any statisticians?

[bandXtrb]

okay, I have to make a recommendation about how college applicants should be considered at a particular school, yes, this is homework


I have these variables for 500 students at the college: High school GPA, Current College GPA, Number of credit hours completed, Gender, Age, Verbal SAT score, MAth SAT score.

I'm really stumped on this. I'm looking for correlations with the College GPA to see what the criteria for admission should be. I tried adding verbal + math sat scores together, and comparing them to college GPA. In a dotplot, it looks like it could possibly be a positive correlation, but its pretty scattered, just like HS GPA compared to college GPA.

Oh, I am working with software called DataDesk. I can do boxplots, dotplots, histograms etc. I can split variables (ie: seperate college GPA by male and female), and append them. Don't know too much more than that.

Any ideas?

 

[Napalm]

I think that you would be best served by factor analysis. Briefly, you set up an analysis that examines the variable(s) you want to predict (y = college GPA??) and the variables that you would have on hand when trying to make a decision for college admission (x1 = gender, x2 = age, x3 = hs gpa, etc...). Then you do a factor analysis that would searches out the predictor (x) that accounts for the most variance in y. Factor analysis then searches out a second variable that accounts for the second greatest amount of variance, and so on until the other variables no longer account for any significant amount of variance. At the end, you will be left with a set of ordered predictive variables with a certain weight that account for the maximum amount of variabilty.

Note that this is different that doing a bunch of simple single correlations between the same variables and rank ordering them in terms of strength. Since many of the x values are themselves intercorrelated, adding one to the matrix may not give you much in terms of additional predictive power. For example, conisder an analysis where you are trying to predict basketball ability in NBA players (i.e., ave points/game) and you have as your variables, length of left leg, length of right leg, reaction time, and strength. Doing simple correlations would likely yield decent correlations for right and left leg length. However, these two variables would be highly, highly intercorrelated and as such the second variable would likely not be used in a correlation matrix since it accounted for virtually the same variance. Get it?

Anyway - use a factor analysis and load the strongest predictor, followed by the next stongest predictor that accounts for unique variance and so on...


Napalm

 

[Capn]

Napalm is probably on the right track, but since I'm lazy this is what I'd do. Go get a program called "Datafit", this is a nice little application that can run regression models on data. It does more than just X-Y regression, it does multivariable regression models too. You can make it run all of the standard forms and rank them by the correlation coefficient, also you can input your own models. So determine what a good dependent variable as college success and see what it spits out.