Incentive Effects of Parents’ Transfers to Children: An Artefactual Field Experiment

Note that primogeniture (the law or custom for the firstborn to inherit all of the parent’s assets) would also accomplish the same result as would any other constant division rule.

As Rosenzweig and Wolpin (1993) show, focusing solely on financial transfers misses the important effects that in-kind transfers and shared living arrangements can have on an individual’s well-being.

Pemberton (1996) proposes that the prevalence of equal giving can be explained by par­ents’ preference to give equally. This argument, however, comes close to assuming its own conclusion. Bernheim and Severinov (2003) propose a model where children view parents’ transfers as signaling parents’ affection for them to explain equal division of bequests as well as unequal division of inter-vivos transfers. Light and McGarry (2004) contend that proxies for altruistic exchange and evolutionary motives all significantly increase the probability of unequal bequests.

These experiments did not control for family structure or composition. For expositional consistency, we use the female pronoun when referring to the parent and the male pronoun when referring to the children.

The participants in this experiment are family members who have an ongoing relationship. Accordingly, even the up-to $100 payout involved in the experiment are very small relative to the life-time incomes of the participants’ children. The curvatures of the utility functions involved are not likely to play a significant role in determining children’s’ preferences over various outcomes with respect to the uncertainty we introduce (see McClelland, Schulze, and Coursey 1993, Rabin 2000). Therefore, when looking at when it might pay for Child B to bet, we will compare expected payoffs rather than expected utility to make the analysis simpler.

Bernheim and Severinov (2003) show that equal division may emerge as a social norm if parents’ transfers signal their affection to the children.

During the actual experiments, the stages were identified by number only.

See the Appendix for complete experimental instructions and screenshot of the computer interface.

The parent’s transfer can be any amount in one cent increments up to $1. To facilitate the parent’s decision-making process and to ensure that the transfers to the two children summed up to $1, the computer software automatically updated the remaining amount to be given to the other child as the parent entered the transfer for one child. After submitting their decisions, parents were presented with a confirmation screen which again displayed both the transfer amounts and the pre-and post-transfer period incomes implied by their decisions. At that point, parents could either confirm their decisions or go back and edit their decisions until they were satisfied. Upon the parent’s confirmation of the transfer decision, the software displayed the outcomes to all members of the family.

In the first eight sessions, there were five actual decision periods following practice. In all other sessions, there were 10 actual periods in the pure transfer stage.

While we do not claim that the incentive problem used in this experiment has an exact analogue in real life, the design is intended to capture the idea that there are choices that parents would prefer their children not make, and they cannot observe whether their children’s outcomes were due to bad choices or bad luck. The fact that parents‘behavior changes in this setting with such relatively small stake suggests that parents are sensitive to these issues and thus would likely change their behavior in higher stakes settings.

In one-child families there could be other somewhat credible beneficiaries such as more distant relatives, friends or charitable organizations. However, as Bernheim, Schleifer and Summers (1985) point out, the task of credibly committing to not giving to the sole child is considerably more difficult than situations where the parent has multiple children.

We expect children in the experiment to have some familiarity with the notion of probabilities because the topic is introduced in the fourth grade and covered in state exams.

In the first 20 sessions, a brief “cheap talk” break in the middle of the moral hazard stage was introduced. During the break, families were allowed to talk about anything they wanted in private for up to five minutes. Since participants were not told about this break in advance, in this analysis, we only use the data from periods before the cheap talk break was announced because there are too few post-cheap talk periods.

Recruiting university staff increased the diversity of the sample as approximately 80% of the parents recruited through the school district had a college degree, while when the latter sample is included, this proportion falls to 66%. Similarly, 85% of the parents from the school district recruitment were married and that percentage dropped to 76% when university staff were included.

Specifically, parents may try to behave in what they think would make the researchers think of them as “good parents”.

This analysis assumes that if parents are compensating in their transfer behavior, they are doing so on a period-by-period basis. In addition, period-by-period decisions may be cognitively easier for parents. However, parents might follow a strategy of trying to equalize the cumulative sum of children’s post-transfer period incomes. To check for this possibility, we constructed vectors of implied transfers under the assumption of cumulative compensating behavior. The fit was worse than both the period-by-period compensation and equal division hypotheses.

We recognize that there is the possibility of parents undoing the effects of their transfers outside the laboratory. However, this type of behavior makes little sense in the context of our design, because parents have the tools within the experiment itself to discourage inefficient behavior by making transfers independent of children’s observed incomes. A post-experiment survey of families in the second recruitment wave asked questions regarding post-experiment activities, and 95.8% of the parents reported that they did not redistribute their children’s incomes after the experiment. The only family deviating from this pattern decided to pool their money to buy new electronic equipment to share in their home.

A random effects model is an appropriate way to account for the possibility of unobservables that are specific to each family. The random effects model assumes that the unobservables are uncorrelated with the covariates, and this is satisfied, because the covariates are constructed as a function of the exogenous random draw.

In addition to the random effects model shown in Table 3, we also ran fixed-effects and OLS models. The estimated coefficients were stable across various models. Using the standard Lagrange multiplier test of the classical regression model versus fixed/random effects models, the classical regression model was not rejected. These results are available by request.

We also ran a model where the predictor was constructed by assuming the parents were trying to equalize the cumulative sum of post-transfer incomes of the children. Given the randomness in children’s pre-transfer incomes, it is not clear if playing catch up in this way is a more effective way of equalizing the sum of the children’s incomes than compensating for differences in each period. In any case, the analogue of the joint restriction above in this case was rejected in both stages.

Clearly, this is not an exhaustive enumeration of all possible signals parents can infer from children’s observed incomes. However, it does constitute a very transparent partitioning of signals.

Given that situations where a bet would not have revealed their decisions were those where children’s random draws were relatively small, one possible way to interpret this is that children exhibit some sort of loss aversion or risk aversion for small stakes (see Rabin 2000 for additional information about this type of behavior).

Participants were asked to enter their first names so that the computer could correctly display information about family members using their names rather than random sequences. The names entered were only stored only during the session and not saved with the experiment data. The group identifier was always a randomly selected four digit integer.