https://orcid.org/0000-0002-8668-2772
![Sample our Economics, Finance,Business & Industry journals, sign in here to start your access, latest two full volumes FREE to you for 14 days]({"type":"image" , "src" : "/sda/5931/TOC-Subject-Sample-2021-Economics-Finance-Business-Industry.jpg"})
Abstract
Time estimates are an important part of many costing systems. As they may be subject to estimation error and hence result in measurement error in product and service costs, it is important to understand the sources of this estimation error as well as mechanisms to mitigate them. We performed a computer-based lab experiment to test the joint impact of task interruption and interactive time estimation on the accuracy of time estimates. As predicted, we find that the negative impact of task interruption on time estimation accuracy is mitigated when participants are allowed to interactively discuss their time estimates before making a final time estimate alone. We explain our findings through the underlying effects of cognitive load and confidence. Hence, managers may improve the accuracy of their time-based costing systems by enabling an interactive time estimation process to offset the detrimental effect of task interruption.
Acknowledgements
We would like to thank participants at the EAA Annual Congress (Milan, 2018), the MCA conference (Groningen, 2017), the ENEAR conference (Leuven, 2017), the FEB Research Day (Ghent, 2017), the Research Day in Accounting (Antwerp, 2017), the EDEN doctoral seminar on Producing and Evaluating Knowledge in Management Accounting (Brussels, 2016), the Limperg course on experiments (Amsterdam, 2015), and brownbag seminars at Ghent University and Emory University for useful comments and suggestions. We are also grateful for the helpful comments on earlier drafts of this paper provided by Werner Bruggeman, Eddy Cardinaels, Patricia Everaert, Maggie Geuens, Laura Gómez, Katlijn Haesebrouck, Karl Schuhmacher, Ivo Tafkov, and Dennis Veltrop. We thank the associate editor Victor Maas and two anonymous reviewers for their very helpful and constructive comments as well as their deep engagement with our work during the entire review process.
Disclosure statement
No potential conflict of interest was reported by the author(s).
Supplemental Data and Research Materials
Supplemental data for this article can be accessed on the Taylor & Francis website, doi:10.1080/09638180.2020.1826339.
Appendix A. Screen Shots Experiment
Appendix B. Alternative Dependent Variables
Notes
1 Moreover, RFID devices may introduce moral value (i.e., privacy) issues (Kurkovsky et al., Citation2011) despite their higher accuracy and direct observation may be prone to the Hawthorne effect (Keel et al., Citation2017).
2 Note that both under- and overestimation of time increase estimation error. Following Brown (Citation1985) and Brown and Boltz (Citation2002), we use absolute error scores to assess the overall level of inaccuracy of time judgments. This is also in line with accounting settings, where over- and underestimation of time (and hence costs) are both detrimental (Labro & Vanhoucke, Citation2007).
3 The effect of cognitive load on time estimates depends on the duration judgment paradigm (Block et al., Citation2010). In the prospective paradigm, participants are informed in advance that they will have to provide time estimates, which is not the case in the retrospective paradigm. Since previous studies have concluded that prospective time estimates are more accurate than retrospective time estimates (Block & Zakay, Citation1997; Brown, Citation1985; Cardinaels & Labro, Citation2008), we decided to situate our study in the prospective paradigm. Hence, we focus on the attentional-gate model and do not elaborate on memory-based models, which predict the effect of cognitive load on retrospective duration judgments (e.g., Ornstein, Citation1969).
4 We believe that, in our study, although task interruption involves task switching, it does not involve mindset switching. Hamilton et al. (Citation2011) conclude that mindset switching results in depleted executive resources and in worse subsequent self-regulation. If task interruption did involve mindset switching, this could be an alternative explanation for the effect of task interruption on estimation error. Although we believe our experimental tasks only involve the actional mindset without switching to a deliberative, implemental, or evaluative mindset (Gollwitzer, Citation1990), we cannot rule out this alternative explanation entirely.
5 We do not state a formal hypothesis for the effect of task interruption on time estimation error, since the attentional-gate model is a well-established model to account for the effect of cognitive load on prospective time judgments (Block et al., Citation2010 used 82 prospective time judgment experiments in their meta-analytic review to test this relation). However, in the results section, we will demonstrate that task interruption indeed increases cognitive load.
6 Although one might argue that interactive estimation could increase judgment accuracy because of the possible deliberation process that stimulates individuals to reflect in a more structured way (as opposed to information collection), prior research shows that intuitive judgments (i.e., without deliberation) are more accurate than judgments after deliberation (Dijkstra et al., Citation2012). In particular, Haberstroh (Citation2008) argues that intuitive frequency estimates are based on an implicit frequency record formed by, for instance, a cognitive counter (see Haberstroh, Citation2008, p. 270). Since we relied on the attentional-gate model, which assumes a cognitive counter (Zakay & Block, Citation1995) to predict how individuals estimate time spent, individuals’ intuitive time estimates can be expected to be more accurate than their deliberate time estimates. As such, prior research (Haberstroh, Citation2008) suggests that deliberation can be ruled out as an alternative process because it would predict the opposite effect of interactive time estimation than the one we predicted and found.
7 This prediction is primarily composed of two simple effects: (1) the simple effect of task interruption given individual time estimation and (2) the simple effect of interactive time estimation given an interrupted task setting.
8 Previous studies on task interruption conclude that interruptions have a negative impact on participants’ performance when they occur in the middle or at the end of the primary task (e.g., Freeman & Muraven, Citation2010).
9 We interviewed five practicing costing consultants to verify the relevance of the different time estimation units in practice. Although, in the past, percentage-time estimates were often used in ABC settings, the interviewees indicated that they now try to avoid percentage-time estimates because of their limited insight into capacity utilization (i.e., reported percentages often add up to 100). Time-based costing systems often rely on total-time estimates as well as on combinations of all three time estimation units. We wish to emphasize the difference between time-driven and time-based costing systems. That is, ‘time-driven’ specifically refers to time drivers in time equations (which require average unit-time estimates), while ‘time-based’ means that, in general, cost allocations are based on time registration. The interviewees explained that in time-based costing systems, which they commonly implement for their clients, employees often need to estimate the total time spent on one project, task or process, such as the time doctors and nurses spend on patient rounds, the time nurses spend on patient room visits while performing different tasks, the time consultants or non-profit organizations spend on a particular project, the time shipping agents spend on a particular shipping or on loading and unloading the transporter, the time employees spend on reconfiguring weaving machines to start a new fabric, or the time installers spend on installing utilities (e.g., water connection, electricity connection, etc.). Hence, according to the five subject matter experts that we interviewed, all three time estimation units are used in practice.
10 For the decoding task only 50.7% and for the dragging-and-dropping task only 41.9% of the interrupted participants pointed out on the manipulation check question (‘Have you been interrupted during the decoding of sentences?’ and ‘Have you been interrupted during the filling of pallets?’) that they had been interrupted. However, task interruption is a factual manipulation (i.e., we programed the interruptions to occur) such that we are sure that all participants in the interruption conditions actually were interrupted. In addition, we checked the data and all participants in the interrupted conditions did perform the interruption tasks. As such, we are confident the manipulation was indeed effective. After a conversation with some of the participants, we realized that the manipulation check question was not clear, because they did not understand what we meant by ‘interruptions.’ Some of them interpreted an ‘interruption’ as an external event and not as an interruption task as part of the study. Hence, the unexpected outcome of the manipulation check question is due to a failure of its formulation and not to a failed manipulation.
11 In our manipulation of interactive time estimation, participants did not provide us with an individual estimate before entering the discussion, i.e., we did not include a ‘box’ in which they had to provide their initial/individual estimate. They only provided their final estimate (after interaction). Although the content of the chat boxes was saved, analyzing participants’ estimates within this chat box is not appropriate to obtain their individual estimates. More specifically, the estimate of the second responder in the chat box may be influenced by the first mover’s estimate (Koehler & Beauregard, Citation2006), such that we are unable to assess whether the provided estimates of the responders are their true ‘first guess’ (i.e., without the influence of the first mover). Therefore, it is not appropriate to assume that participants’ first estimate in the chat box is their true individual estimate. We deliberately made this design choice because we are interested in participants’ interactive estimates and not in whether they changed their first individual estimate.
12 To double-check whether this manipulation was effective, we asked the participants in the interactive condition whether they used the chat box (‘Did you use the chat box?’). They had three options: 1) no; 2) yes; or 3) I tried, but nobody answered. From the 134 interacting participants, 16 (11.9%) stated that they did not use the chat box, 103 (76.9%) participants indicated that they used the chat box and 15 (11.2%) that they tried to use it. Participants who deliberately did not use the chat box do not differ from participants who used or tried to use the chat box on the dependent variable or any of the measured control variables. Hence, we are not able to explain what drives their choice not to use the opportunity of interaction (via the chat box). Excluding these observations (in total 16 ‘conversations’ or 32 observations) from our analyses does not impact our results (qualitatively or inferentially).
13 The ethics committee of the school from which the students were selected provided permission to conduct the experiment. Since the experiment does not involve or require any specific accounting knowledge (Peecher & Solomon, Citation2001) and estimation tasks only require a limited level of skill (Bonner & Sprinkle, Citation2002), it is appropriate to use student participants (Libby et al., Citation2002). Moreover, in practice, any member of an organization can be asked to provide time estimates, irrespective of their experience level (Cardinaels & Labro, Citation2008).
14 As indicated before, interactive total-time estimation makes sense in two possible settings: (1) employees with the same job descriptions and (2) employees who simultaneously perform different tasks. When interactively estimating task durations, in the former group, task comparability is important while simultaneity is not necessary and in the latter, task comparability is not important but simultaneity is necessary. To design the experiment as cleanly as possible and to reduce noise as much as possible, we chose to implement the same tasks for all participants (i.e., comparability) and to have these participants performing these tasks simultaneously.
15 Since the correlation between measured total time ( = 12.5 min + 9.5 min – time spent on interruption tasks) and total-time estimation error is not significant (r = −0.07, p = 0.29), this design choice does not impact our results.
16 Moreover, none of the interactive participants asked questions related to the tasks other than the duration and number of executions in the chat boxes. This observation suggests that participants were convinced that they performed the same tasks.
17 This time interval is based on the second pretest. We observed how much time participants needed to discuss their time estimates via a chat box. We limited the chat time in the final experiment to make sure that participants used their time efficiently and would not start discussing other things than the experimental tasks and their time estimates. One minute before the chat box disappeared, participants got a warning signal. Participants were not obliged to wait for six minutes before leaving the chat box, they could go to the next stage earlier.
18 Results of these analyses indicate that percentage-time estimation is processed similarly to total-time estimation error and yields similar results while unit-time estimation is processed differently (as indeed suggested by Schuhmacher & Burkert, Citation2020). In particular, for unit-time estimation, only a main effect of interactive time estimation is present and we do not find support for our hypothesized ordinal interaction effect.
19 We excluded the two outliers from our analyses: one participant was in the uninterrupted, interactive time estimation condition, and one participant was in the uninterrupted, individual time estimation condition. When we include these observations in our analyses, the contrast test becomes non-significant (F(3, 257) = 1.66, p = 0.20). Since these outliers significantly impact our results, they are labeled as influential outliers.
20 Since our prediction allows for, but does not necessarily imply, a small simple effect of task interruption within interactive time estimation and/or a small difference in estimation error between individual and interactive time estimation within uninterrupted tasks, we ran three alternative sets of contrast coefficients for the predicted pattern ([individual + no interruptions, individual + interruptions, interactive + no interruptions, interactive + interruptions]). As a first alternative set, we ran the contrast [-2, 3, -2, 1] to take into account the increasing (simple) effect of task interruption on time estimation error for both individual and interactive time estimation, assuming that this effect is stronger for individual than for interactive time estimation. Next, we ran two alternative sets that take into account that we expect an increase in time estimation error only in the individual, interrupted condition. In particular, the set [-2, 6, -3, -1] takes into account that within the uninterrupted condition, time estimation error may be lower for interactive than for individual time estimation and it further assumes that the estimation error is smaller in each uninterrupted condition compared to each interrupted condition. The final set [-1, 4, -2, -1] takes into account that within the uninterrupted condition, time estimation error may be lower for interactive than for individual time estimation and it further assumes that time estimation error in the individual uninterrupted condition is equal to time estimation error in the interactive interrupted condition. Our results are robust for allowing these three small differences in the predicted pattern (all p ≤ 0.02).
21 To explain why interactive time estimation results in lower time estimation error, we performed some quantitative analyses on the conversations in the chat boxes. Although word count is not significantly correlated with time estimation error (all p ≥ 0.29), the number of estimates exchanged during the conversation is marginally significantly correlated with total-time estimation error (r = −0.14, p = 0.12; and significantly correlated with percentage-time estimation error; r = −0.23, p ≤ 0.01). Hence, this suggests that the number of estimates exchanged during the conversation may explain why interaction decreases total-time (and percentage-time) estimation errors.