I have been told that there are discrepancies between the results you get when running e.g. the same multilevel models on the same datasets in R and SPSS, and that R is more precise.
Can anyone elaborate on this, confirm/deny, and direct me to literature on this relationship?
Not all researchers are interested in computer programming. According to my understanding of R you would need to know how to program and therefore R lacks the easyness to use that goes with SPSS.
Thanks! I'm proficient with SPSS and can stumble some around in R, so the question is mainly one of the differences in outputs, where they might come from, and what they mean.
Fernando, that actually makes sense. The default settings may be the thing!
Great answer Fernardo! You're absolutely on point with the default options.
With regard to "precision" - If the SPSS user does not account for the study's sampling design or weighted responses using the "complex sampling" option then their results will be less statistically valid than if they had used the complex sampling feature.
Complex sampling improves statistical validity by accounting for the study's sampling design. This is especially important when using large databases that need to account for response weights.
I prefer SAS, use SPSS, and would like to learn R. Looking forward to other responses.
If the statistics by means of R is uniquely programmed for a certain research question it is supposed to be more specific or precise than a software designed to suit many research questions.
For computationally intensive statistical analyses (e.g. multilevel models), almost every statistics program will give slightly different answers. There are a number of reasons for this, many of which are related to the fact that the computer uses some type of iterative method to converge to a solution or find a maximum value. The exact algorithm used tends to differ between software programs. The size of each step in an iteration can also differ. It is also worth noting that many things happen behind the scenes in software programs of which users are often not aware.
For some R packages, the authors will explicitly explain why results differ from SPSS. The psych package by William Revelle is a good example of this. Factor analysis results can be different between R and SPSS, because SPSS performs a Kaiser normalization prior to rotation, and R's GPArotation does not do this. This is explicitly explained in the manual for the psych package. http://cran.r-project.org/web/packages/psych/psych.pdf
Other packages such as lavaan discuss how their results compare to those of Mplus.
As for precision? This does not necessarily have anything to do with differences in answers. In general small discrepancies are common between software programs.
Perhaps you find this interesting:
Keelinga, B.K., Pavurb R.J., 2007. A comparative study of the reliability of nine statistical software packages. Computational Statistics & Data Analysis 51, 3811 – 3831.
R and SPSS will have the same results for the type of functions implemented in SPSS but R is more incorporated in several analysis including complicated computation which give it advantages especially if it is free.
That makes a lot of sense. Especially what you say Paul, since the issue was with fairly complex nested models. I'll get the paper straight away, Rense - might be very useful!
@ Eduard: The SPSS algorithms can be viewed at the link given below--see the panel on the left.
http://pic.dhe.ibm.com/infocenter/spssstat/v22r0m0/index.jsp
The precision comes from the software using either a float or double float in the computations and the algorithm. Most of the time, these differences don't really matter because you have a SE Coeff that is much larger than float of double float.
For example, if you have a coefficient of 12.0000000000000000000000000000000000000000000000001 and a SE for that coefficient of 1.5, the confidence interval for the coefficient is 12.00000... +/- (t-value)*1.5. So, your coefficient will actually be between (roughly) 9.0 and 15.0. So, it doesn't matter all that much if R gives you 12.000...3 and SPSS gives you 12.000...014.
You should be able to increase the precision on each. NIST has sets of data to test stats software with. If you use this data, (which I posted in another link that I can't find right now), you can see for yourself which is better. You will have some issues if your data are "Ill-conditioned" ( A term from linear algebra) This occurs if you have a lot of data that does not vary much. Say, you want to use BMI as a factor and all of your BMI values are between (29.999, 30.001). Otherwise, don't worry too much about that.
the same question was discussed here
I think it can help
https://stackoverflow.com/questions/3787231/r-and-spss-difference
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Statistical Software
- R Statistical Package