I have a very specific question about the application of Cronbach's alpha to data with trichotomous responses (true-false-don't know). I have an acquaintance I'm helping replicate a study using a 45 item knowledge questionnaire. The original article claims that the questionnaire was internally consistent with an alpha of .92. However, if I run reliability analysis on the original responses (coded 0 for false, 1 for don't know and 2 for true) I get an abysmally low alpha of .35. However, if I recode the responses to indicate whether the responder actually got the question right or not (0 for wrong answer, 1 for right answer), and run reliability on these variables, I get an alpha of .87. Which of these scales should I be running the analysis on to begin with? Unfortunately, the original article is extremely poorly written, and does not indicate which course of action the authors took.
Thank you for any advice and responses!