We also thank seminar participants from the Research Unit in Behavioral and Neuroeconomics (Ruben) and the Southern African Labor and Development Research Unit (Saldru). Proponents were randomly paired with same/other race respondents, and asked to play the trust game after viewing a photograph and hearing a 10 second audio clip of the respondents reading a standardized script in English. We find that accentuation is a statistically significant predictor of trust and is strongly nonlinear in the race of the paired subjects for males but not for females.
At the same time, only ten percent of the population speak English as their home language in South Africa, but it remains the lingua franca in the post-apartheid period (SSA 2011).3 Under the contemporary school system. Another somewhat unique feature of our design is that there are no cross-gender pairings between proposer and respondent. We find that accent is a statistically significant predictor of trust and is highly non-linear in the race of the paired male subjects: offers drop by 11.3% if Player B has a native English accent and does not share the same race as Player A, but increase by about 6.6% if they share the same race as Player A.
The rest of the paper is structured as follows: in section 2, we review the relevant evidence on accent discrimination (as opposed to language discrimination). The interaction consists of two phases: during the first phase, both of the players are given the same endowment by the experimenter, after which Player.
Measurement of confounding factors by 3rd Party evaluations
The instructions stated that they were equipped with 50 South African Rand and they now had a chance to send any amount of this R50 to their partner, Player B.12. They were also told that any amount they decided to send would be doubled before they passed it on to Player B. The instructions also provided a private link and password so Player A's could only access their partner's track. A similar process was followed for all Player Bs as well, whereby they saw and heard their partners before a decision was made on the amount they would return, if any.
The evaluators did not participate in the trust game and thus had no interaction with the participants. They were asked to make their best assessment on a host of characteristics related to the candidate they were assessing, including our key variables of interest. Similar to the design of the trust game, the gender of the evaluators and candidates in the excerpts was kept constant.
Modeling fractional responses
Evaluators only had to guess what those scores would be, and the closer their rating was to the data, the more money they would receive. These videos contained images and voice clips of 21-22 randomly selected trust game participants. In particular, evaluators were asked to rate the candidates on attributes such as credibility, trust, attractiveness and timidity, as well as on language attributes ie: speaking English as a first language after watching their clips.
The evaluations were then done by the evaluators who watched the tapes and then answered the evaluation questions. To make the ratings as objective as possible, we assigned 6 raters to the voice recording - 3 white and 3 black. Another approach, also suggested by Wooldridge (2010), would be a log-odds transformation of the dependent variable.
The idea here is that since a fraction is mathematically equivalent to a probability, you can see /(1−y) as an odds ratio. Log-transformation of this pseudo-odds ratio will thus map the result back to reals, and then OLS can be used. Again, however, the problem lies in the fact that this quantity will be undefined for the boundary observations.14 A further concern is that even if 0< y <1, the interpretation is not straightforward.
This approach is a simple extension of the binary logit or probit model, but essentially the log-likelihood function as applied to a Bernoulli distributed random variable will take exactly the same structure for a fractional response and has the added advantage of to map predictions into (0) ,1) (see Papke and Wooldridge (1996) for the details). A further limitation is that one is forced to assume either the logistic or normal distribution for responses to 0< y <1. A further limitation is that although the Bernoulli log-likelihood belongs to the linear exponential family, fractional logit or probit does not directly handle the fact that the conditional variance is a function of the mean.
In a series of articles, Ferrari and Cribari-Neto (2004) and Simas et al. 2010) developed regression models for beta-distributed random variables using a parameterization of the beta law indexed by the mean and dispersion parameters.
Zero-One Inflated Beta Regression
Ospina and Ferrari (2012) extended this original framework to handle fractional reactions involving limit mass points. This new approach, which they call zero-or-one inflated beta regression, forms the basis of our approach, albeit with some minor differences.15. Following Ferrari and Cribari-Neto (2004), we outline the framework using a different parameterization of the beta density.
Since interest in the regression context focuses on modeling the conditional mean, it makes sense to set. To use these distribution functions in a regression context, it is standard to use the GLM framework. In our specifications, we chose the logit link function.16 Up to this point, the estimation would proceed more or less in the same way as fractional regression, with the exception that instead of using the binomial distribution to fit the likelihood function, we use the beta distribution and model the zeros and ones as discrete choices.
Since we are really dealing with a selection problem where the outcome is observed after an experimentally induced treatment, a discrete selection approach for the cutoff points makes more sense, as opposed to assuming that the zeros and ones result solely due to sampling variability. 16 The link function is always needed as we want to avoid predictions outside the unit range. Ospina and Ferrari (2012) showed that this model can be generalized to a context involving extreme values on the closed unit interval.17 Their extension of the GLM framework to cover this case requires degenerate probability statements that produce a mixture density which in turn effectively boils down to an additive term to the log-likelihood function given above.
Specifically, we can think of 3 possible cases: (a) the case of zero inflation (which is the case considered by Ospina and Ferrari (2012)), where a new parameter is added to account for the probability of observing a value at zero. Although adding both zero and one inflation complicates the likelihood function, this complication is only additive (in the sense that two new terms are added to the likelihood function). As shown by Ospina and Ferrari (2012), this estimator can be operationalized by separately fitting a logit/probit regression for the binary outcomes p0 and p1 and then using the resulting predicted probabilities for y = 0 and y= 1 to construct the terms (1− p0 ) and (1−p1).
Descriptive Statistics
In all subsequent regressions, our dependent variable is the fraction of the endowment that player A offers to player B after receiving a signal packet.
Empirical Estimates
Although not statistically significant, the inclusion of the interaction term suggests that offers increase by approximately 7% if Player B's native English speakers. In other words, a native English accent has some advantage in negating the out-group bias. Offers are 17.8% higher if Player B is of the same race as Player A (Co-ethnic pair = 1) and Player B is also English speaking.
In contrast, offers are 11.3% lower if Player B is English-speaking but not of the same race as Player A (Co-ethnic pair = 0). On the other hand, Table 6 shows that women make about 9% lower offers in co-ethnic pairs and this does not seem to depend on whether or not Player B's mother tongue is English. Although being White and speaking English as a mother tongue is almost synonymous, this is not true for Black Player Bs.22 Therefore, if the racial effect is non-linear in accent, it will manifest most sharply among Black Player Bs.
However, this apparent in-group bias of Black Player As is mitigated if Black Player B has a native English accent. 2293% of white students are native English speakers, while only 38% of black students are native English speakers in our sample. For Black males paired with other Black males, Table 8 shows that Player A trusts Player B about 11.5% more if Player B has a native English accent.
This paper focused on how native English accent interacts with the race effect, particularly for black subjects. The results suggest that accents do matter and this result is particularly pronounced for black men as they show higher levels of confidence when hearing a native English accent. Croson, and Sara Solnick, "Trust and Gender: An Examination of Behavior and Beliefs in the Investment Game," Journal of Economic Behavior and Organization.
Grossman, “The Relative Cost of Fairness: Gender Differences in a Punishment Game,” Journal of Economic Behavior Organization, August. Greig, Fiona, “Gender and the Social Costs of Claiming Value: An Experimental Approach,” Journal of Economic Behavior and Organization. Houser, Daniel, Daniel Schunk, and Joachim Winter, “Distinguishing Trust from Risk: An Anatomy of the Investment Game,” Journal of Economic Behavior and Organization.
Kamas, Linda and Anne Preston, "Can Social Preferences Explain Gender Differences in Economic Behavior?", Journal of Economic Behavior and Organization. Dependent variable: proportion of grant offered in the trust game; Additional controls: player A receives financial aid; Player A is South African; Player A's age; The square of player A's age; The interaction is anonymous. M T E stands for "Mother Tongue is English".
