In my current work, we have calculated the p value for association using multiple linear regression in SPSS for 6 SNPs (three different models -additive, dominant, recessive) with several quantitative traits eg: axial length. There is no option for correction (post hoc as in ANOVA) available in SPSS linear regression window. So
i) Is correction for the p values obtained from this is absolutely necessary,
ii) Is there an option in SPSS that i have missed,
iii) can a manual bonferroni correction (0.05/ the no of SNPs) be sufficient/ appropriate
Kindly help out with the same.
Thanks in advance
It is unclear to me what you would want to correct for / why you would want to correct. Can you explain this in a bit more detail?
If there is a relation between your dependent variables, you may want to think about using SEM instead of spss multivariate regression.
You can always correct "by hand" by dividing your alpha level (.05) by 3, if you did 3 analyses. However, to indicate whether that is necessary or not, we would need more information about your analyses, and specifically about the question why you would want to do a correction.
Hi Gerine
In my analysis a set of SNPs were evaluated for association with axial length (SPSS-linear regression method) using three different genotype models (dominant, recessive, additive) and the least p value from the three were finally taken for reporting. Few of these SNPs showed significant association with axial length (p
I am not in medical science, so it could be that there are different methods used in medical science. However, it seems odd to me to just report the analysis in which you found the lowest p-value. Wouldn't it be better to report all three? This sounds a bit like fishing...
Furthermore, if you have multiple DV's that are predicted by the same predictors, it would make sense to use path analysis. This can be done using AMOS (spss add on), Mplus, R, SAS or a variety of other statistical packages.
If you are sure this is not the way to go, you could apply the Bonferroni method manually. You would do this by deviding the alpha level by the amount of tests you used (alpha / k, where k is the amount of hypotheses, in your case alpha/3). There is some discussion about whether or not this is a good idea, for instance, see the paper attached to this answer.
Article A farewell to Bonferroni: The problems of low statistical po...
Thanks for replying.
As per the literature available from similar work, this is how it is done for genetic association studies.
I will definitely have a look at the link as well..
As a general rule, there are no corrections for MLR. Assuming that your results abide by the assumptions of ANOVA.
How many responses do you have? Are your responses categorical or continuous?How many of your factors are categorical and how many are continuous?
In this study i have used 6 independent variables (categorical variables; dominant and recessive model will have 2 categories and additive has 3) and tested for 3 continuous dependent variables.
Hi Ferdinamarie.
In general, there is no need to adjust p-values in MLR as they are already ajusted for the presence of all the variables in the model. Of course you need to check the assumptions of the linear model (errors with normal distribution, homocedasticity, inexistence of multicollinearity, inexistence of serial correlations among the errors...) as this can influence the estimations of your model.
I hope this will help you!
Best
Frederico Rosário
If you have 3 response(dependent) variables, you might want to look at using MANOVA . If all of your input variables are categorical, I don't think MLR is the correct method to use. GLM would be a better way to go.
In response to Frederico and Andrew:
It is true that there is no correction needed for the predictors, if you test them in the same model. But, there may still be correction needed for the dependent variables, as in spss you cannot test several dependent variables at once. This can be done using path analysis, as I indicated earlier.
Regarding MANOVA, I am not sure whether that would be the way to go here, because the dependent variables are continuous and the independent variables are categorical rather than the other way around. That being said, you could in some cases switch them around, especially when you have a correlational design without control variables.
- Badges
- Science topic
- Similar topics
- Mathematical Sciences
- Statistics
- Statistical Software
- SPSS