A first pairwise comparison experiment with video material in D-PAC is successfully completed. The goal of this experiment was twofolded: (1) test the tool on the scalability using videos; (2) and test the inter-rater reliability.
A group of 134 students in Education Sciences had to judge 9 clips on the quality of the simulated scientific semi-structured interview demonstrated. The pairwise comparisons were all scheduled synchronously. So, in total 134 assessors were simultaneously interacting with the D-PAC system which was sending out video clips to these assessors. During the experiment no technological issues arose, leading to a very positive conclusion on the scalability of the D-PAC tool.
In order to test the inter-rater reliability the group of assessors was split in three random groups consisting out of 46, 44 and 44 assessors. All of the groups assessed the video’s in a comparative manner. The only difference between the groups was in terms of providing feedback when every comparison was completed. Group 1 was not specifically instructed to give any argumentation or feedback during the process. The second group was asked to give a short overall argumentation for their choice after each comparison. Group 3 was asked to write down some positive and negative features of each interview after each comparison. The amount of comparisons each group made was 520, 354 and 351 comparisons, respectively.
Based on the pairwise comparison data we calculated the Scale Separation Reliability for each of the three groups of assessors separately. The results are given in Table 1. From this table it can be seen that the reliabilities are high (.91 – .93).
Table 1. Scale separation reliability and average number of comparisons per video
|Scale Separation Reliability||Average number comparisons per video|
To provide an answer on the question of inter-rater reliability we calculated the correlations between the estimated abilities (based on the Bradley-Terry –Luce Model) of each of the three assessments (see Table 2). The Spearman rank correlations between the two assessments in which assessors had to provide an argumentation (Group 2) and where assessors had to provide feedback (Group 3) is the highest (.87). The Spearman rank correlations between the scores resulting from the assessment without any argumentation (Group 1) and the two other conditions are somewhat smaller (.82 and .84). Overall these correlations are high.
Table 2. Spearman Rank Correlations between scores coming from the 3 groups of assessors
|Group 1||Group 2||Group 3|
GiveGiven that each of the 36 possible pairs were assessed by multiple assessors within and between the three groups, we were able to calculate the agreement between assessors for each possible pair. In Figure 1 the agreement is plotted per pair, split up for the three groups of assessors. As shown, the average agreement in each group overall is around 77%. For some pairs the agreement is only 50%, for other pairs the agreement is 100%. These differences can, of course be partially attributed to the fact that in some of the pairs are more difficult to judge than some other pairs. Comparing the results of the three groups showed no significant differences.
To conclude, this pairwise comparison experiment first of all demonstrates the robustness of the tool to deal with large numbers of assessors assessing video clips simultaneously. From the resulting scales and pairwise comparison data learned us that the inter-rater reliability seems to be rather high as well.