Risk theories typically assume individuals make risky choices using probability weights that differ from objective probabilities. Recent theories suggest that probability weights vary depending on which portion of a risky environment is made salient. Booth and Nolen (2012c) used experimental data to show that salience affects young 6 Next, we wanted to investigate if the gender differences alter when environmental factors reflecting nurture are incorporated into the probit estimation. The specification in column [2] adds controls for school type and experimental group composition. In this specification, the gender gap becomes even more pro‐nounced – girls in coed schools choose to enter the lottery 36 percentage points less than boys from coeducational schools. Furthermore, we have evidence that nurture has an effect on risk preferences. First, the coefficient for being in an all‐girls group is statistically significant and positive: girls randomly assigned to all‐girl groups are more likely to choose to enter the lottery. Because our estimates show that girls assigned to single‐sex peer groups are less risk averse than those who are assigned to mixed‐ gender groups, evidence is provided in support of our hypothesis that girls in same‐gender experimental groups are less risk averse than girls in mixed‐gender experimental groups.12 Notice that the same‐gender peer group is only affecting girls; the all‐boys coefficient is insignificant. Second, the single‐sex school coefficient is statistically insignificant but the coefficient to single‐sex schooling interacted with female is significant and positive. Therefore school background only affects the risk preferences for girls at this level of risk aversion, and has no effect on boys. The risk preferences of boys are not affected by either environmental variable, whereas the risk preferences of girls are significantly affected by both environmental factors. Column [3] adds in some ‘ability’ controls, namely the score students obtained in the mandatory completion of mazes in Rounds 1 and 2 of the experiment. These rounds were paid on the basis of piece rates in Round 1 and a tournament in Round 2. These ‘ability’ variables are denoted as the number of mazes correctly completed in Round 1 (R1) and the difference men and women differently. We found that men are significantly more likely than women to switch from a certain to a risky choice once the upside of winning is made salient, even though the expected value of the choice remains the same. Quite why this might occur remains a topic for future research. However, our finding of gender differences in the probability of being affected by salience has an additional implication, namely part of observed risk differences between men and women can be manipulated by framing operations. 12 The all‐girls coefficient is robust to different types of analysis. For instance, if regressions are run on sub‐ samples comprising only students from coed schools, or only students from single‐sex schools, the all‐girl coefficient is still significant. Thus, there is a positive effect of being in an all‐girls group for each type of student. Furthermore, if dummy variables for mixed‐gender groups with three or two boys are used as controls, the significance of the all‐girls coefficient does not go away. 7 between the number of mazes correctly completed in Rounds 1 and 2 (R2‐R1). The estimates of interest ‐ gender, group‐type and single‐sex schooling and its interaction with gender ‐ are robust to the inclusion of these variables. Column [4] adds in a number of additional controls that are listed in the note under the table (and whose coefficients are not included in this table). Again the results of interest are robust to this change of specification. Table 1: Dependent Variable (0,1) If Student Chose Option 2 in the Lottery COEFFICIENT [1] [2] [3] [4] Female (=1) ‐0.16*** ‐0.36*** ‐0.37*** ‐0.34*** [0.05] [0.07] [0.07] [0.11] Single‐Sex (=1) ‐0.13 ‐0.13 ‐0.06 [0.10] [0.10] [0.18] Female * Single‐Sex 0.33*** 0.33*** 0.30** [0.06] [0.06] [0.12] All‐Girls Group (=1) 0.12* 0.12* 0.14** [0.06] [0.06] [0.06] All‐Boys Group (=1) ‐0.05 ‐0.04 ‐0.05 [0.10] [0.10] [0.11] Maze Score R1 ‐0.01 [0.03] Maze Score R2 ‐ R1 0.02 [0.02] Marginal Effect for Female = ‐0.07 ‐0.06 ‐0.03 Single‐Sex = Female * Single‐Sex = 1 [0.05] [0.05] [0.06] Observations 260 260 260 260 Columns [1]‐[3] use entire sample. Col [4] uses only students from single‐sex schools, students who took 11+ exam, and students from Suffolk. Controls: Mother went to University (=1); Father went to University (=1); Number Brothers; Sisters; Student aged 14 (=1). This information was obtained from a post‐experiment questionnaire. Robust SEs in brackets; *** p<0.01, ** p<0.05, *p<0.1. 8
Description: