Hello Iliya,

can you report all the correlations among all involved variables (optimally by attaching, not posting!) a correlation matrix?

Best regards,

--Holger

Hello Iliya,

can you report all the correlations among all involved variables (optimally by attaching, not posting!) a correlation matrix?

Best regards,

--Holger

My short answer would be yes (if I understood it right). In a multiple regression model you can have both negative (inverse) and positive statistical significant associations. A similar question has been answered extensively in a previous thread:

https://www.researchgate.net/post/Is-it-possible-to-have-a-negative-correlation-but-positive-regression-in-SPSS

Positive and negative correlation both are possible in the study (depends on a dataset). both have their own significance and interpretation.

Hi Holger,

Thank very much for your suggesstion.

Attached is the correlation matrix.

Thank you.

Thank Jesper Fabricius, I will look at it. Thank u Shanti Verma

Hi Holger Steinmetz

Hi Holger,

Thank very much for your suggesstion.

Attached is the correlation matrix.

Thank you.

Hi Iliya,

first, N=4? Seriously? Don't do that.

Second, I did not quite understand your problem, especially the " the correlation came back with a strong positive". Do you mean the final table with the predicted value? Of course it is strong and positive. It has huge explanatory power (taken the circumstances into account).

Beyond that, it is quite easy to understand the connections between regression coefficients and correlations by applying the path tracing rules, invented in the 1920s by path analysis inventor Sewell Wright. The rationale lies in the concept of a compound path--that is the succession of arrows. In a nutshell the correlation between to variables can be decomposed to all direct and compound paths between the two variables (with the exception that a compound path must not created by two arrows pointing to each other).

For a simple regression model with two predictors, x1 and x2 that have effects (paths) of b1 and b2 on Y,

cor(x1,y) = b1 + r12*b2

were r12 = cor(x1,x2)

Simple as hell. Now play around with different scenarios of b1 being positive but r12 is negative. If both are equal in size but of opposite direction, cor(x1,y) will be zero (despite an nonzero b1), if b1 is positive but small and r12 is negative but huge, cor(x1,y) may even be negative and so on and so on.

That is the reason why correlation does not imply causation and while correlation coefficients are actually quite uninformative...

BTW: The above simple example can be generalized to any complex regression or structural equation modeling.

HTH

Holger

Holger Steinmetz brilliantly to the point, how far ahead was Sewell wright? And @illya , I would suggest you divide your dependent variable by 1000 or 10000,; software struggles to cope with very large absolute numbers, rounding issues often cause problems as estimation involves products and squares of numbers. SPSS is better than most in this regard, as its architecture is old, when these things really mattered. And I presume n of 4 is a tryout; if not forget it. I have read your question again and this is meant to be helpful, you need to talk to someone about your understanding of multiple regression, it is not about correlation.

First, I agree with Holger Steinmetz that trying to do regression with N = 4 is inappropriate. Further, two of your variables are correlated at above.99!

This strikes as a case where description is more appropriate than statistics.