Log data and response selection behavior were examined. A total of 216,625 log records from a cohort of 1,241 students were analyzed, including 225 participants who took multiple exams. Data were derived from one exam from each of five different departments. Examinations from the years 2021 and 2022 were used, with an average participation of 248 individuals per exam (range: 234–282). Care was taken in the selection process to ensure that the summative exams were from various stages of the curriculum, the allocated time per item was varied, and the exams exhibited good characteristic values (Difficulty, Cronbach’s α). The five exams were invigilated electronic state examinations conducted in-person using the Q-Examiner^® software (IQuL GmbH, Bergisch Gladbach, Germany); additional details about the exams are provided in Table 1.

The item formats utilized in this study comprised of multiple-choice question (MCQ) in single-choice format, Kprim, Key Feature, picture diagnosis, and picture mapping, which are described in more detail in Table 2.

Based on the log data, the response time per item of each examinee was determined, and subsequently, the mean response time for each individual item was calculated. For the respective subjects and examinees, the median time spent on each item and the submission time of the exam was recorded. To standardize and facilitate comparability, the submission time was calculated as a percentage of the maximum available time for the examinations.

The overall difficulty of the examinations was calculated as the mean value of all examinees’ scores and is presented as a percentage of the maximum attainable score.

For each of the 346 individual items, the parameters considered were the item format, response time, character count as well as the two psychometric characteristics: discrimination index and difficulty. The character count was defined as the sum of all letters and numbers, including special characters, punctuation, and spaces from the item stem and the answer options. These parameters were then checked for correlation. Correlations between the response time and the parameters difficulty, discrimination index and character count were examined separately for every item format in order to exclude the influence of the variances between the formats. Only the statistical data for the two formats MCQ and Kprim are presented, as a sufficiently high number of items for a meaningful statistical evaluation of the other formats was not achieved. As a next step, the ratio between the length of the question stem and the answer options was examined. For this purpose, items of the two item formats MCQ (n = 231) and Kprim (n = 81) were considered. The relative proportion of the character count of the question stem to the total character count of the item was calculated for each, and its relationship with the parameters difficulty and discrimination index was assessed.

Regarding the response selection behavior, data on changes made to the selected answer option were only available for items of the MCQ format. This meant that a total of 231 items from the five examinations were available for evaluation. For every item, the number of changes in answer selection, along with the corresponding switch between the originally selected option and the newly chosen answer were recorded for each examinee. Furthermore, in cases where the response selection was changed several times, the last change was identified, as it was the one that was ultimately graded. Due to variations in the exam conditions concerning item and participant numbers, the analyses were conducted based on relative proportions. For each exam, the following parameters were calculated:

To assess the quality of answer modifications in detail, the number of changes between distractors and attractor or between different distractors was examined. For clarity, distractors are referred to as incorrect answer options (incorrect) and the attractor as the correct answer option (correct).

The difficulty of an item is defined as the percentage of participants who answered the task correctly and can thus range from 0 to 100% (11). The recommended range for item difficulty is 40–94% (12). Discrimination index describes an item’s ability to differentiate between participants with high performance and those with low performance. Items with good discrimination are answered correctly by good candidates and incorrectly by poorer candidates (13). The values of the discrimination index can vary from −1 to +1, with values above 0.2 considered adequate (12).

This study was approved in advance by the Data Protection Officer at the TiHo. All utilized and collected data were processed and analyzed anonymously. Students had to agree a data protection declaration upon matriculation, which permits the use of data collected during examinations in anonymized form in accordance with the requirements of Art. 6(1)(e), 89 GDPR in conjunction with § 13 Lower Saxony Data Protection Law (Niedersächsisches Datenschutzgesetz, NDSG).

Access to the raw data was restricted to the authors of this paper only, and all data was stored and processed on secure servers within the institution. To protect students’ data privacy, all personal identifiers were removed from the data before analysis, including matriculation numbers and any other information that could potentially be used to identify individual students. Instead, a unique, anonymized identifier was assigned to each data point to maintain the integrity of the dataset. Descriptive and statistical analysis was performed using aggregated data to prevent the identification of students based on their response behavior during the exams.

The descriptive analysis was performed using the spreadsheet software Microsoft® Office Excel 2010 (Microsoft Corporation, Redmond, WA, USA), while advanced statistical analysis was conducted using SAS® Software, Version 9.4, and SAS® Enterprise Guide® 7.1 (SAS Institute Inc., Cary, NC, USA).

Concerning quantitative data, all normally distributed numerical values are presented as mean values, including the standard deviation (SD) where applicable. For non-normally distributed values, the median and interquartile range (IQR) are provided.

To examine the correlations among the quantitative parameters, these were initially tested for normal distribution using the Kolmogorov–Smirnov test. Subsequently, a Spearman’s rank correlation analysis was performed on non-normally distributed data. For correlations between qualitative with quantitative parameters, the Kruskal-Wallis test was applied, followed by the Dwass-Steel-Critchlow-Fligner pairwise comparison method. A significance level of 5% was used, indicating that a p-value <0.05 implied that the influence of the parameters was significant.

Log data was initially analyzed by examining the time of submission of exams by the participants in their respective subjects. The absolute number of submissions per submission time, along with the available time in minutes, the number of items, and the overall difficulty of each exam are presented in Figure 1. Across all five of the analyzed exams, half of the candidates completed the exams within 64% (IQR: 11%, range: 46–77%) of the maximum available time. Three out of four students had submitted their exams within 77% (IRQ: 17%, range: 58–90%) of the exam time. Additionally, 90% of the participants finished the exams within 92% (IQR: 17%, range: 70–98%) of the exam time.

For each exam, the average allocated time per item was calculated for further analysis. Subsequently, correlations among the parameters “available time per item,” “submission time,” and “overall difficulty” were examined. A significant (p < 0.0001) negative correlation (r_s = −0.4312) was observed between the available time per item and the submission time. Hence, participants did not utilize the additional available time per item in exams that allocated more time per item.

Furthermore, a significant (p < 0.0001) positive correlation (r_s = 0.1106) was found between the available time per item and overall exam difficulty. This implies that the more time students had to answer the items, the higher the frequency of correct answers. There existed a significant (p < 0.0001) negative correlation (r_s = −0.2738) between the submission time and overall exam difficulty. Accordingly, students completed easier exams earlier.

The median item response time including range for each of the five academic subjects is shown in Table 1.

Regarding the five item formats MCQ, Kprim, Key Feature, picture mapping, and picture diagnosis, the median response time was evaluated. For items of the MCQ format (n = 231), 43 s (IQR: 26 s) were spent, for Kprim (n = 81) 74 s (IQR: 26 s), for Key Feature per subquestion (n = 21) 44 s (IQR: 30 s), for picture mapping (n = 8) 87 s (IQR: 84 s), and for picture diagnosis (n = 5) 77 s (IQR: 35 s). The effect of item format on item response time was significant (p < 0.0001). A more detailed distribution of the required time per item and format is depicted in Figure 2. Pairwise comparisons show significant differences between the formats MCQ and Kprim (p < 0.0001) as well as between Key Feature and Kprim (p < 0.0001). There are no significant differences between any of the other formats. Further analysis reveals significant differences between item formats referring to their difficulty (p = 0.0003) and discrimination index (p = 0.0243). Pairwise comparisons show significant variations in difficulty for MCQ and Kprim (p = 0.0009) as well as for Kprim and Key Feature (p = 0.0026). For variations in discrimination index the test was significant for MCQ and Key Feature (p = 0.0072) as well as Kprim and Key Feature (p = 0.0228).

In addition to the influence of the item format on item response time an assessment of the impact of the character count of an item was included. Median character count of the item formats was 246 characters (IQR: 200 characters) for MCQ items (n = 231), 268 characters (IQR: 222 characters) for Kprim (n = 81), 446 characters (IQR: 243 characters) per subquestion for Key Feature (n = 21), 253 characters (IQR: 113 characters) for picture mapping (n = 8), and 292 characters (IQR: 109 characters) for picture diagnosis (n = 5). Variations in character count of the various item formats were significantly different (p = 0.0003) but only for the pairwise comparison of MCQ and Key Feature (p < 0.0001). In Figure 3 the formats MCQ (n = 231) and Kprim (n = 81) were included separately to display the time spent on each item based on individual character count. A significant positive correlation was found between the two parameters item response time and character count for both MCQ (p < 0.0001, r_s = 0.3809) and Kprim (p < 0.0001, r_s = 0.5986), which indicated that more time is needed to complete items the more characters they contain.

Figure 4 illustrates both the difficulty of individual items separated by item format and their respective average response time, revealing a significant and negative relationship between the parameters “difficulty” and “item response time” for MCQ (p < 0.0001, r_s = −0.6607) and Kprim (p < 0.0001, r_s = −0.5038). Accordingly, answering more difficult questions required more time.

However, no significant correlation (MCQ: p = 0.2101; Kprim: p = 0.7390) was found for the relationship between character count and difficulty.

Figure 5 depicts the discrimination index of individual items separated by item format and their respective average response time. A significant negative correlation was identified between the parameters “discrimination index” and “item response time” for MCQ (p < 0.0001, r_s = −0.3779) and Kprim (p = 0.0099, r_s = −0.2851). This indicates that questions on which students spent more time generally showed poorer discrimination, suggesting that the time students took to respond to an item might be linked to its ability to discriminate effectively between high- and lower-performing students.

Furthermore, Figure 6 depicts the discrimination index of individual MCQ items in relation to their respective character count, where a significant (p = 0.0236) negative correlation (r_s = −0.1496) was found for both parameters. Consequently, the discrimination index tended to decrease slightly for MCQ items with a higher character count. However, for Kprim items, no correlation (p = 0.5534) between the discrimination index and character count could be verified.

Regarding the evaluation of the effect of the length of the question stem, no significant correlation was found between the parameters “relative proportion of the question stem” and “difficulty” for MCQ (n = 231, p = 0.5627) and Kprim (n = 81, p = 0.619) nor for “relative proportion of the questions stem” and “discrimination index” for both MCQ (n = 231, p = 0.9718) and Kprim (n = 81, p = 0.4449).

Table 3 shows the calculated parameters concerning the relative proportions of changes made to the chosen answer option for each of the five subjects.

The relative proportions of changes between incorrect and correct options are presented for each individual exam in Table 4. The mean values for the changes were 42.89% (SD ± 3.81%) for incorrect to correct ones, 27.34% (SD ± 6.71%) for incorrect to incorrect ones, and 29.77% (SD ± 4.01%) for correct to incorrect ones.

Within this study, log data of electronic examinations at the TiHo as well as the response selection behavior of the examination participants were evaluated for the first time in order to examine whether recommendations for the design of online examinations can be derived from them.

When examining the submission times of participants for different exams, a wide range of exam completion times becomes evident. Within a time frame of 58–90% of the maximum time limit, 75% of the participants managed to complete their exams. The time range for 90% of the students was between 70 and 98% of the available time. This highlights that the time required for exam completion cannot be reliably standardized. However, trends can be derived from the statistical analyses conducted. Notable was the observation that participants tended to finish exams earlier in relation to the maximum available time when there was more time allocated to answer individual items and when the overall exam was easier.

In exams with the longest allocated time to complete items in subjects D (120 s per item) and E (90 s per item), it is apparent that students generally did not utilize this longer time frame. The majority of candidates managed to complete the exam before 70% of the exam time had elapsed. Thus, theoretically, there is room to reduce the time allocated per item from the 90 s per item recommended in the literature (12, 14) and used by the TiHo. However, this reduction needs to consider the correlation between the available exam time and the overall difficulty of the exams reported in this study, as students tend to answer items more accurately when they are granted more time per item, which is also described in Lovett’s publication (15). The positive impact of additional time per item on students’ performance was already demonstrated in previous studies (16–19). Here, the effects are most apparent for students with relatively weaker performance (16, 17, 19), or those experiencing test anxiety (20). This effect might be attributed to reduced exam stress (21), more time to consider items, ample time for item completion (22) as well as reduced test anxiety (20, 21), which is said to lead to difficulties in concentration and impaired information processing skills (20). In this context, a reduction in exam time should be carefully considered with regard to exam fairness. Particularly with regard to the mentioned exam fairness, it must be noted that increasing time pressure can put certain other groups of examinees at a disadvantage. For example, studies indicate that increasing time pressure has a more substantial negative impact on the performance of female examinees than male participants (23, 24). Furthermore, students with learning, cognitive or psychiatric disabilities such as ADHD (15, 25), Asperger’s syndrome (25) or dyslexia (9) are entitled to accommodations for disadvantages, which is often in the form of an extended time limit (15), since the negative impact of time pressure is especially evident for students with disabilities (26).

Examining the actual time taken by candidates to complete the exams on the level of individual items also indicates that candidates did not utilize the time limit of 90 s per item. The average response time ranged between 47 and 57 s per item for all veterinary subjects. This observation aligns with the findings of other analyses, consistently reporting a time of 60 s or less per multiple-choice item (27–30). In some cases, the required response time per item was approximately 40 s or less (28–30). Therefore, there is an opportunity of reducing the exam time limit to 60 s per item, a practice already standard in some other exams (28, 31, 32). On the one hand, this would allow more items to be included in the same time frame, thus improving the validity and reliability of exams (12, 33–37), and, on the other hand, a stricter time limit can be used as a tool to reduce interchange between candidates in unproctored online examinations (9, 38).

It is important to note that three additional aspects need to be considered when determining time limits for exams and individual items.

Firstly, exams in the medical field are intended to test students’ knowledge, understanding, and application of knowledge and should thus be conducted as power tests (12). When time constraints are introduced or exams are conducted under high time pressure, factors such as candidates’ stress resistance and cognitive performance affect overall performance (9, 12). Consequently, the difficulty of exams increases (39, 40), and a potential time shortage can result in unanswered questions or blind guessing, raising concerns about the validity of exam results (12, 22, 41). In light of the examination objective in veterinary medicine, so-called speed tests where the time factor significantly impacts candidates’ performance should be avoided.

Secondly, the influence of exam time limits on performance, primarily observed in weaker students, should be considered to preserve exam fairness. A shorter time frame can lead to poorer test results.

Lastly, the ratio of different item formats should be mentioned (42) since significant differences in response time for formats were shown. Especially for answering the three formats Kprim (74 s), picture diagnosis (77 s), and picture mapping (87 s), students needed on average more than 60 s of response time, while MCQ (43 s) and Key Feature (44 s per subquestion) required less than 60 s. Other studies also conclude that formats like picture diagnosis and picture mapping are more time-consuming for candidates than MCQ (43, 44). Therefore, when determining the time limits for exams, it is essential to consider the item formats used and their ratio (42). This ratio is particularly important if aiming for the aforementioned reduction of response time to 60 s per item. Since students require significantly longer than 60 s on average for Kprim and image-based formats, there is a risk that an exam with a high proportion of these formats becomes a speed test. Hence, when limiting it to 60 s per item, attention should be given to predominantly select MCQ items, as less than 60 s are sufficient for this format, thereby balancing the additional time required for Kprim and image-based formats.

Furthermore, such exams can be evaluated to determine whether time significantly impacts students’ performance. The Education Testing Service defines a multiple-choice exam as a power test if all participants answer at least 75% of all items, and 80% of participants respond to all of the items (45). Otherwise, it is considered a speed test if these criteria are not met. However, this rule of thumb was developed based on paper-based exams and the assumption that examinees do not complete the exam if they run out of time. In the context of electronic multiple-choice tests where an incorrect answer does not result in a negative grading, it is more likely that examinees randomly select answer options for the remaining items to still have a chance to randomly choose the correct option. This is referred to as rapid guessing behavior (22, 44). Modern analysis methods use the log data from electronic assessments to identify rapid guessing behavior. Schnipke (22) graphed the standardized natural logarithm of response time to detect examinees with accelerated response times and a lower frequency of correct answers, indicating that these students might be running out of time. In addition, other complex models and methods, including those based on Item Response Theory, have been developed to assess the time influence on students’ exam performance and behavior (41, 44).

Regarding the analysis of character counts, response time, and item parameters of the tasks and their correlations, it was demonstrated that students need more time for items with high difficulty, poor discrimination index, and a higher number of characters. Additionally, it was shown that MCQ items with a higher character count tend to have a poorer discrimination index, but that the character count has no effect on difficulty. These effects have been described in other studies that also concluded that examinees require more time for poorly discriminating items (30, 46) as well as for difficult items (46, 47). As a result, to improve discrimination and reduce required response time, items should be kept as clear and concise as possible, which aligns with the formal requirements of multiple-choice items in the literature (12, 28, 48–52). However, focusing on the brevity of items should not be at the expense of the learning content to be assessed, as it is outlined in Bloom’s taxonomy (53).

Bloom’s taxonomy categorizes cognitive skills into six levels: remembering, understanding, applying, analyzing, evaluating, and creating (53). Questions designed to test basic recall or understanding can indeed be kept short and precise, reducing both the time needed for students to respond to these items and the time required for question authors to create them. For example, a straightforward multiple-choice question requiring students to recall a specific fact or definition can be brief without sacrificing its effectiveness. In contrast, questions that aim to assess higher-order cognitive skills, such as applying knowledge to new situations, analyzing data, or evaluating concepts, often necessitate a more detailed question stem (50, 52). For instance, regarding the clinical sciences of veterinary medicine, presenting a comprehensive scenario that provides sufficient context to assess students’ critical thinking and problem-solving abilities is essential. Such scenarios might include clinical findings, laboratory results or detailed case vignettes (50). These elements are crucial for testing students’ abilities to integrate and apply their knowledge but inevitably lead to an increased character count of these items. Subsequently, a careful balance is necessary when creating questions that are able to assess these higher-level cognitive skills. While the objective is still to keep items as concise as possible, it is equally important to ensure that all necessary information is included to allow students to fully understand and respond to the question.

Consequently, an efficient question design is recommended to reduce the time needed for authors to create items and for students to answer them, which can make it feasible to reduce the time per item and thus include more items in the same timeframe as before, which positively impacts the quality criteria of exams (52).

Applying these findings to the design of online assessments, a time limit of 60 s can be set for this format as long as the exam is primarily composed of MCQs and Key Feature items, which are short and concise. The shorter allocated time per item can help reduce the risk of cheating. However, it is essential to ensure that students do not feel overly rushed. Some authors even discuss explicitly designing exams to create significant time pressure on participants as a means to reduce cheating attempts (9, 38, 54). This statement should be viewed critically based on the aspects discussed earlier. When it comes to students with disabilities (see section 4.1), reducing the time per item must be approached with caution, as this can create additional challenges in administering exams (15, 25). In compliance with local laws and regulations, appropriate accommodations must be provided, which typically involve extending the exam time to allow students to utilize these accommodations effectively (15). Decision-making processes should be established to determine whether, and to what extent, accommodations are offered to disadvantaged students, as well as how much extra time is actually necessary (15). Additionally, the technical and organizational challenges of granting extra time to only some students must be considered (25).

The analysis of the log data from the five exams indicates that for almost all of the items there was at least one student changing their answer and that the majority of students belonged to the group that revisited and modified the originally selected response option. However, it is noteworthy that each individual student changed their originally selected answers only for a very small proportion of items. These findings based on the data from veterinary students corroborate the results of other studies, which also conclude that nearly all students changed their original answers to a different one, yet each individual participant only makes corrections for a small fraction of all exam items (55).

One potential influencing factor on this low frequency of modified answers could be the widespread belief that the initial intuition when reading the question is the correct answer. Benjamin et al. (55) reported in their literature review that almost all students believe that revising their answer does not impact the exam result; 75% were even of the opinion that the result would deteriorate. The teaching staff shared a similar belief, with over half of the respondents stating that exam results would worsen due to answer modifications (55). In contrast, studies demonstrate that the majority of answer revisions are from incorrect to correct responses, resulting in improved exam results (55–60). Hence, this phenomenon is termed the “first instinct fallacy” (55, 57–59). The analysis of the response selection behavior in this study reaches the same conclusion, as the majority of all corrections were either from incorrect to correct options or between two different incorrect options, resulting in a positive or neutral impact on exam scores.

Accordingly, the available time for exams should not be stipulated within such a short timeframe that participants lose the opportunity to adjust their response selections due to time constraints. For online examinations, one study recommends that students should only be allowed to select an answer for an item once only (38). We do not follow this recommendation based on the results of our analysis, as not being able to change the originally selected answer could potentially worsen the exam results.

As an alternative, a so-called “block setting” can be used for electronic exams. In this approach, the exam is divided into equally sized groups of items, and the items within the blocks as well as the blocks themselves can be individually randomized for each examinee. Access is only granted to the items in the currently worked-on block, which can be accessed and modified multiple times. After completing the block, the following block is unlocked and the previous item block is locked. This means that students no longer have access to the previously worked-on items. This approach allows exams to be structured better, the students focusing their concentration on the items within the block. Randomizing the blocks and the included items can be used to reduce cheating (38).

A limitation of this study was the small sample size for the image-based item formats, namely picture diagnosis and picture mapping. The reason for that was that for the sake of comparability exams were chosen in which all five item formats were used and had already been established for several years. However, the image-based formats were less frequently employed in these exams. In future studies, a greater emphasis should be placed on a more in-depth evaluation of the two formats picture diagnosis and picture mapping.

Since the focus was primarily on analyzing log data and response selection behavior of electronic examinations, more research is needed regarding the evaluation of the psychometric quality of item formats in online examinations and thereby identify suitable formats for distance exams. As both lecturers and students in veterinary medicine call for more application- and competence-based item formats (61, 62), these would be of particular interest, for example key feature and the image-based formats referenced in this work.

Regarding the items included in this study and their item analysis parameters, it is noticeable that some items display a low discrimination index, and a few items exhibit a negative discrimination index (see Figure 5). This indicates that these items might not effectively differentiate between higher-performing and lower-performing students, according to Krebs (12) as well as McCowan and McCowan (13). Negative discrimination indices imply that students understood these items in the opposite way than originally intended, meaning good students answered incorrectly while lower-performing students chose the correct answer (12). Therefore, such a negative discrimination index is an indicator of potential flaws in item construction or a lack of alignment with the learning objectives. To address this issue, the content of items with negative discrimination was reviewed for correctness, relevance, formal errors, and cueing, including a distractor analysis. However, no significant issues were identified. The low discrimination indices can partly be attributed to some items being relatively difficult (p < 40%) and some items being relatively easy (p > 80%) (see Figure 4). The best discrimination indices are found in items with medium difficulty, and deviations upwards or downwards result in significantly poorer discrimination values (12). Regardless of the reason, it must be noted that such items directly impact the calculated correlations of this study.

Open-book exams are a format commonly used for electronic distance assessments. Over the course of the COVID-19 pandemic, they gained importance as online proctoring is not necessarily required for this format (4), which resolves data protection issues associated with remote examinations. Comparative analyses already showed that the psychometric parameters of closed-book in-person exams and online open-book exams do not significantly differ (54, 63–65). Hence, according to the mentioned literature, this innovative format can be considered suitable for summative assessments. However, it is crucial to take into account the current developments in the field of artificial intelligence (AI). Generative AIs, particularly ChatGPT, are capable of successfully passing challenging final and licensure exams, including several law bar exams (66), the United States Medical Licensing Exam (67), the German medical state exam (68), and other assessments (69, 70). In Progress Tests, ChatGPT answered over 60% of the items correctly (71). It can be assumed that this performance can also be reproduced in the field of veterinary medicine. Given that such AIs are readily available to students as well (72), the feasibility of distance assessments needs to be critically examined and re-evaluated (72–74).