Hi Em Olson

well, if 20 people answered the question incorrectly at t1 and only 10 people answered the question incorrectly at t2, you have your answer, given the same people were present at both time periods. The date is still paired, you just do not know who is who.

By the way: If you use a 6-digit code for each individual (e. g. first two letters of their parents' first names and their own date of birth), you will be able to avoid this specific problem.

Best

Marcel

I think if questions result in ordinal or scale variables you can't really test the effect of the intervention because you didn't set a within-subjects design.

Hello Em,

This is one of those instances in which hindsight renders research issues so much more clearly than foresight.

The best you can do is treat the two sets of responses as independent of each other (since there's no post hoc method for matching individual responses to yield the more preferred dependent/paired analysis).

A test for equality of proportions (e.g., Ho: p1 = p2) would do the trick for addressing whether or not a comparable proportion of persons answered the question correctly across the two occasions.

Here's an online calculator that will compute the probability (non-directional): https://www.socscistatistics.com/tests/ztest/default2.aspx

Good luck with your work.

As David Morse says, treating as independent is an option, though not preferable to running the study correctly. The problem is that an assumption of the independent test is that the data are independent, and your data are not. The power can go up or down. Consider the following (in R so anyone can use since it is free):

> vals for (i in 1:1000){

+ set.seed(i)

+ pre

> for (i in 1:1000){

+ set.seed(i)

+ pre