comparing extremes of an independent variable

#1
Dear all,

I have been cracking my head for 2 weeks but i havent been able to figure out this, so i decided to ask for help.

I have done a regression between my independent and independent variable and found nothing.
Now my supervisor wants me to find out if what i wanted to find, can only be found at the extremes.

So participants with the lowest scores (10% left) and participants with the highest scores (10% right).

My supervisor didnt give me much help, she only said something about using ANOVA?

I hope someone can help me.

kind regards

Niccolo
 

kiton

New Member
#2
Dear all,

I have been cracking my head for 2 weeks but i havent been able to figure out this, so i decided to ask for help.

I have done a regression between my independent and independent variable and found nothing.
Independent and dependent variables, I assume. As otherwise, you are indeed likely to find nothing :)

Now my supervisor wants me to find out if what i wanted to find, can only be found at the extremes.

So participants with the lowest scores (10% left) and participants with the highest scores (10% right).

My supervisor didnt give me much help, she only said something about using ANOVA?
What you can try here is run a quantile regression (QR) -- one that would allow you to estimate the impact of your predictor on the DV in the lower (e.g., 10th) or higher (e.g., 90th) percentiles of the distribution, if that is what you are looking for.

Check this paper to get a sense of what QR is all about: Cade, B. S., & Noon, B. R. (2003). A gentle introduction to quantile regression for ecologists. Frontiers in Ecology and the Environment, 1(8), 412-420.
 

hlsmith

Omega Contributor
#3
Can you tell use more about your data. Do you have two continuous variables or if not how are they formatted?


Have you plotted you data at all to see if there may be trends or possible visual differences. Visualizing data is important in figuring out which direction to go.
 

rogojel

TS Contributor
#4
My supervisor didnt give me much help, she only said something about using ANOVA?

I hope someone can help me.

kind regards

Niccolo
Hi,
what your supervisor might have in mind is to build two groups, LOW and HIGH say, and to run an ANOVA with the two groups (actually a simple t-test is enough in this case).
regards
 
#5
Can you tell use more about your data. Do you have two continuous variables or if not how are they formatted?


Have you plotted you data at all to see if there may be trends or possible visual differences. Visualizing data is important in figuring out which direction to go.
I will elaborate more yes, thx for the answers!

I have 5 independant variables (personality traits) and my dependant variable is Telomere length. All continous.

Doing a multiple regression i found that my independant variables had no significant effect on my dependant varable. Now i need to figure out if the effects do appear when looking at the extremes.

Hope this makes it more clear?

Im going to read that QR thing, first time i hear about it, thx!
 

hlsmith

Omega Contributor
#6
So your description still reads a little weird. So you have 5 different continuous variables, correct? Do you want to see if all of these may differ by extremes?


What is your sample size?


You may want to strategically break the trait in categories, then make lowest level the reference group for comparisons.


So you think traits predict lengths. I would wonder in general about that hypothesis and in addition wonder if you have the independent and dependent variables actually switched?
 
#7
So your description still reads a little weird. So you have 5 different continuous variables, correct? Do you want to see if all of these may differ by extremes?


What is your sample size?


You may want to strategically break the trait in categories, then make lowest level the reference group for comparisons.


So you think traits predict lengths. I would wonder in general about that hypothesis and in addition wonder if you have the independent and dependent variables actually switched?
the sample size is substantial, is almost 10 thousand participants.

maybe giving a concrete example would help to explain myself.

Scores on a persality (neuroticism) may have a negative effect and thus shortening the length of the telomeres. Because i havent found any statistical significance during my regression. I postulate that the effect of personality traits may only be seen when u look at the extremes. And thus comparing lowest 10% and highest 10% of the scores of said trait onto telomere length.

This exact thing, i have to do with all 5 personality traits. while taking the age of my participants into consideration.

all my variables are continuous (dependant and independant ones)

the english lenguage isnt my forte, i hope i explained myself a bit better this time
 

hlsmith

Omega Contributor
#8
That did help.


Quantile regression deals with grouping the dependent variable, which you want to chop up your independent variable. You can chop up your IV and control for age, you just need to let people know if you disseminate results that this was a post hoc analysis. I am not familiar with multivariate regression, but I believe you want to go that route because you are examining multiple things against the dependent variable. So yeah, go ahead and dissect the middle values and run your model with all chopped up variables!
 
#9
That did help.


Quantile regression deals with grouping the dependent variable, which you want to chop up your independent variable. You can chop up your IV and control for age, you just need to let people know if you disseminate results that this was a post hoc analysis. I am not familiar with multivariate regression, but I believe you want to go that route because you are examining multiple things against the dependent variable. So yeah, go ahead and dissect the middle values and run your model with all chopped up variables!
Thank you! is there any way to achieve that on SPSS? Because i read that QR is done in R (which i know little to nothing about)
 

kiton

New Member
#11
hlsmith is right in that QR "chops" the DV, not the IVs. As such, if I understood his proposition correctly, you can "chop-out" the middle values of your independent variable(s) and keep only those that are at "extremes". For example, assuming the values of your IV(s) range from 1 to 100, recode 11 through 89 as missing values (or use -if-) -- then run the regression using this "chopped" variable(s). Note, your sample size is likely to decrease but having 10k cases you'll have plenty left, I guess.
 
Last edited: