Hello Drazenka,

The best choice will depend on several factors: (a) the reliability of each of the measures; (b) whether the relationship (if one exists) is linear or non-linear in form; and (c) sample size.

Reliability: Any time you form a difference score, the reliability of those differences will always be less than the reliability of either of the scores from which the difference is formed, unless both scores are perfectly reliable or the scores are uncorrelated (in which case, a difference would be pretty meaningless). My guess is that you could likely get a pretty reliable vertical jump measure by averaging the value of, say, six attempts on a given occasion (but it would still be subject to disturbances such as current injury status, health, time after eating, level of fatigue, etc.). I'm not sure how the aggression scores would fare, without more information about how you quantified them.

Relationship: If the relationship is linear in form, then ordinary Pearson or Spearman correlation would do a good job of quantifying the association (or incorporate into a regression model). If non-linear, you could either try polynomial regression, or categorize cases into subsets/groups (e.g., quartiles, quintiles) based on one variable and compare the groups on the other variable (via eta-squared in anova or R-squared in regression).

Sample size: the categorization might not work well with more than a few levels without an adequate sample size.

Good luck with your work.

I may be interpreting your question wrong. But are you trying to predict a scale variable from a binary categorical variable? If so you can dummy code you categorical variable and run regression.

I may be interpreting your question wrong. But are you trying to predict a scale variable from a binary categorical variable? If so you can dummy code you categorical variable and run regression.

Dear Mr. David Morse,

thank you for the answer.

I understand what you stated.

1) Indeed, the test-retest reliability of physical fitness measures (vertical jump and Yo-Yo tests) has been demonstrated to be good to excellent ( > .80). We used the Aggression Questionnaire developed by Buss and Perry, that has been shown to have good overall reliability (.82), but poor to good subscales reliability (from .53 to .82).

2) We evaluated the relationship between PF and aggression measures assessed at pre-test and post-test, which drew to the different conclusions depending on whether the relationship was evaluated at a pre-test or post-test. E.g., Pearson's correlation showed a significant relationship between the Yo-Yo test and physical aggression subscale when assessed at the pre-test, but non-significant when assessed at post-test. This is giving me a headache.

3) Finally, the sample consists of 105 subjects (n=54 and n=51 per group).

I appreciate your answer, thank you.

Dear Mr. Lee Curley,

Thank you for your answer.

We would like to predict one scale variable from another scale variable (since the Buss-Parry Aggression Questionnaire addresses each aggression subscale using a cumulative score of the 5-point scale items that represent one subscale).

So, if I understood,, we could use regression analysis, but which model will be adequate depends on which curve fits data the best (linear vs. non-linear).

Thank you.

In addition to my previous response to you, Mr. David Morse,

(3) I would like to clarify that the aforementioned groups didn't refer to the '' variable-based'' groups/levels, but the intervention group (n=54), and control group (n=51).

I appreciate your time, thank you.

Drazenka Macak Sorry for my confusion, yes if you have a scale criterion variable and a scale predictor variable a regression would be the best way to proceed. Is your data linear or non linear?

In relation to the groups above (intervention vs. control; this is where I was confused), you could add these categorical variables within the model, if you dummy code. Hope this helps.

Also thank you for clearing up my query.

Lee.