Manipulation checks and descriptive statistics
We first tested for successful randomization of the participants across cells. The Kruskal–Wallis H test indicates successful randomization across cells (H(3)Footnote 2 = 0.233, p > 0.10). Because all the participants are responsible for the initial decision, we control for whether the participants feel responsible.
In relation to escalation of commitment, our results align with self-justification theory which predicts the participants who report feeling responsible for their initial decision show a stronger tendency to justify their behavior than participants who do not feel such a responsibility. Following Whyte (1991), we analyze the influence of the participants’ reported feelings of responsibility for the first decision. The mean level of responsibility is 7.78 (S.D. = 2.229) on an 11-point Likert scale. We use the one-sample t test to investigate whether the mean deviates from the scale medium of 5 (t(200) = 17.262, p < 0.001), indicating that the participants felt a high level of responsibility.
We also check whether our negative-feedback condition indicates a higher level of responsibility compared to the positive-feedback condition. A comparison of means shows that, when facing negative feedback, participants feel more responsible (M = 8.16, S.D. = 1.904) than they do when facing positive feedback (M = 7.41, S.D. = 2.557; t(200) = 2.369, p < 0.05). We verified this result using the Mann–Whitney U test (U = 4129, p = 0.023, r = 0.16) and the Kruskal–Wallis H test (H(1) = 5.151, p < 0.05). Both tests indicated a significant difference. Additionally, we check whether responsibility differs between information load conditions. We use the Kruskal–Wallis H test and find no significant difference (H(3) = 0.980, p = 0.806).
In relation to information load, we checked whether the participants used the KPIs provided. We find that the participants used the additional KPIs provided (β = 0.486, p < 0.001). The Kruskal–Wallis H test shows that we find different uses of KPIs across cells (H(8) = 60.290, p < 0.001).
In relation to investment behavior, Table 2 shows that, in the negative-feedback condition, the allocation of resources to the initially chosen business division (escalation tendency) decreases from the 2 KPI condition over the 4-KPI condition to the 8-KPI condition and then rebounds in the 12-KPI condition. In the positive-feedback condition, the allocation of resources to the initially chosen business division continuously increases with the number of KPIs over the four conditions. Table 3 provides the results of the independent samples test, which indicate different behavior in negative and positive feedback situations.
Table 2 Summary statistics Table 3 Independent samples test Effect of information load in a negative feedback situation
We checked for the theoretical mechanism regarding the effect of self-justification on the allocation of resources, i.e., the escalation of commitment. When we analyze whether the negative or positive feedback that the participants receive after their initial funding decision influences the amount of resources that they allocate in the second part of the experiment we find that the participants who are given negative accounting information (negative feedback M = 12.22) allocate more resources than those who are given positive accounting information (positive feedback M = 10.70; t(201) = 1.915, p < 0.05, d = 1.52), in line with existing research (Schulz and Cheng 2002; Sleesman et al. 2012; Staw 1976). We verified this result using the Mann–Whitney U test (U = 3922, p < 0.01, r = 0.20) and the Kruskal–Wallis H test (H(1) = 7.633, p < 0.01). Both tests indicate a significant difference.
H1 predicts that, when facing negative feedback, increasing the information provided reduces the escalation of commitment up to a certain point.
Planned polynomial contrasts indicate a significant quadratic relationship between information load and escalation behavior (contrast estimate = 2.004 vs. hypothesized value = 0, S.D. = 1.042, p = 0.058).
Drawing on Shields (1983), we calculated a regression to check for the U-shaped relationship. An ordinary least squares (OLS) regression with allocation of resources as the dependent variable and squared KPIs as the independent variable shows a significant U-shaped relationship (β = − 0.112, p < 0.01, F = 3.314, p < 0.10). We also checked whether other functional forms of the relationship between allocation of resources and information load might represent the relationship. Neither a linear OLS regression (F = 0.476, p = 0.492) nor a cubic regression (F = 1.594, p = 0.196) revealed a significant result. Thus, H1 is supported.
We conduct analysis of variance (ANOVA) to test for an interaction effect between the type of feedback and the information load. A 2 (consequence of initial decision: positive feedback/positive accounting information vs. negative feedback/negative accounting information) × 4 (information load: 2, 4, 8, 12 KPIs) ANOVA reveals a positive and significant interaction between the type of feedback and the information load (Table 4). Figure 1 shows the results.
Our findings indicate that, in the negative-feedback condition with 2, 4, and 8 KPIs, more information reduces the escalation tendency. If even more information is provided (12 KPI condition), we find that the escalation tendency increases again.
Research question RQ1 asks how information load and self-justification interact in influencing escalation of commitment when facing negative feedback. We expect that nonlinearities are involved in the interaction between information load and self-justification. RQ1 focuses on the negative-feedback condition as we do not expect self-justification to play a relevant role in the positive-feedback condition. Nonetheless, we include the positive-feedback condition in the following analysis for reasons of comparison.
We use a median split discussed in the recent literature (e.g., Iacobucci et al. 2015). We split the sample into two approximately equal groups (low/high self-justification) and add the median participants to the lower self-justification group (Iacobucci et al. 2015). Table 5 shows summary statistics regarding the two groups of self-justification. It is striking that we observe a broad range of values for self-justification in the positive-feedback condition.
Table 5 Summary statistics regarding self-justification Figures 2 and 3 show the allocation of resources with low and high self-justification of participants under different information load conditions for the negative-feedback and the positive-feedback condition respectively. Our results indicate that the type of feedback affects self-justification. In the positive-feedback condition, we find increasing levels of allocation of resources in the high-self-justification group when information load increases. We check for differences between the low and high self-justification groups in the negative-feedback condition using the Mann–Whitney U test (U = 1149.5, p = 0.737) and the Kruskal–Wallis H test (H(1) = 0.113, p = 0.737). Both tests indicate that there is no significant difference between groups in the negative-feedback condition. Moreover, we check for differences between the low and high self-justification groups in the positive-feedback condition using the Mann–Whitney U test (U = 847.5, p < 0.01, r = 0.29) and the Kruskal–Wallis H test (H(1) = 5.769, p < 0.01). Both tests indicate a significant difference between groups.
We used a moderated mediation model drawing from Hayes (2015) to calculate possible interactions between information load and self-justification. We used information load (i.e., KPI) as the independent variable, self-justification as the mediator, allocation of resources as the dependent variable and type of feedback as the moderator of the three relationships in the mediation model.
Table 6 shows the results of the moderated mediation model. We find that self-justification is affected by type of feedback, as expected. Furthermore, although we do not find a direct effect of information load on self-justification, we find a negative and significant interaction between information load and self-justification. In a detailed conditional effects analysis of the focal predictor at values of the moderator, we find that this effect occurs in negative-feedback cases only (w = − 0.139, p < 0.05, 95% CI = [− 0.258; − 0.020]), while there is no effect in positive feedback cases. This indicates that information load reduces self-justification in some manner.
Table 6 Regression results for moderated mediation analysis In relation to the additive effect of both variables on allocation of resources, we find that both information load and self-justification increase allocation of resources next to the effect of negative feedback. Our results reveal two significant interactions. First, we find a negative moderation of the type of feedback on the relationship between information load and allocation of resources, indicating that negative feedback reduces the participants’ tendency to invest more with increasing information load. Second, our results show that there is a negative and significant moderation of type of feedback on the relationship between self-justification and allocation of resources, indicating that, in a model with information load, self-justification has no consistent increasing effect on allocation of resources in escalation situations.
Effect of information load in a positive feedback situation
Hypothesis H2 predicts that, with the participant facing positive feedback, increasing information load positively affects the resources allocated to a promising course of action up to a certain point and decreases from that point on. The ANOVA in Table 4 has already shown that there is a significant interaction effect between the type of feedback and the information load. Also, in the previous section, we discussed the interaction of information overload and self-justification also for the positive feedback situation. Those outcomes did not examine the functional form of the relationship between information overload and allocation of resources in the positive feedback situation. For that analysis we undertook a procedure similar to that of Shields (1983). We calculated a quadratic OLS regression with allocation of resources as the dependent variable and both a squared information load and a linear information load (i.e., KPI) as independent variables. The OLS regression is significant (F = 7.163, p < 0.01), but we find that the quadratic term is not significant (β = 0.066, p = 0.203), while the linear term is significant (β = 0.889, p < 0.05). Hence, H2 is not supported. Figure 1 in the previous section already suggested this finding.
In the case of positive feedback, we find that increasing the information load leads to higher investments (F(1, 101) = 12.601, p < 0.001, R2 = 0.102, std. β = 0.333, p < 0.05).
An unexpected result was that the relationship is not quadratic but linear. To check for a linear relationship, we calculate an OLS regression with allocation of resources as the dependent variable and a linear information load as the independent variable. This OLS regression is highly significant (F = 12.601, p < 0.01), and the coefficient is positive and significant (β = 0.547, p < 0.01).