Statistical analysis selection question

[loveoforganic]

---

Members don't see this ad.

This is probably pretty basic, but my stats background isn't too strong.

1) What's the proper method to look at differences in one variable (a rating scale with only a limited number of integer values) between two samples while controlling for the effects of a second variable (another integer value, but not on a rating scale)? Is it just a MANOVA?

2) If you're looking at a correlation between two non-dichotomous variables, where both of them are on a ratio or interval scale, you use either pearson or spearman correlations, correct? Pearson would be used if your data is parametric and spearman if it's nonparametric? When deciding if your data is parametric or not, do you just eyeball it or is there a test with some kind of acceptable "fit" value?


Thanks in advance!

 

[JockNerd]

1) What's the proper method to look at differences in one variable (a rating scale with only a limited number of integer values) between two samples while controlling for the effects of a second variable (another integer value, but not on a rating scale)? Is it just a MANOVA?

MANOVAs are for multiple DV multi-group analyses. The "controlling for" is a key that you need to use a covariate approach. Probably ANCOVA.

2) If you're looking at a correlation between two non-dichotomous variables, where both of them are on a ratio or interval scale, you use either pearson or spearman correlations, correct? Pearson would be used if your data is parametric and spearman if it's nonparametric? When deciding if your data is parametric or not, do you just eyeball it or is there a test with some kind of acceptable "fit" value?

Non-normality basically, abiding by standard guidelines (Skew raw value less than 2, kurtosis raw value less than 5), or ordinal data would be cases where you'd use Spearman's.

I'll jump on a little soapbox that analyses methods should be decided upon before data are even collected 🙂

 

[loveoforganic]

Thanks a bunch! Just in reference to your soapbox comment, how can you decide on the correct correlation before data comes in? In your proposal, would you just say pearson used unless skew <2 and/or kurtosis <5, in which case spearman will be used?

 

[JockNerd]

Thanks a bunch! Just in reference to your soapbox comment, how can you decide on the correct correlation before data comes in? In your proposal, would you just say pearson used unless skew <2 and/or kurtosis <5, in which case spearman will be used?

It was more in reference to the anova than the correlation.

 

[loveoforganic]

Gotcha!

Thanks again 🙂

 

[Ollie123]

Alternatively, you can run transformations to try and make something more normal. Which one you use depends on what the data looks like, but I'm blanking right now on which options go with which data. I usually won't transform unless the data is massively skewed though (skin conductance data is notorious for this). It may matter for correlations (embarassingly, I don't actually know) but for ANOVA/ANCOVA etc. they are pretty robust to departure from normality - which doesn't mean anything goes, but it does mean that I generally won't change anything if my skewness comes out at say....2.5 or 3 because I consider it "close enough" that I can't justify changing the variables, removing any ability to make comparisons to other research, etc. However, I just submitted an abstract where skin conductance came out with skewness in the neighborhood of 400, and kurtosis of ~2000. That one got transformed😉

 

[JockNerd]

Alternatively, you can run transformations to try and make something more normal. Which one you use depends on what the data looks like, but I'm blanking right now on which options go with which data. I usually won't transform unless the data is massively skewed though (skin conductance data is notorious for this). It may matter for correlations (embarassingly, I don't actually know) but for ANOVA/ANCOVA etc. they are pretty robust to departure from normality - which doesn't mean anything goes, but it does mean that I generally won't change anything if my skewness comes out at say....2.5 or 3 because I consider it "close enough" that I can't justify changing the variables, removing any ability to make comparisons to other research, etc. However, I just submitted an abstract where skin conductance came out with skewness in the neighborhood of 400, and kurtosis of ~2000. That one got transformed😉

Skew will mess up a Pearson's correlation pretty badly... but, as I think about this, the amount of skew you can have with scale ratings is limited. If you're dealing with something like age, though, a single high value will mess up everything. I recently skimmed a cruddy article in which they reported that, among doctoral students, age was positively related to more positive attitudes toward homosexuality. The problem: age was distributed something like M = 22, SD = 3, Range = 21-46. So, one liberal 46-year-old throws everything off. In such a case, a Spearman's resolves the issue statistically, but not conceptually; in the end, the design must be made to address the research question--people should not run around plugging data into analyses all willy-nilly. In my example, if an inference about age and attitudes toward homosexuality was intended to be assessed, it should have been incorporated into the design (e.g., intentionally sample multiple age cohorts of doc students) rather than being chucked in there as a poorly-designed afterthought.

Ollie, with the skin conductance example you gave... did you look at next-highest-value transformation for the outlier(s)? That would be my first thought.

 

[estatistics]

even this is an old thread, i would like to add a comemnt!

Normality is essential for parametric testing.
Complicated designs, such Manova better be off these!
Simpler designs the best ones! Dont complicate your life!

1) You must check for outliers
2) You must check for Normality
3) Do simpel tests to see the trend of your data
4) If you have multivariate normality, then you may try GLM structures (Manova is part of GLM)

You may downgrade your data to simpler form e.g. (0/1) if this a better approach (from ratio/interval to ordinal and then to Dichotomous, to nominal).

As a statistician / psycholigst, sometimes i have obtained better results by downgrading variables' properties than trying to fix them with transformations! I dont know however your data structures!

According to your second question, the fellow members, i think they covered the topic of correlations very well!

sincerely,
Elias "Estatistics" Tsolis
---------------------------------------------------------------------------------------------
The above opinion is only my subjective personal opinion / experience - and it is not a professional or scientific opinion.
--------------------------------------------------------------------------------------------
My Bsc was on Psychology and a Msc in Org. Psychology.
Also My hobbies: I Like Music and I do an analysis to music rythm.
Also, I like to analyze my dreams. I have made a theory on this.
I like Online Learning in Coursera (you may find me as Estatistics).
Moreover, I like Statistics!
My ResearcherID is A-1480-2016 and my related ORCHID is 0000-0003-3426-9393.

 

[MCParent]

My Bsc was on Psychology and a Msc in Org. Psychology.

As a statistician / psycholigst