Research on the Progress Bar Distortions
In today’s software, the usage of progress bars (of any type) is very high and users come across them everywhere, from their personal computers to public places, like ATMs. Some of the processes that are being visualized by the progress bars consist of multiple subprocesses, of which some might have a variable speed that may cause decelerations or stops of the progress bar.
This work aims to broaden the research made by Harrison et al. [1] about progress bar distortions that can reduce an overall percieved time of progress bar loading. This has been done by designing and creating a website with questionnaire, and then conducting two separate experiments (laboratory and public). Based on collected data, the work describes mutual relationships between different distorting functions, and rates their strength to reduce the perceived time. In addition, a new phenomenon of “second shown progress bar” has been discovered.
This article is a shortened version of the original work, available here.

1 Concept

Progress bars are graphical visualizations showing the progress of currently executed processes so that the user can easily imagine the amount of work already done and how much time should he or she expect that it will take to finish the rest. Progress bars are widely used in entertainment as well as in professional software. Therefore, users can come into contact with this element everywhere from computer games, mobile applications and ATMs to software they use for their work. The problem with progress bars is, that often unexpected delays of processes occur, which results in pausing or decelerations on progress bars. According to Harrison et al. [1], these effects tend to make progress bar be perceived as longer then if the progress bar was loading for the same time period but with a linear character of loading. Due to this fact, Harrison suggested that distortions of the projected data might improve the overall impression of the process length. Moreover, he claims that using a Peak-and-End effect [2] can even make progress bar look faster than it is. The aim of this work is to verify some of the questionable results of the original work by Harrison et al. and also to broaden the research by investigating unexamined distorting functions and their mutual relations when they are used on progress bars. Another focus of this work will be on the behavior of users facing the distortions . To increase the relevance of results, the study will consist of two experiments performed in two different environments. Collected data will then be also used to research how an environment, specialization, and age can affect the final results. By studying these functions and their properties, it will be possible to say if some functions can suppress a negative behavior of progress bars without knowing what the behavior will be like (e.g., if it will stop, slow down or load linearly).
The original work has a few imperfections that I do not consider as crucial, but they might become problematic while broadening the research. Therefore, it is important to correct those imperfections and based on these changes develop the conceptual model of the research itself.

1.1 Problems with “Rethinking the Progress Bar” by Harrison et al.

The first one is that Harrison et al. used an unrepresentative sample of respondents of 22 people (14 male, 8 female), all from two computer research laboratories. One of the most important UX principles is that the research should be done on a broad amount of people with a different personal background. The group of respondents having only IT specialists is very narrow to claim that the conclusion the team of Chris Harrison is generally applicable. Apart from that, the IT specialists are a very specific focus group for any kind of UX testing. Since they use different kinds of user interfaces on daily a basis and do know lots of shortcuts and facilitations, they are used to elements that an average user is often not to and are able to cope with difficult situations more easily. Therefore, the chosen sample might have affected the results, especially on such a small amount of respondents.
Another questionable factor is that users the research has been realized in their own offices. The author has not mentioned whether the environmental conditions during the testing were controlled or not (i.e., were there other colleagues during the testing?). If they were, on the one hand, the moderator provided guided testing so that users could ask anytime they did not completely understand the assignment or had any problem. Also assuming from the paper that there were no distractions during the testing, the users were not disturbed by any other factors from the surrounding world. On the other hand, the ecological validity could be improved by allowing the users face everyday distortions (i.e., listening to music, people around talking) .
During the testing, the progress bars loading took 5.5 seconds each. From short empirical research, I conducted, I assume that such length is not sufficient enough to show the characteristics of different distortions. That is connected to another problem that on certain distortions, especially on the wavy ones, such a short time period has not behaved naturally and users would not get in touch with such behavior of progress bars in real life very often.
According to the pilot testing (see Chapter 3) of this research, some of the users started to lose their attention even before they finished their 20th 8-second-length comparison. Therefore, I assume letting users compare 45 different pairs of progress bars could affect the relevance of the results.
Linear function
Figure 1: Linear function
Power function
Figure 2: Power function
Wave after Power function
Figure 3: "Wave after Power" function
Wave after Pause function
Figure 4: "Wave after Pause" function
Pause after Power function
Figure 5: "Pause after Power" function

1.2 Broadening the Research and Conceptual Changes

Although I assume the original work has imperfections, I do not suppose that these imperfections could have an essential effect on the statistical preference.
Supposing that the original conclusions are correct, to see what functions are dominant over the others, and to be able to add them to an ordered system created by Harrison et al., new functions were created. The dominance of function means that the function can suppress the effect of other function that makes the progress bar to be perceived as slower or faster than it actually is. I have selected one representative from each of the hierarchy groups [1, p. 117, Figure 3] and composed its functions together.
See the final set of functions in Figure 1 — Figure 5. The composition order of functions was selected in the way to make their characteristics more visible.
In contrary to the original work, the domain of rating values has been enlarged to be able to distinguish the rate how much does one function seem faster to the user than another. The domain has been modified from original values -1, 0, 1 to new -5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5.
To deal with the problem of an unrepresentative sample of respondents and the environmental conditions, the changes were made in the following way. There were used two samples of users in two different environmental conditions.
The first group has been answering the questions in controlled conditions of the school laboratory. The answering was conducted in groups from 4 to 12 respondents, depending on the date. Every respondent has been spaced out so that he or she did not sit next to another respondent (to not be affected by others).
The second group was formed by public experiment and was intended to have steadier specialization and age distribution and to be conducted at home on the users’ computers. These factors should increase the ecological validity of the whole research by allowing me to compare two different experiments both in controlled and uncontrolled environmental conditions. The public link of a website with the experiment has been spread to people and shared on social media.
To keep the user’s attention and their interest in finishing the test (mostly for the public testing), I edited the number of questions the users had to answer. According to the pilot testing (see Chapter 3), the users in the post-testing interview were sceptical about the amount of questions asked and told me, that the end questions were really the threshold to start losing attention and if they were not asked to complete the test personally, they would have probably not finished the test. To prevent this behaviour during the public testing, the questions were reduced to 10 per respondent in public testing and kept 20 (even though in original research was 45) for laboratory testing.
To prevent the results from being affected by the order in which the comparison pairs were shown, it was needed to shuffle their order. Unlike in the original work, the concept of assigning the questions has changed to assigning the users different patterns of comparison pairs that were shown. Every pattern consists of comparing every distortion to all others with the exception of comparing a distortion to itself. Therefore, for 5 different mathematical functions used in this research, every set (pattern) had 10 questions that, for laboratory testing, were also duplicated in reverse comparison order. That was done to also keep track of the information if an order within the comparison pair affects user’s perception as Harrison et al. have claimed. That means laboratory patterns had 20 questions, where respondents compared both distortion A to distortion B and distortion B to distortion A. To get answers of the same relevance from both types of testing, the public testing (having 10 questions only) had not only 4 patterns (as laboratory one) but 8 different patterns so that some users got to compare distortion A to B and some compared distortion B to A. The order of comparisons within the patterns has been pseudorandomized.
Changes were also made to the graphical user interface of the application. The testing is web-based to allow users to complete the questionnaire from their home. At the very beginning, users go through introduction, instructions and personal info form followed by a preview comparison. In laboratory testing, a user also has a chance to ask the guide if any information is not clear. During each comparison, the first progress bar shows, once it finishes it disappears and the second one shows. After completion of the second bar, it disappears too and shows the comparison panel. Users then choose on a scale from 5 to 0 on both sides (11 buttons in total) which functions and how much did they perceive them to be faster. This feature lets users express their feelings more accurately than just by preferring one or another option. The maximum value of 5 was considered as large enough to sufficiently express the degree of affection by any of the sides based on the amounts of options in most used questionnaires in UX studies [3].

1.3 Expectations

The main focus is to find out which functions are dominant over the others, meaning that they can suppress the perception effect of the other function. Especially whether Peak-and-End of a function can suppress the trends of other functions that were perceived as slower than linear function. An expectation is that the function “Early Pause after Power” will dominate over all other functions.
Another focus is to find out what are the partial comparisons between functions and if there may exist a cycle in the mutual comparisons. The expectation is that the functions created by composing two other functions together, where one of them is “Power”, will be perceived as faster than a linear function, with exception to the “Wave after Early Pause” that will be perceived as slowest of all the others.
The last hypothesis is that there will exist a statistical difference between answers of people of similar demography (men vs. women, technicians vs. others, millennials [4] vs. older users).

2 Design

After defining the concept of the research, I started to design the interface using UX methods to create an easily usable interface without bugs and ambiguous elements.
Sketches
During the design process, I focused on creating a friendly environment that has been clear and easily learnable to anybody. The reason for focusing on learnability is that the application was going to be presented online without the possibility of asking the moderator of the experiment for help. Therefore, the instructions had to be clear and sufficiently detailed, but on the other hand, the interface had to be very clean so that the users focused only on the task that was given at the time (reading, watching the progress bar, etc.).
For more information about the design process of GUI (sketches, lo-fi and hi-fi prototypes), see the full-text work.
Wireframe

3 Testing

The whole testing had two parts. The first part was pilot testing that was carried out on a small sample of users with the aim of revealing the weaknesses of the experiments. It has been conducted on four users with a laboratory version of the questionnaire. The users had no crucial troubles with understanding the instructions during the research. During the post-research interview, they had some comments and suggestions (see full-text version for more information).
After finishing the pilot testing, the bugs and other weaknesses were fixed, and both parts of the main tasting started.
The main testing consisted of two separate experiments—public testing conducted on anonymous users of Internet and laboratory testing conducted on respondents invited to the environment of a laboratory. The findings found in pilot testing were implemented into the final design of the testing application. Source code of the website is available in University Archive.
Screenshot of application

3.1 Laboratory Testing

The data collection for laboratory testing was conducted in a school laboratory within 10 groups counting from 1 to 12 respondents per group. Every respondent was using his own computer and had at least one seat on each side free so that they were not affected by others. In the laboratory were no significant distractions during the testing.
At the beginning of the research, the users were informed that during the testing they are going to be comparing progress bars. They were asked to turn off their mobile phones, to not communicate with others and once they finish their questionnaire to stay calmly in their seat to not disturb others. The questionnaire database was then checked by the moderator online to know when all the respondents had ended.

3.2 Public Testing

The data collection for public testing took 14 days. A public link was sent to relatives and friends, and they were asked to share the link with their circle of friends. Apart from sharing a link within the friends, the link was also posted to 5 well-known Facebook groups with a request to help with a bachelor thesis and short introduction of the thesis itself.

4 Results

4.1 Laboratory Results

To analyze the users’ preference between two functions (both when the order of the compared function matters and when it does not) the same method as Harrison et al. has been selected with modification that it is used on a scale from -5 to 5 (instead of -1, 0, 1). The ratings of each function pair a user has given are summed up and later divided by the number of the values which produces a mean value the users have given. An example might be function 1 versus function 2. Function 1 is evaluated by these values: -3, 2, 4, 0, 3, -1 (therefore function 2 gets 3, -2, -4, 0, -3, and 1), the sum is then divided by 6 and the mean value would be 0.83—a positive value, therefore the function 1 is preferred. This method takes into account that numerous low negative values in combination with a few large positive values may produce mean of 0.
To investigate the general preferences of participants, I tested suggested differences in preferences with the use of inferential statistics to support my expectation about the general superiority of the suggested function “Early Pause after Power”. The data were normally distributed, so I used parametric tests for the deeper analyses.
Photo of testing

4.2 General Evaluation of Laboratory Testing

I used one-way ANOVA (analysis of variance) to indicate the suggested effect in reported preferences of participants. In the controlled laboratory conditions 46 people were tested, of which 33 were male and 13 female. They were aged from 19 to 26 (m = 21.93, med = 22, sd = 1.531), both from technical (34) and other (12) fields.
ANOVA indicated a significant effect between functions (F = 28.478, df = 4, p < 0.001).
The analysis of the values calculated with respect to the order of functions within the comparison couple has shown that users tend to prefer the first function they have seen. Users have selected the first function in 48% of all answers, 31% the second one and 21% did not prefer any. Especially in comparisons of “Linear” with “Early Pause after Wave”, and “Wave after Power” with “Power” the trend of dominance of one of the functions is very well visible.
The overall values with no respect to the order have proven the hypothesis that the “Early Pause after Power” function is the most dominant of them all. The second strongest (fastest perceived) one is “Power” function, then “Wave after Power”, “Linear” and the last is “Wave after Early Pause”.
Graph of average values in aboratory testing
Figure 6: Mean Values in Laboratory Testing
Values with a significant difference from comparisons without respect to the mutual order have been then used to form a topological order of the functions. The final order of functions is shown in Figure 7—the leftmost function was perceived as the slowest, solid line represents a significantly different relation, dashed represents relation that was not significantly different.
Topological order of functions in aboratory testing
Figure 7: Topological Order of Functions in Laboratory Testing

4.3 Public Results

Similarly to the laboratory testing the mean method (see Chapter 3.1) has been chosen. Public testing has been conducted on anonymous respondents using an online questionnaire.
To investigate the general preferences of participants, the inferential statistics have been used again to support the expectations. The results were normally distributed again; therefore parametric tests were used for the deeper analysis.

4.4 General Evaluation of Public Testing

Similarly to the laboratory condition test, I used one-way ANOVA to indicate suggested effect regarding reported preferences of general public users. During 14 days of testing in the public environment, 192 people conducted the whole questionnaire (answers of the people, who did not complete the whole questionnaire, were eliminated). They were aged from 16 to 68 (m = 27.545, med = 23, sd = 10.598), both from technical (84) and other (108) fields.
ANOVA indicated a significant effect between functions (F = 53.827, df = 4, p < 0.001), similar as in the laboratory testing.
The phenomenon of users selecting the first function more often has occurred during public experiment again. However, the percentual difference is smaller than in the laboratory experiment (the first function has been selected in 41% of cases, second one 35% and in 24% was selected zero).
There was no effect of visible dominance of the first seen function calculated via mean values during the public experiment. In the contrary, there has been a new behaviour of preferring the second function during comparison of “Linear” function versus “Wave after Power” function.
While not calculating with respect to the order of functions within the comparison couple, the hypothesis of “Early Pause after Power” function being the fastest has been proven again. Other functions had the same order within each other too, the only difference was their mutual mean values and significance.
Graph of average values in Public testing
Figure 8: Mean Values in Public Testing
Similarly to the laboratory experiment, values with a significant difference from comparisons without respect to the mutual order have been then used to form a topological order of the functions. The final order of functions is shown in Figure 9—the leftmost function was perceived as the slowest, solid line represents a significantly different relation, dashed represents relation that was not significantly different.
Topological order of Functions in Public Testing
Figure 9: Topological Order of Functions in Public Testing

5 Discussion

Thanks to comparison of new functions with two of the original ones, it is now possible to place the new functions on a number line created by Harrison et al. [1, p. 117, Figure 4]. Even though it is not possible to classify the new functions with an exact number, the topological order of functions gives us an idea of how they would fit into the original order (Figure 10). Original [1] (green points), edited (pink points representing the investigated functions). Bottom branches represent indefinite mutual order within the functions on the same interval of the axis.
Order of Original Functions and New Ones
Figure 10: Order of Original Functions and New Ones
The analysis comparing laboratory testing with the public one has not uncovered any significant differences between results collected from users in these two different environments. The only difference between these two experiments was visible during the analysis of relative values with respect to the mutual order, specifically the comparison of “Linear” and “Wave after Power” functions. In the laboratory experiment, the testing in both orders has resulted in favour of “Wave after Power” function.

5.1 Second Shown Progress Bar Phenomenon

Compared to that, in public testing, the results were indecisive. Both functions were once preferred over the other, but surprisingly, only if they were shown as second compared function. This phenomenon has not been spotted or described yet neither by Harrison et al. and would be interesting to study it deeper. Considering the resulting values of a laboratory experiment were both very close to zero, I do not assume that the phenomenon of the second function shown was a statistical exception.

5.2 Effects of Distortions

The final order of tested functions says that the “Power” function in combination with any other has an effect of making the progress bars be perceived as faster than if the data were projected linearly. Therefore, it is possible to say that the “Power” function has a dominance no matter what is the second function that it is in combination with. Whether it is a function with characteristics of making the progress bar look slower or faster. In both situations, the Peak-and-End effect dominates and improves (makes faster) the final overall impression.
This fact enlarges the field of use of distortions in practical use. On one hand, as Harrison et al. have already stated, the progress bars with progress corresponding to the approximated state are sufficient to visualize the ongoing process. On the other hand, thanks to the discovered facts, it is possible to use a distortion on any progress bar without knowing its future behaviour (no matter if it will stop for a while, periodically slow down or have some other combination of behaviours). In the commercial sphere, this feature could be well-suited to enhance user experience.

Conclusion

The goal of this work was to broaden and confirm the relevance of existing research on distortions used to manipulate users’ perception of progress bars made by Harrison et al. [1].
Experiments comparing both new and original functions showed that the Peak-and-End phenomenon has a large positive effect on the perception of a user. Moreover, it can inhibit negative effects like periodical slowing down or pausing. Also, a hypothesis that the “Power” function is better perceived then the “Linear” one has been confirmed. All these results are applicable in the software development to improve the user experience, especially by shortening the perceived progress bar duration during their unexpected behaviours.
The work has also explored a new phenomenon of “second shown progress bar” that, in some cases, contradicts the principles that resulted from the original work by Harrison et al.
The collected data also served as a source for various exploratory analyses . These analyses were carried out on the basis of demographical differences (gender, specialization, age) and showed differences between genders, and also between millennials and older respondents—females and older users tend to use a larger domain of answers (for more information see full-text work).

Future Work

Findings of this work opened various topics to be discussed in the future. The discovery of the “second shown progress bar” phenomenon established a new field that could be investigated. Usage of different functions with similar trends as “Linear” function and “Wave after Power” function during the comparisons might reveal new findings, that would help to describe it. Also, some new behaviours might occur if the strength of a “Power” function is greater (for example as the “Fast Power” function in the original work [1]).
Usage of two different environmental conditions showed that both experiments have similar results. Therefore, for future work, it would be sufficient to research only in one of these environments.
Also, it would be interesting to investigate longer progress bar durations that might show new phenomena which would enlarge the usage of distortions for improving user experience again.
Because females behaved differently during both experiments (see the original work), another research could be focused on this topic.
A big contribution for future research might also be a possibility of usage of the web application implemented for this work (available in University Archive). It is easily adaptable to many kinds of other experiments focused on progress bar distortions.

References

[1] HARRISON, Chris; AMENTO, Brian; KUZNETSOV, Stacey; BELL,Robert. Rethinking the Progress Bar. In: Proceedings of the 20th An-nual ACM Symposium on User Interface Software and Technology. New-port, Rhode Island, USA: ACM, 2007, pp. 115–118. UIST ’07. ISBN 978- 1-59593-679-0. Available from DOI: 10.1145/1294211.1294231.17
[2] LANGER, Thomas; SARIN, Rakesh; WEBER, Martin. The retrospective evaluation of payment sequences: duration neglect and peak-and-end effects. Journal of Economic Behavior Organization. 2005, vol. 58,no. 1, pp. 157–175. ISSN 0167-2681. Available from DOI: https://doi.org/10.1016/j. jebo.2004 .01.001.17
[3] HARTSON, Rex; PYLA, Pardha. The UX Book: Process and Guidelines for Ensuring a Quality User Experience. 1st. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 2012. ISBN 0123852412, 9780123852410.
[4] “Millenials”. https://en.wikipedia.org/wiki/Millennials. Retrieved 2018-12-9.
Full-text work available here.