I am comparing results of three experiments,in which 3 classes were taught same course with slight modifications (course A to class X , B to Y , C to Z) , a questionnaire will be used to assess factors like (interest , motivation and course completion ) in each experiment . Each question has (Y/N ) options only. As I am new to statistics I am not getting the idea from where to start . which technique it requires ?
3
-
$\begingroup$ Do you have some clear hypothesis you want to address or are you fishing for any difference on any question between any of the groups? $\endgroup$– BjörnCommented Nov 9, 2017 at 6:03
-
$\begingroup$ Yes , i have hypothesis that adding some elements of games enhance interest of students , students of class X were taught simple course without games and class Z was taught game based course . Now the questions (i.e were you interested in the course(Y/N) determine factors needed $\endgroup$– shamrozCommented Nov 9, 2017 at 6:21
-
$\begingroup$ And there were just 2 (or 3) classes not multiple classes assigned (by randomization?) to each approach? If this were a cluster randomized trial with clusters being the classes, the analysis would be straightforward. $\endgroup$– BjörnCommented Nov 9, 2017 at 6:39
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
If there are just 3 classes, each with a different intervention, then relying on randomization (even if you did it will be problematic). It seems to me that one sensible approach would be to estimate propensity scores for whether subjects will be in one class or another. Those propensity score would be based on any relevant factors that might potentially influence that and later outcomes (e.g. previous test scores, socioeconomic status of parents etc.). Because there's 3 classes/interventions, it's not quite as simple as a single score, but you may end up with 2 of them (e.g. for being in class X rather than Z and in class Y rather than Z) - if you go this way, you probably want to read up on propensity scores for multiple interventions.
Then if your main outcome of interest is "Were you interested in the course? (Y/N)", you could use logistic regression for that stratified by e.g. the deciles of the propensity scores (perhaps challenging when there are few subjects in each class and you need to deal with two scores).
An experiment that would allow for a much easier analysis would be one, where you assign a large number of classes randomly to each of the interventions. In that case you could simply use e.g. logistic regression with a random effect for class (on the intercept), but if you just have 3 classes and each has a different intervention, it becomes hard to disentangle the effects of being in a particular class from that of the particular intervention.
