Check Inter-Rater Reliability¶
A second kind of error is an error in judgment. One coder thinks that the child touched an object but a second coder does not think that the child touched the object. One coder interprets the child’s facial expression as distress but the other coder sees it as neutral. If coders frequently cannot agree about the codes for the same section of video, your coding scheme lacks inter-rater reliability. Inter-rater reliability will be low if there is a problem with the coding criteria (criteria are ill defined; criteria do not map well onto the behaviors) or if the coders are not well trained, or both. So, before you commit yourself to a coding scheme, test inter-rater reliability. If it is too low, you may need to revise your coding scheme or retrain your coders. Disagreements among coders are inevitable, even those who are practiced and familiar with the coding scheme. The question is whether the inter-rater reliability is sufficiently high to warrant confidence in the coded data.
How To Test Reliability Formally¶
What level of agreement is sufficient to consider a code to be reliable. The literature has no gold standard, but labs typically have their own gold standards. Generally, for categorical codes, you should use Kappas (which control for the base rate of the behaviors) rather than percent agreement. If you rely solely on percent agreement, then low frequency events will not be counted fairly. For example, if you are coding child affect as positive/negative and children rarely express negative affect (say, only 2-3 times per 100 trials), one coder could score positive for every trial without even looking at the video and you will have 97% agreement. The Kappa statistic takes low frequency events into account. For continuously scored behaviors, you can use the Kappa statistic to check inter-rater reliability frame by frame. For isolated events, you can use a Pearson correlation coefficient. You can estimate Kappas in Datavyu using scripts and the R interface, or you can estimate Kappas using statistical software after you export your data. Regardless, disagreements among coders are serious and must be reported in the write-up of the research.
Rules of Thumb¶
By definition, careless errors will lower inter-rater reliability. Therefore it is important to check for and eliminate careless errors before you test for inter-rater reliability.
Coders will experience “drift,” meaning that their coding will change slightly as they become increasingly experienced at looking at the particular behaviors. Therefore it is important to check inter-rater reliability at every point in the study—on initial sessions, in the middle of the study, and on the final sessions.
How much video should the reliability coder view to ensure inter-rater reliability. A good rule of thumb is 25%. But because every child is different, your reliability coder should score 25% of each child’s data, rather than 25% of the children. Which 25%. You can check inter-rater reliability at random intervals or regular intervals—whatever is most appropriate for sampling over the dataset. In some cases, particular trials or segments of video are especially important. In these cases, the reliability coder can score a larger percentage of the data—up to 100%.
Spread your best eyes over the entire dataset. For a large amount of data where several people will split the job of coding a particular pass, have your most experienced and knowledgeable coder score reliability so that your “best pair of eyes” is looking at representative data over the entire dataset.
This video displays one way to check for inter-rater reliability for a single column in a spreadsheet.