Regression Analysis - Different Scores?

Moderators: statman, Analyst Techy, andris, Fierce, GerineL, Smash

simG
Posts: 7
Joined: Wed Jun 08, 2011 6:07 pm

Regression Analysis - Different Scores?

Postby simG » Thu May 10, 2012 10:28 am

Hi!
I ran a regression analysis on my data set and asked for the predicted values in standardized scores for my data.
I also computed these scores by myself (using the regression beta values equation).
The two computations didn't give the same results!! How can that be?

another question - how big of an sample size do I need in order to generalize from my sample and use the regression model on another sample?

Thanks,
Simon
Penguin_Knight
Posts: 473
Joined: Thu Apr 05, 2012 5:58 pm

Re: Regression Analysis - Different Scores?

Postby Penguin_Knight » Thu May 10, 2012 12:16 pm

simG wrote: I ran a regression analysis on my data set and asked for the predicted values in standardized scores for my data.
I also computed these scores by myself (using the regression beta values equation).
The two computations didn't give the same results!! How can that be?
Try use SPSS to get the unstandardized original predicted value.
simG wrote: another question - how big of an sample size do I need in order to generalize from my sample and use the regression model on another sample?
It's a big topic and I am not sure if I know all the criteria. At the very least, I believe you'll need:
1) Internally valid, that your model is unbiased. (aka, accurate and have no problem with unmeasured variable nor confounding)
2) Sample size and power calculations were performed to ensure your analysis can detect the suspected association.
3) Sample is randomly drawn.
4) Population you wish to generalize to is reasonably similar to your sample.

And there could be a lot more, here is a starting point for you to read up:
http://en.wikipedia.org/wiki/External_validity
simG
Posts: 7
Joined: Wed Jun 08, 2011 6:07 pm

Re: Regression Analysis - Different Scores?

Postby simG » Thu May 10, 2012 12:54 pm

Thanks for the reply.
That what I did and the results aren't the same.
as for criteria for generalization, I know there are a lot, I'm asking about sample size specifically.
Simon
Penguin_Knight
Posts: 473
Joined: Thu Apr 05, 2012 5:58 pm

Re: Regression Analysis - Different Scores?

Postby Penguin_Knight » Thu May 10, 2012 2:12 pm

simG wrote: That what I did and the results aren't the same.
What degree of "not the same"? Like totally wrong or off by a bit in the decimal places? Maybe some example would help. Also, if you are familiar with syntax, could you lay out your regression syntax and the calculation syntax? And if not too much trouble, can you also show the regression output?

Another possible problem is maybe you have used the "Standardized beta" in the regression coefficient output, while the actual ones we should use are the coefficient listed under the first column called "B".
simG wrote: as for criteria for generalization, I know there are a lot, I'm asking about sample size specifically.
Then that'd depend on your power, suggested effect size, and variation among your sample and population. If your slope is steep and the variation is small, then a small sample can give unbiased estimates. If your slope is shallow and the variation is big, then a bigger sample size is needed. This is not a question that can be answered by just the sample size alone.
simG
Posts: 7
Joined: Wed Jun 08, 2011 6:07 pm

Re: Regression Analysis - Different Scores?

Postby simG » Fri May 11, 2012 3:06 pm

Thanks for the thorough reply.
What degree of "not the same"? Like totally wrong or off by a bit in the decimal places? Maybe some example would help. Also, if you are familiar with syntax, could you lay out your regression syntax and the calculation syntax? And if not too much trouble, can you also show the regression output?
In some cases, its way off course. unfortunately, the results and syntax are at my office adn not accessible now.
Another possible problem is maybe you have used the "Standardized beta" in the regression coefficient output, while the actual ones we should use are the coefficient listed under the first column called "B".
I thought of that, but that's not the problem. to be exact - the problem occurs only with the standardized scores.
Then that'd depend on your power, suggested effect size, and variation among your sample and population. If your slope is steep and the variation is small, then a small sample can give unbiased estimates. If your slope is shallow and the variation is big, then a bigger sample size is needed. This is not a question that can be answered by just the sample size alone.
Is there somewhere I can learn about this subject?

Simon

Who is online

Users browsing this forum: No registered users and 1 guest

cron