Hierarchical gets used in a few ways (a reason to avoid it). The other responders, thus far, have interpreted it as when you put in lots of variables into the right hand side of the equation. Here, conditional association is not the same as unconditional association, so sometimes effects go in opposite directions.

Hierarchical is also used for what are sometimes called mixed or multilevel models (hierarchical would be a subset of these). Here, researchers account for the data being clustered. Here is an example where the sign of the correlation differs from the regression coefficient. If a lot of tasks are given to people of very different abilities, then if you correlate all the times with the accuracy they can be positively correlated because the experts are fast and accurate and the novices the opposite. However, sometimes controlling for person effects in a hierarchical model means the effect turns negative because of the speed-accuracy tradeoff. This won't always happen in speed/accuracy data. Other examples exist also. Kreft and de Leeuw discuss the relationship between income and education either accounting for or not accounting for employment sector. Within sectors more education is positively associated with income, but many sectors that require a lot of education (e.g., teaching) are not highly paid.

The reason for the reversal in sign in these two interpretation of hierarchical come down to the same reason: a bivariate association can change when you account for other variables.

Correlation analysis look at relationship between two variables at one time. In heirarchical regression when a model looks at many variables at one time the result is influenced by relationship between the multiple variables.

This phenomenon was called "the lack of coincidence" by Polish statistisian Z. Hellwig. The main reason is too many variables in the model and multicollinearity.

Yes, multicolinearity - the strong correlation amongst the predictor variables can cause all kinds of whacky results in a regression equation. My rule of thumb for what to do when I have two predictors which are highly correlated is to choose the one which is most easily interpretable and enter that into the equation as a stand-in for the common variance of the 2 variables. Others probably have different strategies.

Hope this is helpful.

This appears to be a case of suppression and it is one of the reasons why multiple regression exists. If the correlations and beta weights always matched then you'd never need to run a regression.

From my book "Serious Stats":

"How does suppression come about? It can happen in slightly different ways, but is fairly easy to explain in a regression with only two predictors (X1 and X2), where X2 and Y are uncorrelated (i.e., rY2 = 0). X2 can act as a suppressor variable if it explains error variance in the regression of Y on X1. You can think of this as improving the reliability of X1, making it a better predictor of Y. Another way to look at it is in terms of the residuals of the simple linear regression model predicting Y from X1."

and from Darlington 1968 (p. 163):

“A typical example ...might be a situation in which a test of reading speed is used in conjunction with a speeded history achievement test to predict some external criterion of knowledge of history. Since the history test is contaminated by reading speed, assigning a negative weight to the reading-speed test would help to correct for the disadvantage suffered by a student with low reading speed who is competing with faster readers.”

Hierarchical gets used in a few ways (a reason to avoid it). The other responders, thus far, have interpreted it as when you put in lots of variables into the right hand side of the equation. Here, conditional association is not the same as unconditional association, so sometimes effects go in opposite directions.

Hierarchical is also used for what are sometimes called mixed or multilevel models (hierarchical would be a subset of these). Here, researchers account for the data being clustered. Here is an example where the sign of the correlation differs from the regression coefficient. If a lot of tasks are given to people of very different abilities, then if you correlate all the times with the accuracy they can be positively correlated because the experts are fast and accurate and the novices the opposite. However, sometimes controlling for person effects in a hierarchical model means the effect turns negative because of the speed-accuracy tradeoff. This won't always happen in speed/accuracy data. Other examples exist also. Kreft and de Leeuw discuss the relationship between income and education either accounting for or not accounting for employment sector. Within sectors more education is positively associated with income, but many sectors that require a lot of education (e.g., teaching) are not highly paid.

The reason for the reversal in sign in these two interpretation of hierarchical come down to the same reason: a bivariate association can change when you account for other variables.

Biba, your library probably has a copy or two of Mosteller & Tukey (1977).  I suggest that you borrow it and take a look at Chapter 13 (Woes of Regression Coefficients). 

HTH.

http://www.pearsonhighered.com/product?ISBN=020104854X

Hi,

Your question: Why are Beta weights in hierarchical regression negative while correlation analysis between the same variables is significantly positive?

If you want to find out correlation means that is the relationship between two variables at a particular time.

If you want to do hierarchical regression analysis, in that more than two variables (three or more) are involved.

So, you have to get negative in regression and positive significant in correlation.

You should consider the p-value only not consider the positive or negative relationship.

For significant you have to consider only the p-value, that you will get it from SPSS or SAS or any other statistical software.

All the best.