Saturday, November 22, 2008

Centering in SPSS

i cannot seem to find directions for the keystrokes necessary to center a variable in SPSS. How does one translate the equation into something that SPSS recognizes- i've hit a wall. I've done it manually but am not sure whether or not the results are precise enough.

6 comments:

Danny said...

I'll try to answer the question because it's been up for a while...

Here goes (maybe someone can confirm this information when you have time):

When you are centering a variable you are creating a new variable which is equal to every data point for your variable minus the mean of that variable. In doing so, you are creating a new column of values in your data set.

The first thing you need to do is determine the mean for your variable. To do this follow these SPSS commands:

Analyze > Descriptive Statistics > Descriptives > pull over the variable you are looking at - let's say MPG for example > click Options > check Mean > Continue > OK

Your output will reveal a Descriptive Statistics box that will include the mean value for your variable.

Jot that number down and go back to your data set view.

To center a variable you follow these SPSS commands:

Transform > Compute Variable > enter a label/name for the variable you want to center in the "target label" field - so for example, if you are centering MPG you might call your new centered variable centmpg > in the "numeric expression" field you enter the equation for your variable minus the mean - so for MPG, you pull over the MPG variable and enter the minus sign and the value for the mean you jotted > click OK.

You will find your new centered variable, in this example labeled centmpg, in the last column of your data set.

Remember: as mentioned in previous blog questions (including one of my mine!) when centering, center all continuous predictors and moderators, and never center the outcome.

Kris said...

Danny is correct. Centering a variable consists essentially of recomputing the new variable of whatever name you have chosen (although, I would recommend a new name rather than starting each variable with "cent" as this can slow down your process when you maybe can only see the first few letters of the variable name in some SPSS screens. Maybe MPG.c or something would keep the original variable name visible and still designate the difference) minus the mean.

Just to elucidate this one more time:
As Danny mentioned, this is what the top of the dialogue box for the "Compute Variable" should say:

(target variable name of your choice) = (variable to recompute) - (mean value)

MPG.c = MPG - 10.2 (or whatever your mean is from the descriptive statistics)

Good description Danny.

Just remember to keep all your variables straight (by this I mean clear). Maybe consider having a separate file with the names listed and the changes/transformations you have made so you can be sure that you are using the correct variables for the analyses that you want (certainly more helpful as your analyses become more complicated w/ more variables and more sophisticated procedures).

Unknown said...

Centering variables in regression analysis does not fix multicollinearity. How could it? Do the math. Regression is based upon correlations which in turn are based upon the product of standard scores (hence "product-moment" correlations). Standard scores are already centered.

There are numerous articles written about the uselessness of centering variables in moderated regression. Look them up!

Vishal Lala said...

Richard, mean centering is used for specific multicollinearity problems. You are right is saying mean centering is not your multicollinearity panacea but when faced with multicollinearity from interaction variables (X1X2) or power functions (X1^2), the ingredient variables and the newly formed variable tend to be correlated. Of course, it is no surprise for X and X^2 to be correlated or X1 and X1X2 to be similarly correlated. It is in these situations that mean centering saves the day.

Unknown said...

Thanks so much for the explanation on centering variables - will enable me to accurately conduct my moderated regression - you saved the day.

Ben Suavio said...

Still helpful in 2013