«Building and Rebuilding Trust with Promises and Apologies Eric Schniter and Roman Sheremeta and Daniel Sznycer Online at ...»
Perhaps as a consequence of promise-keepers’ demonstrated trustworthiness in Game 1, the Game 2 investment rate of 92.3% (143/155) in Game 1 promise-keepers was higher than the overall investment rate of 83.4% (191/229) in Game 1 (Fisher's exact test, p-value0.01, n1=155, n2=229). Specifications (1) and (2) in Table 3 indicate that messages with content and promises closer to the even split of $10 positively affected investment in promise-keepers.5 We expected that, due to the greater reliability of available behavioral cues which demonstrated their trustworthiness (relative to the reliability of a cheap signal), Game 1 promisekeepers would be less inclined than promise-breakers and distrusted trustees to construct messages for the purpose of persuading investors to choose IN in Game 2, and so would send both shorter messages and a higher proportion of empty messages in Game 2. Supporting that expectation, Game 1 promise-keepers’ messages contained fewer words than those from the set including both Game 1 distrusted trustees and Game 1 promise-breakers (M1=11.41 (SD1=11.94) versus M2=22.9 (SD2=22.37); Wilcoxon rank-sum test, p0.01, n1=155, n2=74). Game 1 promise-keepers’ messages were also more frequently empty than those from the set of both Game 1 distrusted trustees and Game 1 promise-breakers (20.6% versus 10.8% of the time;
Fisher’s exact test, p-value=0.05, n=229).
Investments in Game 1 promise-keepers paid off for investors choosing IN in Game 2.
These investors received an average of $8.62 from trustees, as opposed to the $5 earned from Note that in estimating these regressions we cannot include both returns in Game 1 and promises in Game 1 since for promise-keepers they are perfectly correlated. Moreover, we had to omit variable Promise211 × Promise2 since there are only three observations greater than $11, which makes Promise211 and Promise211 × Promise2 almost perfectly correlated.
OUT (Wilcoxon signed rank test, p-value0.01, n=143), with 83.9% (120/143) of the promises kept or exceeded, and 16.1% (23/143) broken. Compared to $0 earned by those Game 1 promisekeepers that were not trusted in Game 2, promise-keepers also profited from trusted promises in Game 2, earning $11.38 on average (Wilcoxon rank-sum test, p-value0.01, n1=12, n2=143).
Specifications (3) and (4) in Table 3 indicate that Game 2 promises lower and higher than the even-split are associated with lower and higher amounts returned by Game 1 promise-keepers, respectively. Whether the messages have content or not, on the other hand, has no effect on returns.
4.2.2. Promise-Breakers A major question our data address concerns what happens after a breach of trust when a fresh opportunity for cooperation arises: How trustees behave, how investors respond, and what outcomes are achieved. Here we focus on the 18.8% (36/191) of pairs with broken promises in Game 1 (i.e., where the amount returned was lower than the amount promised). These broken promises represent breaches of trust and the relationships that immediately follow are considered to have damaged trust (because trust-based expectations were not met). A central question motivating this study is whether signals such as new promises and apologies can (i) restore investors’ willingness to trust and (ii) facilitate the achievement of higher joint payoffs.
Figure 4 displays the distribution of promises made by 36 Game 1 promise-breakers in Game 2 resulting in IN and OUT decisions. Promise-breakers promised $12.11 in Game 2, which is significantly higher than their average promise of $10.58 in Game 1 (Wilcoxon signed rank test, p-value=0.01, n1=n2=36). The extent of upgraded promises (Promise2-Promise1) by promise-breakers is also significantly higher than the extent of upgraded promises by promisekeepers (M1=1.53 (SD1=3.70) versus M2=0.44 (SD2=1.79); Wilcoxon rank-sum test, pvalue0.01, n1=36, n2=155). Assuming that many of the investors whose trust had previously been damaged would be inclined to choose OUT, it appears that promise upgrades partially restore trust, since 69.4% (25/36) of investors whose trust was damaged in Game 1 chose IN again.
In addition to promise upgrades, we also find that Game 1 promise-breakers frequently used messages. Table C1 in Appendix C reports all messages that were sent by 36 promisebreakers. Analyzing the messages, we find that 83.3% (30/36) of the messages have some content.6 Game 1 promise-breakers’ messages contain more words than messages from Game 1 promise-keepers (M1=19.06 (SD1=19.03) versus M2=11.41 (SD2=11.94); Wilcoxon rank-sum test, p-value=0.03, n1=36, n2=155), suggesting that the behavioral cue of trustworthiness made verbal persuasion for continued investment less determinant of re-investment and less necessary for trustees.
To further classify all 36 messages, we used an incentivized laboratory coordination game (Houser & Xiao 2011). Three coders recruited from the subject pool and blind to the hypotheses7 were asked to code each message twice: first based on whether or not it conformed to a “broad” definition of apology (an explicit or implicit acknowledgment of offense), and second based on whether or not it conformed to a “narrow” definition of apology (an explicit or implicit acknowledgment of offense, along with remorse, regret, or sorrow stemming from acknowledgment of the offense). All 6 messages without content were coded by all coders as not conforming to the broad definition and not conforming to the narrow definition of apology. Of the 30 messages with content, 28 were coded by the majority of coders as conforming to the broad definition of apology and 13 were coded by the majority of coders as conforming to the narrow definition of apology.8 When using a broad definition of apology, which was coded with “substantial” agreement (Kappa of 0.70), we find that 82.1% (23/28) of apologizers were retrusted in Game 2 in comparison to only 25.0% (2/8) of non-apologizers (Fisher’s exact test, p-value0.01, n=36), suggesting that messages with apology are more likely to restore trust after broken promises than empty messages or messages without apology. 9 Due to the lower interrater reliability for messages coded according to the narrow definition, we will consider only the broad definition of apology in the subsequent analyses.
We find that 80% (24/30) of messages with content restored trust (i.e., where investors chose IN in Game 2 after having suffered broken promises in Game 1), as opposed to only 16.7% (1/6) of messages without content. These differences are significant (Fisher’s exact test, p-value0.01, n=36).
The instructions for coders and details about how they were paid are attached in Appendix B. Coders each earned an average of $28.33 for matched codings, plus $7 for arriving on time and participating.
We use a standard approach from content analysis methodology to calculate Cohen’s Kappa interrater agreement coefficient (Cohen 1960; Krippendorff 2004). Kappa values between 0.41 and 0.60 are considered “Moderate” agreement, and those above 0.60 indicate “Substantial” agreement (Landis & Koch 1977). We find Kappa values of
0.70 and 0.53 for the broad and narrow definitions of apology, respectively.
When using a narrow definition of apology, which was coded with moderate agreement (Kappa of 0.53), we find that 84.6% (11/13) of apologizers were retrusted in comparison to only 60.9% (14/23) of non-apologizers (Fisher’s exact test, p-value=0.13, n=36).
Thus far, we have only considered the independent effects of new promises and apologies in restoring damaged trust, but recognize that these remedial strategies are often used jointly.
Among Game 1 promise-breakers, the size of the upgrade in amount promised is significantly larger for participants issuing apologies than for those who did not (M1=$1.68 (SD1=3.10) versus M2=$1 (SD2=5.55); Wilcoxon rank-sum test, p-value=0.05, n1=28, n2=10). When the apologetic promise-breakers are compared to all other trustees the difference is even larger. The upgrade in amount promised for apologetic trustees is almost four times greater than among all other trustees (M1=$1.68 (SD1=3.10) versus M2=$0.44 (SD2=2.68); Wilcoxon rank-sum test, pvalue=0.05, n1=28, n2=201), indicating that apologetic trustees upgraded their promises most.
Next, we estimate probit regressions (see Table 4) to identify how these remedial strategies work in conjunction. Specification (1) indicates that the two most significant predictors of trust in Game 2 are promise adjustments (specifically promise upgrades) and apologies.
Specification (2) shows that in addition trust is negatively affected by the extent of under-return relative to Game 1 promise. Moreover, based on the likelihood-ratio test, we find that the promise adjustment, upgraded relative to the extent of under-return on previously broken promise (i.e., Promise2-(Promise1-Return1)), positively and significantly influences trust in Game 2 (likelihood-ratio test, p-value=0.05). These results indicate that investors’ decisions to re-invest are sensitive not just to the existence of broken promises (specifications 1 and 2 in Table 2), but also to the extent of under-return relative to Game 1 promise, apologies, and upgraded promises (specifications 1 and 2 in Table 4).
We have argued that signals such as apologies and promises should have evolved only if they provided net benefits to both the senders and receivers of those signals on average and over the evolution of the communication system. We evaluate whether Game 1 promise-breakers’ signals resulted in benefits for both investor and trustee in Game 2, and whether these signals were reliable indicators of subsequent trustee behaviors. Investors in Game 1 promise-breakers were returned $7.28 on average, which is significantly higher than the OUT payoff of $5 (Wilcoxon signed rank test, p-value=0.05, n=25). Moreover, Game 1 promise-breakers returned significantly more in Game 2 than in Game 1 (M1=$7.28 (SD1=4.86) versus M2=$4.60 (SD2=3.72); Wilcoxon signed rank test, p-value0.01, n1=n2=25). This is also true when we look
at investments in the subset of 23 out of 28 trustees who issued apologies and where retrusted:
they returned significantly more in the second game (M1=$7.52 (SD1=4.81) versus M2=$4.61 (SD2=3.64); Wilcoxon signed rank test, p-value0.01, n=23), which is also significantly higher than the OUT payoff (M1=$7.52 (SD1=4.81) versus M2=$5; Wilcoxon signed rank test, pvalue=0.03, n=23).
Although investments in Game 2 paid off, we still find that 60.0% (15/25) of re-trusted promise-breakers subsequently broke their promises again in Game 2 – almost irrespective of the apologies and new adjusted promises. From specifications (3) and (4) in Table 4, it appears that neither promises adjustments, new amounts promised, nor apologies are predictive of return in Game 2. The only variable that predicts return in Game 2 is return in Game 1.
4.2.3. Distrusted As mentioned above, 16.6% of trustees (38 out of 229) were not trusted in Game 1 (see Figure 1). The source of this distrust appears to be: (i) the higher variance around the even-split point of the distribution of distrusted promises (relative to that of the distribution of trusted promises; see right panel of Figure 2), and (ii) a lower degree of default trustfulness among Game 1 distrustful investor (accounting for the fact that a sizeable number of even-split promises were rejected). In particular, in Game 1, 55.3% (21/38) of distrusted trustees promised less than $9 while another 10.5% (4/38) of them promised more than $11. As with our Game 1 predictions of trusted promises, we expect that previously distrusted trustees would adjust their Game 2 promises towards the modal and more trusted promise of $10, that these adjustments would affect decisions to invest, and that investments made based on adjusted promises would benefit both the investor and trustee.
First we evaluate whether Game 1 distrusted trustees adjust their promises as we expected, and if adjustments of promises by Game 1 distrusted trustees affect investment decisions. Trustees who were distrusted in Game 1 promised an average of $8.92 in Game 2, which is similar to their average promise of $8.61 in Game 1 (Wilcoxon signed rank test, pvalue=0.45, n1=n2=38), yet most investors (84.2% or 32/38) who did not trust in Game 1 chose IN in Game 2. Figure 5 displays the histogram of promises made in Game 2 by the 38 trustees who were distrusted in Game 1. Distrusted trustees changed their distribution of promises towards more equal splits: 66.7% (14/21) of trustees who promised less than $9 in Game 1 increased their Game 2 promises and 100% (4/4) of trustees who promised more than $11 in Game 1 decreased their Game 2 promises. Correspondingly, among previously un-trusting investors, 92.6% (13/14) of those who received increased promises and 100% (4/4) of those who received the decreased promises chose IN in Game 2.
Next, we analyze whether new trust in previously distrusted trustees can be statistically attributed to how distrusted trustees utilized messages and adjusted promises. We expect that distrusted trustees would construct longer messages (to persuade investors to choose IN in Game
2) than trustees who had already established reputations of trustworthiness. Table C2 in Appendix C reports the messages that were sent by 38 trustees who were distrusted in Game 1.
Analyzing these messages, we find that 94.7% (36/38) of the messages used by distrusted trustees have some content. Messages from Game 1 distrusted trustees contain more words than messages from Game 1 promise-keepers (M1=26.58 (SD1=24.83) versus M2=11.41 (SD2=11.94);
Wilcoxon rank-sum test, p0.01, n1=38, n2=155). These data suggest that distrusted trustees use both promise adjustments towards 50/50 divisions of income and longer messages to persuade investors to trust them. The estimation of specification (1) in Table 5 indicates that the investment decision in Game 2 is positively correlated with the amount of $10 promised in Game 2 (p-value=0.06), but that message length is not significant. The Game 2 rate of trust-extension for Game 1 distrustful investors was 84.2% (32/38). This is very similar to the original unconditional investment rate of 83.4% in Game 1.