The experiment consisted of two parts, a single-task and a multi-task part, both of which involved working on the slider task and the counting zeros task. The selection of these two tasks was based on a pre-test experiment, which was conducted among 18 subjects. The pre-test experiment explored the within-subject correlations between five real-effort tasks. Five tasks, the circle task [41], the counting zeros task [42], the ball-catching task [43], the slider task [44], [45] and the number adding task [18] were presented to subjects in a random order. Based on a within-subject analysis, we found that the highest correlation was between the slider task and the counting zeros task both in terms of actual performance and perceptions of the difficulty level. Aiming for characteristically very similar tasks to control for heterogeneous preferences for and performances in each task, we selected these two tasks as the most suitable for the purposes of our main experiment. The experiment was submitted for ethical review and was granted favorable opinion by the University of Surrey Ethics Committee (UEC/2015/042/FBEL). Written consent was obtained from all participants and the data was analyzed anonymously.

In the single-task part of our experiment, subjects sequentially received instructions about the two tasks they had to complete and had 5 minutes to complete each task. The order of the presentation of the two tasks was randomized to avoid any potential order effects. They had to complete each task separately and the performance in each task served as a control for heterogeneous abilities. At the end of a task, subjects received feedback on the number of correctly positioned sliders and completed counting zeros tables, which eliminated any potential uncertainty in subjects’ knowledge of their absolute ability levels. They were paid a piece rate of £0.10 per correct completion both in the slider and counting zeroes tasks, and the total earning in the single-task part was the sum of piece-rate earnings from both tasks.

After the single-task part, subjects completed a mid-study questionnaire that elicited demographic information (gender, age, and nationality) and self-reported economic preferences (general risk attitudes, competitiveness, confidence and attitudes towards donating to charities and social institutions) on a 7-item Likert-scale. By eliciting a number of demographic and psychological constructs we aimed to remove the salience of gender that could activate gender-related stereotypes and eliminate any possible experimenter demand effects [46]. The main reason for using mid-study rather than end-of study questionnaire was to ask subjects to choose a charity organization that they regularly donate to or whose activities they support from a list of charities offered. They could also fill in the charity of their liking that was not on the list. The chosen charity was then used in the next part of the experiment, where we induced social-value incentives as donations to chosen charities. Additionally, we chose to elicit the four economic preferences as they were identified by the previous literature to predict competitiveness, social preferences and image-seeking behaviour of subjects in single-task settings. The self-reported measures of economic preferences were similar to the validated survey instruments developed by [47] who show strong correlations of the measures with many real life and laboratory decisions. At the end of the questionnaire, we asked subjects which task they enjoyed most such that they could choose either one, both or neither. This question was aimed at controlling for subjects’ personal preferences for the tasks.

After the completion of the mid-study questionnaire, the multi-task part of the experiment started. Subjects had to multi-task for 10 minutes between the slider and the counting zeros tasks presented to them on the same screen. This part of the experiment consisted of four between-subject treatments where all subjects in a session participated in the same treatment. In all four treatments, subjects had to position a minimum of 10 sliders to receive a fixed rate of £3 and additionally they could earn a piece rate of £0.10 per completed counting zeros table. If subjects positioned less than 10 sliders their earnings from the multi-task part would be equal to £0. The fixed rate of £3 for a minimum correctly completed 10 sliders served two purposes. First, it helped to control the experimenter demand effect [46] that could affect the effort allocation decision in the slider task: subjects might have felt obliged to position a certain number of sliders since the task was on their screens. Second, our design mimicked the real world multi-task work environment where one of the tasks has a minimum required work load to be completed, otherwise the worker could get fired. For example, academics are required to provide a certain number of teaching hours [48], surgeons have to see a certain number of patients in a week and manufacturing workers have to adhere to minimum quality standards.

Across the four treatments, we manipulated whether subjects received additional incentives for their effort in the slider task. The Baseline condition lacked any additional incentives other than those described above and served as a control condition. In the other three treatments, we manipulated the additional incentives applied to the fixed rate paying slider task whilst keeping the piece rate in the counting zeroes task constant. In the Charity treatment, the slider task was a social-value generating task where subjects could earn a donation of £0.10 for a charity of their choice per each slider they positioned correctly exceeding the minimum of 10 sliders. While completing the mid-study questionnaire, subjects were not aware that we would use the chosen charities in the later parts of the experiment. This treatment measured the effect of social-value incentives on effort allocation decisions between the tasks compared to the Baseline. The Charity treatment mimicked the environment of working in corporate social responsibility or corporate philanthropy projects, where certain job tasks and projects improve societal welfare [49], [50]. It also mimicked social service jobs such as academic multitasking where academics have to undertake multiple job tasks for the university such as teaching, publishing articles, gaining research grants and administration. There is some debate in the literature whether these roles are complementary or [51], and the nature of teaching-research nexus depends on the academic discipline (soft sciences versus hard-applied sciences) and the level of teaching (undergraduate versus postgraduate) as shown by [52–53]. In situations where the academic tasks are substitutable, the effort spent on teaching has a direct third party benefit (to students) compared to research activities that are considered to have more of a private value in terms of career progression and job mobility [4]. [5] further show that in gender differences in career progression in academia hence may persist.

We assessed the effect of social-image incentives on effort allocation decision across the tasks in the CharityImage treatment where we designed the treatment according to [54] definition of social-image being positional, desirable and non-tradable. Social image is achieved via favourable comparison to others in socially recognized situations, hence positional. Social image is also desirable, because good social image brings along some benefit. Finally, social image is non-tradable in the sense that it cannot be directly purchased. Instead, it must be gained individually and, therefore, it must be acquired through actions that are visible. Hence, in social value generating settings, people concerned with obtaining high social image will strive to appear more generous. In addition to donations earned for every slider exceeding the minimum of 10, subjects were told that the names of the top three highest donors in a session would be publicly announced and they would be awarded with a Thank You! certificate. The certificate was signed by the project grant holder and thanked the participants for their donation. The “Thank you!” certificate is included in Appendix A in S1 File. This design choice was inspired by a field experiment [34] and a lab experiment [35] which show significant effect of public recognition and awards on effort. We chose to use a certificate award rather than the visibility of donations as our manipulation of social-image incentives to reduce the “wallflower effect” (an aversion to deviate from average behavior) previously reported in experiments [38].

Finally, in the Prize treatment, we measured the effect of competitive financial incentives on effort allocation decisions between the two tasks. In this treatment, in addition to Baseline incentives subjects could earn a £5 prize if their performance in the fixed rate paying slider task was in the top three performances in their session. The amount of £5 was determined so that the prize is slightly less than the total donation made in a Charity session divided by three, thus making the treatments comparable. We ran one session of the Charity treatment before the first session of the Prize treatment to identify the average amount donated in a session and calibrate its financial prize equivalent.

By comparing Charity to Baseline, we can determine how social-value generation affects effort allocation; by comparing CharityImage to Charity, we can determine whether social-image affects effort allocation and finally by comparing Prize to CharityImage we can identify the effect of competitive private monetary rewards as opposed to competitive social-value generation effects on effort allocation. After the presentation of the multi-task instructions, subjects had to answer control questions to make sure that they have understood the incentives correctly. Subjects could start the multi-tasking part only after they have answered all control questions correctly. The full questionnaire and the experimental instructions are presented in Appendix A in S1 File. Table 1 summarises the incentives used in the multi-task part.

It is worth noting that compared to previous literature that focuses on gender differences and incentives, we employ a continuous decision variable rather than a binary [22], [25]. This is similar to more recent studies that have looked at how subjects choose the percentage of their compensation to be based on piece-rate versus competitive incentive [55] or what subjects’ willingness to accept competitive incentive is [56]. Similarly our experiment uses continuous effort allocation decision between two differently incentivized tasks and extends the incentives to social-value and social-image incentives as well as competitively incentivized tasks.

At the beginning of the experiment subjects knew that only one of the (single-task or multi-task) parts would be randomly selected to be paid out. By using the random incentive system to pay subjects, we controlled for income effects and other potential interdependencies between the single-task and multi-task parts. The random incentive system is a widely used experimental procedure. For a discussion of its rationale and possible limitations see [57]. At the end of the multi-task part, subjects received feedback about their performance in both parts, the amount of donations they contributed to their chosen charity (in the Charity and ChairtyImage treatments) and whether they were in the top three of the session in their slider task performance (in the CharityImage and Prize treatments). In the CharityImage treatment, we also publicly announced the participation numbers of the top three donors, made them stand up and awarded them with a “Thank You!” certificate. All subjects were then privately paid in their cubicles.

We recruited subjects through advertising our study via weekly student-union newsletter, posters and handouts distributed around the campus. In total 210 subjects participated in the experiment with 54% female and 26% from social science departments. There were 42 subjects in the Baseline (4 sessions), 51 in the Charity (4 sessions), 65 in the CharityImage (5 sessions) and 52 in the Prize (5 sessions) treatments. The number of subjects in a session varied between 10 and 15 subjects depending on the number of fifteen registered subjects to show up to the experiment. While subjects knew that the experiment was for 15 participants when they signed up, they were not aware of how many exactly showed up as we seated them in their cubicles as soon as they arrived. This ensures that perceptions of probability of winning in Prize and CharityImage treatments were not affected by the actual number of subjects in a session. The experiments were conducted at a large university in the United Kingdom and were programmed using the software Ztree [58]. We ensured that the timing of sessions for each treatment was counterbalanced to account for time and day-of-the-week effects. Subjects received £5 for participating in one-hour session and additionally earned an average of £6.2.

In the single-task part of the experiment, the average number (standard deviation) of completed sliders was 42.8 (12.3) with a minimum of 13 and maximum of 76 sliders. The average number (s.d.) of completed counting zeros tables were 28.2 (6.2) with a minimum of 9 and maximum of 48 counting zeros tables. In Table 2, we test whether any of the observable characteristics of subjects significantly predict the performance in the tasks. We observe a significantly lower performance in the Slider task and higher performance in the Counting Zeros task by females. The result is consistent with the previous finding [44], [45] who also find lower performance of women in the Slider task; we are however not aware of previous evidence showing a higher performance of females in the counting zeros task. We also find a lower performance of older subjects in the slider task but not in the counting zeros task. These results have some potential implications on future experimental studies that use slider and counting zeros tasks in gender and age varied subject pools, especially if either of these variables are the main point of interest. We thus use subjects’ single-task performance as an individual ability control in the parametric analysis of multi-task part, which eliminates any potential issues studying the gender interaction with the effectiveness of incentives.

We find a significant correlation in performance between the two tasks consistent with the pre-test results that identified the Slider and Counting Zeros tasks as the most similar tasks. Absent any strategic considerations, the positive correlation between the tasks supports that there was not an endogenous conflict between the tasks per se. This enables us to exogenously induce conflict into the multi-task part of the experiment by explicitly manipulating incentives and restricting work time. We find no significant correlation between favouring a task and performance in a task. The insignificance of the correlation between favouring a task and scoring high in a task is encouraging and potentially means that personal preferences play a negligible role in exerted effort levels under piece-rate incentives. Additionally, we check whether single-task performance measures were the same across the experimental sessions and treatment conditions. We cannot reject the hypothesis that effort in the single-task part was the same across the sessions and the treatments (all p-values>0.100). Overall, the results of the single-task part are reassuring and enable us to control for individual heterogeneity in performance and personal preferences independent of any other strategic considerations when analysing the effectiveness of incentives in the multi-task part.

In Fig 1 and Table 3, we present the distribution of effort allocated to the fixed rate paying slider task and piece-rate paying counting zeros task in the multi-task part of the experiment across the four treatments. On inspection, the graphs show that all three incentive conditions are effective in increasing effort levels in the slider task that they are applied to. A non-parametric test on the equality of distributions shows significant differences between Baseline and the other three treatments. In the Baseline condition, the average number of completed sliders is 18 (s.d. 8.31), slightly above the minimum level of required 10 sliders. Since we did not give our subjects any feedback on the number of correctly positioned sliders while they were performing the task, subjects took an extra care not to fall below the minimum threshold 10 sliders by completing on average 18 sliders. The average number of completed sliders is the highest in the Prize treatment with 51 (s.d. 40.23), followed by the CharityImage treatment with 37 (21.02) and the Charity treatment with 35 (s.d. 18.02) sliders. A Wilcoxon-ranksum test on the equality of distributions shows significant differences in average effort level in the slider task between the treatments and Baseline conditions (p-value<0.000). Comparing the distributions of effort levels in the slider task across the three treatment conditions, the results show significantly higher effort level in the Prize treatment compared to the Charity and CharityImage treatments (p-value = 0.012 and 0.011, respectively). Table A in Appendix B in S1 File presents the results of an OLS regression across the three different model specifications to test whether the differences between the treatments are robust to model specifications. We find robust effect of incentives on the effort allocation in the slider task to the inclusion of additional variables such as single-task part performance, demographics and reported economic preferences. Panel (b) of Fig 1 presents the equivalent analysis of the effort level allocated in the counting zeros task demonstrating a mirror image of the effort allocations in the slider task. The only difference in the multi-task effort levels in the two tasks is in the differences in variance between the two charity treatments. The variability in effort levels in the slider task is higher in the two charity treatments than in the Baseline whereas variability in effort levels in the counting zeros task is not significantly different between the two charity treatments and the Baseline condition. Additionally please see Figure A in Appendix B in S1 File the cumulative density function of the completed sliders in each treatment.

The visual evidence of the effectiveness of incentives on effort allocation decisions in the slider and counting zeros tasks for men and women is presented in Fig 2. In the Baseline condition, both women and men complete around 18 sliders. Men complete slightly higher number of counting zeros tables but this difference is not statistically significant: mean number of counting zeros tables completed is 54.5 for men and 49.5 for women (Wilcoxon-ranksum p-value = 0.355). The reported p-values are from a Wilcoxon-ranksum tests across the genders and treatment conditions, unless otherwise specified.

In the Prize treatment, we observe differences between men and women in the multi-task part: men and women on average complete 62.6 and 38.8 sliders (p-value = 0.0315) and 29.4 and 37.9 counting zeros tables (p-value = 0.1607), respectively. The parametric analysis of the data, reported in Table 5, provides further evidence of significant differences between men and women in their effort allocation decisions in the slider task. The coefficient of the interaction term FemalePrize is negative and significant under all three model specifications that control for individual single-task slider performance and other self-reported preferences of subjects. The result is consistent with the previous literature showing that when given a choice of incentives, men self-select into competitive work environments whereas women choose non-competitive payment schemes [18], [60]. Differently from these studies that focus on binary decision of men and women on whether to take up competitively incentivized task or not, our results provide evidence on how much effort to allocate between the two incentivized tasks, using a continuous dependent variable. This is similar to the recent papers [55], [56] who also use continuous measure of competitiveness and find that women have significantly lower willingness to accept competitive incentive schemes and only one-fifth of the most competitive quartile are women. Our subjects also had a choice of allocating positive amount of effort between competitively paid slider task or piece-rate paid counting zeros task and women choose to spend less effort on the slider task than men: 80% of women choose to complete less than 50 sliders while this figure is 40% for men.