# Correlation vs. Means Plot

#### caligula4stats

##### New Member
Hey guys,

for my master thesis I conducted an experiment and received for two variables means plots, that I will attache below. They were the result of two one-way ANOVAs. I wanted to test if the scores in the two variables differs significantly in three conditions.

The next step would be a mediation analysis. Beforehand, I conducted a correlation analysis. To my surprise (and adding up to my desperation) the variables seem to be negatively correlated. However, the means plots paint a completely different picture, don't they?

So before I go into mediation and regression analysis, how to tackle these findings? What to believe? Does that make sense?

I double checked, that the scales I computed are properly (one had to be reverse-coded). Furthermore, I controlled for and excluded outliers, the scales are both normally distributed, Pearson was used.

I have been fighting with my data for the past week and think I reached the limits of my abilities in SPSS, please help a poor fool out!

Best, Nils

#### Attachments

• 20.1 KB Views: 4
• 21 KB Views: 4
• 47.9 KB Views: 3

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Didn't quite follow what you are looking at, but you would be better served looking at scatterplots. Also how are the variables formatted?

#### caligula4stats

##### New Member
Didn't quite follow what you are looking at, but you would be better served looking at scatterplots. Also how are the variables formatted?
Hey hlsmith,

thanks for your reply! The scales were composed over the means of 14 (likability) and 5 (ant_dev) items, the latter was reverse-coded. Is that what you mean by format?

X = manipulation (three conditions)
Y = IV (likability)
M = Mediator (ant_dev)

What puzzles me is the following: I conducted two ANOVAs, with M as dependent and Y as dependent. Now I would look at the means plots and think "oh high M scores are correlated to high Y scores, what a surprise!".
Then I go into further analysis and see that the correlation between M and Y is negative (correlation analysis), that M is a strong negative predictor for Y (linear regression) and that there is a significant indirect mediation effect of X on Y through M.

I attached the scatterplot below. Maybe you can explain your approach using the scatterplot?

Best, Nils

#### Attachments

• 27 KB Views: 1

#### caligula4stats

##### New Member
Hey hlsmith,

thanks for your reply! The scales were composed over the means of 14 (likability) and 5 (ant_dev) items, the latter was reverse-coded. Is that what you mean by format?

X = manipulation (three conditions)
Y = IV (likability)
M = Mediator (ant_dev)

What puzzles me is the following: I conducted two ANOVAs, with M as dependent and Y as dependent. Now I would look at the means plots and think "oh high M scores are correlated to high Y scores, what a surprise!".
Then I go into further analysis and see that the correlation between M and Y is negative (correlation analysis), that M is a strong negative predictor for Y (linear regression) and that there is a significant indirect mediation effect of X on Y through M.

I attached the scatterplot below. Maybe you can explain your approach using the scatterplot?

Best, Nils
Just so you can all get a bigger picture of what was conducted, I will now include all remaining "puzzle pieces" in a bundle.

#### Attachments

• 27 KB Views: 1
• 89.8 KB Views: 1
• 70.4 KB Views: 1
• 66 KB Views: 1

#### hlsmith

##### Less is more. Stay pure. Stay poor.
Can you color code scatterplot based on 3 groups? What do your residuals look like and the dependent variable is a categorical/interval variable?

I would say it gets a little weird treating what I think are pooled Likert style variables as continuous variables. I usually get a little flummoxed in these more social science analyses. So you are saying cross-sectional collected instrument's construct is mediated by another construct and the outcome may be categorical. I guess, I'm just saying watch out for your modeling assumptions and possible model misspecification. Though, these types of data structures are not my forte, so keep that in mind.