0
$\begingroup$

I'm trying to analyze the behavioral data of my research experiment and I'm a bit lost... o_O

Briefly, subjects undergo 4 different conditions (of increasing complexity 1<2<3<4).

At the end of each conditions, they perform a little test.

These are the results :

  Subject Cond 1 Cond 2 Cond 3 Cond 4
     01     96    100     88     80
     02    100    100     92     72
     03    100    100     80     72
     04    100    100     92     76
     05     96    100     80     68
     06    100     96     92     76
     07    100    100     96     92
     08    100     96    100     72
     09    100     96     92     84
     10    100     88     88     72
     11    100     92    100     84
     12    100     96     92     76
     13    100    100     88     80
     14    100     96     88     80
     15    100     96     88     76

I want use a statistical test to affirm when the difficulty increases, scores decrease.

I think I should use a non-parametrical ANOVA (Friedman).

But, the results of condition 1 and 2 caps at 95-100.

Do I need to perform any correction before applying the Friedman ANOVA? Someone told me about an ArcSine correction but I don't know if it's valid or not...

$\endgroup$
  • $\begingroup$ It seems the test at Cond 1 is not providing useful information about differences among subjects. I don't see how a transformation will cure this. Massive ties may interfere with the power of the Friedman test. // Is it useful just to compare Conditions 2 and 3 with each other? Tests for those conditions seem to have useful information. Wilcoxon signed-rank test gives P-val 0.001 with CI $(10,16)$ for shift. // Also Cond 3 has median signif below 100 (P-val 0.002). $\endgroup$ – BruceET Sep 27 '18 at 17:27
0
$\begingroup$

The Friedman's is a non-parametric rank-sum test. In other words, it ranks the results of each subject across all evaluated conditions.

Therefore, any correction applied to your data will not change the ranks on the Friedman's test and, thus, your result will be the same.

I recommend you to simply apply the Friedman's test and verify if there are significant differences among your conditions.

$\endgroup$
  • $\begingroup$ Thank you for your answer. "verify if there are significant differences among your conditions" With Wilcoxon signed-rank tests ? Second question, if I have two tables to compare (a table for kids and a table for adults), can I just do a wilcoxon too between each columns of each table ? $\endgroup$ – FinnMcCool Sep 28 '18 at 9:13
  • $\begingroup$ The Wilcoxon's test compares two treatments (in your case, two conditions) between them. The Friedman's test is intended to compare three or more treatments (as is your case). The Friedman's test takes into account the differences among all of the evaluated treatments. Therefore, it is not equivalent on running many Wilcoxon's (one for each tuple of treatments). $\endgroup$ – Iago Carvalho Sep 29 '18 at 1:46
  • $\begingroup$ Which condition do you want to test by comparing the adults with the children table? $\endgroup$ – Iago Carvalho Sep 29 '18 at 1:48
  • $\begingroup$ I would like to test Condition 1 in adults vs Condition 1 in children (and same for each condition). $\endgroup$ – FinnMcCool Oct 4 '18 at 12:21
  • $\begingroup$ Since your data does not meet the requirements of a parametric test, you can use the Wilcoxon's test. You will compare the Condition 1 in adults (treatment 1) with the Condition 1 in children (treatment 2), and do so for all conditions. However, there is one important thing to remember: your Wilcoxon's tests should be unpaired. $\endgroup$ – Iago Carvalho Oct 4 '18 at 14:21

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.