2

I'm a bit puzzled how to analyse self-report questionnaires (for example psychometric studies).

For example, consider a standard Hamilton Rating Scale for Depression which comprises 21 items with numerical scores [0-4] for each item.

For example, I have a cohort of 100 subjects, who replied to this questionnaire before and after some intervention.

Question: How to assess if there is a significant difference before and after the intervention?

Conceptually, one may apply paired t-test to both groups (before and after, as they are related) and then compute p values etc.

But the problem is that the outcomes (the overall score of such questionnaires) are not ratio data, but rather ordinal.

I know, that many researchers ignore that fact and apply standard methods designed for ratio data (real numbers). I am wondering, to which extent it is justified and is safe for statistical inference?

Arnold Klein
  • 618
  • 8
  • 18
  • 1
    If your items are Likert items, you might find this thread useful: http://stats.stackexchange.com/questions/10/under-what-conditions-should-likert-scales-be-used-as-ordinal-or-interval-data – T.E.G. Mar 06 '17 at 16:44
  • wil you present data and statistical results ? A definition of "cohort" as perceived for your study. –  Mar 07 '17 at 04:04
  • @subhashc.davar yes I will. I collected some data on people and would like to make statistical inferences. As far as I understand, for such sample size people would not bother too much and will treat numbers as ratio data, with standard stats approaches (t-test, anova, non-ordinal regression and computation of summary stats as mean and std. what do you think? – Arnold Klein Mar 07 '17 at 07:22
  • @Arnold Klein May I have a look at your questionnaire to understand ratio data and ordinal data. –  Mar 07 '17 at 07:34
  • How do you define rstio data ? –  Mar 08 '17 at 16:18
  • 2
    Voting to close because "standard methods of statistical analysis" is tremendously over-broad. If you have a *specific* form of analysis in mind, you might find your question editable to a more acceptable form. – Alexis Mar 08 '17 at 17:58
  • @subhashc.davar, sorry I didn't have time to upload data. I will do it in due course and hope you can help. – Arnold Klein Mar 08 '17 at 18:32
  • @Alexis, you are right, my bad. By statistical analysis I mean to compute parametric statistical tests (to check difference between different distribution of patients) on ordinal data. – Arnold Klein Mar 08 '17 at 18:34
  • Arnold, your edit is a big improvement. – Alexis Mar 08 '17 at 19:53

1 Answers1

2

There are no statistical assumptions regarding measurement scale. The question is one of interpretation and what you can say theoretically if you conclude there is a true difference in means. It is theoretically possible for mean differences to be very misleading, especially if the scale is extremely non-interval and the distributions are unusually shaped. As a practical matter, most differences between means can be taken at face value. If your browser lets you run unsigned Java applets you can explore the practical consequences with this simulation. http://onlinestatbook.com/2/introduction/measurement_demo.html

David Lane
  • 1,194
  • 1
  • 8
  • 9