Hi,
This is my first ever research I am doing for my MSc so please bear with me.
N=422, looking at the histogram and the skewness coefficients it can be confirmed it is skewed. Scatter plots tell me there are 3 weak to moderate relationships between IVs and DV linear and 1 curvilinear relationship but also seem to be homoscedastic. Kolmogorov-Smirnov values for all variables were significant, once again confirming non-normal distribution.
I tried removing outliers but did not help the distribution. I transformed them in spss, following Pallant (2016) recommendation by using Reflect and Logarithm, Reflect and inverse as these shapes matched my histograms. The results are still far from normal.
Following SPSS survival manual (Pallant, 2016), is seems there are limited options to what I can do at this point. I can report the descriptive part of my analysis for the dissertation and what usually happens is people report correlations next and then they test their model with some form of regression. And if they test moderating effect of a variable they seem to do SEM.
Now I chose Spearman's correlation and described the association between the variables (rather than correlation) then I learned that it can also be used for hypothesis testing? But it can only be used for hypothesis such as H1 There is a relationship between this and that. It that correct? If so, that is fine, three of my hypotheses are formulated according to this. But what about the moderating effect? Is there a non-parametric way of doing it?
As an answer to my question regarding regression with data like this, my supervisor said OLS does not assume normal distribution (referring to the Gauss Markov theorem) so I could look into that and that it is recommended not to rely on normality tests but the diagnostic plots of residuals. Or stick to Spearman's but then he says I need to change my approach accordingly.
So once I use Spearman can I in addition use OLS? And if so, is there a non-parametric way of testing for moderating effects of a variable on the other relationships? Sorry about the long message, just wanted to give you enough context.
What is your research question? What are you trying to accomplish? If this is survival data can you plot it so we can see what happens as time increases? Are any observations censored? Please note that survival methods and ols regression don't generally go together. Best wishes, David Booth
Yes. Don't rely on normality tests. Don't remove "outliers" unless they really are bad data. Yes, look at residual plots, but also know that some skewness is probably not problematic. If the data really aren't from a conditionally normal population, then a different kind of model is probably warranted. (See generalized linear models.) ... It's fine to present Spearman correlations as an initial step, and then go on to more complex modelling.
If you have SPSS 26 or later, take a look at the QUANTILE REGRESSION command.
- https://www.ibm.com/docs/en/spss-statistics/26.0.0?topic=regression-quantile
- https://www.ibm.com/docs/en/spss-statistics/26.0.0?topic=reference-quantile-regression
Thank you Mr David Eugene Booth I am using the Theory of planned behaviour by Ajzen to see whether Covid 19 changes intentions to attend events (football games in this scenario so my sample comprises of football fans). I am trying to see if it moderates the relationships of the model and it is a cross sectional survey study.
Thank you Mr @Salvatore S. Mangiafico, thanks for confirming some of the things I was worried about. I did read that outliers are often not outliers but the representation of the sample and should not be removed unless are due to missing data, error or they are really sitting out on their own.
Thank @Bruce Weaver, unfortunately my license for spss 26 has expired and now I am using spss 25 and I am not sure if it's a good idea to delete and reinstall that in the middle of data analysis, but I will look into the links.
A couple of things here. If you want to try quantile regression I have an R function that does that. The R function is hqreg and you can google this for more details. A shout-out to Bruce with a Thank you for that suggestion.
Now Theory of planned behavior ideas could be tested by the appropriate models in hqreg. I am attaching an R program that is not entirely complete but will do most of the analyses as is. Quantile regression I have yet to implement but you may try it if you like. If you are not familiar with R just download Jared Lander, R for everyone available from the z-library for complete instructions.. I have added some teaching comment s to my program which should help you to get started. If you have questions please ask.
David Booth. PS R program will be attached to my next response. Have some difficulties with my phone. Back ASAP
Matyas, attached is my latest update of my hqreg R program. Please contact me me if you have any questions or concerns. Google hqreg R function to get started. Best wishes, David Booth
Re quantile regression, another option for SPSS 25 is the R extension command SPSSINC QUANTREG.
- https://www.ibm.com/docs/en/spss-statistics/25.0.0?topic=r-extension-commands-spss-statistics
- https://github.com/IBMPredictiveAnalytics/SPSSINC_QUANTREG
Thank you Sal Mangiafico , thanks for confirming some of the things I was worried about. I did read that outliers are often not outliers but the representation of the sample and should not be removed unless are due to missing data, error or they are really sitting out on their own.
Thank Bruce Weaver actually just realised i do have spss 26, Birkbeck gave us licences for a bunch of different versions and I have trouble with my old mac to use any of them but my somewhat newer windows 10 is handling it well, I have the quantile function, will def. check it out..
Thank You again David Eugene Booth I will have a read first thing tomorrow!
Matyas -
One thing that seems odd to me is that with skewed data you generally notice a large difference in size of data. In such a case, you generally should notice heteroscedasticity. Do you say you have homoscedasticity because of some hypothesis test? I would not pay much attention to that.
The presence of heteroscedasticity is not really a yes/no kind of thing, but more a matter of degree. (That is a problem with hypothesis tests, but often we can find a way to visualize the effect size.)
Cheers - Jim
PS - You might want to look at some of the following, including an update on hypothesis tests:
https://www.researchgate.net/project/OLS-Regression-Should-Not-Be-a-Default-for-WLS-Regression, and the various updates there, in reverse chronological order.
James R Knaub thank you for picking up on that, actually you are right, I meant to write heteroscedastic, and as you said it is a matter of degree and it is hard to tell for someone new to research to tell to what degree it is. thank you for the recommendation.
For 'degree' of heteroscedasticity, you could consider the following:
https://www.researchgate.net/publication/333642828_Estimating_the_Coefficient_of_Heteroscedasticity
with
https://www.researchgate.net/publication/333659087_Tool_for_estimating_coefficient_of_heteroscedasticityxlsx
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics
- Regression