Hi!
I would like to ask the following:
I have between other independent variables severity: severe, high tisk, medium risk, low risk and no risk; and the variable outcome: discharged, death, admitted, transferred.
The dependent variables is Emergency Room lenght of stay in minutes, as usual Lenght of stay is skewed, not normal distribution.
I proceed wth logistic regression and it seems that there might me a collinearity between these two variables.
My question is wheter to use Spearman test to check collinearity bewtween these to independent variables.
Thank you
Hi Maria,
To check for multicollinearity, run a simple (linear) regression as there is no facility for checking collinearity in logistic regression.
1) To do so, transfer you independent variables (excluding interactions, if there is any) and the dependent variable to their respective spaces in Linear Regression (SPSS)
2) Select Enter in Methods
3) Click Statistics and in the window (Regression coefficients) check Estimates, Confidence Interval, Model Fit and Collinearity Diagnostics then run the analysis
[Each of the statistics in 3) above measures a specific assumption, but let's stick with your question on multicollinearity]
4) Look at VIF (Variance Inflaction Factor) under Colliniearity Statistics in the output. IF any variable shows VIF of >10 then that variable has multicollinearity problem.
Good luck!
The collinearity diagnostics computed by the REGRESSION procedure in SPSS are appropriate for single-df variables only. If you have categorical variables with 3 or more categories, you need to compute generalized VIF (GVIF). AFAIK, SPSS has no routines to compute it. You may need to use R. HTH.
- https://stats.stackexchange.com/questions/70679/which-variance-inflation-factor-should-i-be-using-textgvif-or-textgvif
- http://math.furman.edu/~dcs/courses/math47/R/library/car/html/vif.html
Bivariate correlations (or more generally, associations) are not terribly useful for diagnosing multicollinearity, because you can have multicollinearity in the absence of any alarming pairwise associations, and conversely, you can have alarming looking pairwise associations that do not signal problematic multicollinearity. See the example by Jerry Dallal, which is mentioned in this old post to the SPSS usenet group.
https://groups.google.com/d/msg/comp.soft-sys.stat.spss/iHjKuJQAZoo/4soJFLgqqUsJ
I repeat here the data from Dallal's example:
X1 X2 X3 Y 18 88 106 13 72 45 117 43 36 63 99 50 75 26 101 77 22 83 105 23 99 71 170 68 69 53 122 6 6 49 55 51 86 99 185 37 85 64 149 10 87 7 94 32 93 32 125 69 44 88 132 4 34 34 68 13 84 28 112 18
Here is the correlation matrix (Pearson r):
| x1 x2 x3 y
-------------+------------------------------------
x1 | 1.0000
x2 | -0.3068 1.0000
x3 | 0.6573 0.5156 1.0000
y | 0.2882 -0.2825 0.0358 1.0000
There is nothing terribly alarming there. But now try to regress Y on X1, X2 and X3. If you do, one of the variables will be excluded due to complete linear dependence. Why? Because x3 = x1+x2.
So I repeat what I suggested earlier: I think Maria needs to find a way to compute Generalized VIF (GVIF) if she wants to check for multicollinearity.
HTH.
Dear Ahmet Aytekin:
Thank you so much for your answer.
- Collinearity between outcome and severity can be done with Kendall's Tau, I think the kind of kendall is C rather than C because the number of categories of these variables is not pair. When I did this the results indicated p: 0.000 and value 0.304 These means a relationship between those variables but a low one.
- About Multicollinearity, I found a way to check it with VIF. I entered the other variables, but the ones that had more than 2 categories I converted into dummies. I hope this is a correct way. But I am concerned is when doing this, the category discharged of the variable outcome was excluded on the final model.
Maria
Dear Ahmet Aytekin :
Thank you so much for your kind answer.
Regards,
Maria
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS