Any statisticians around?

[VintagePunk]

Preferably those who know statistics in a research context?



I need information/advice regarding the skewness of a distribution, and what level of skewness is acceptable before the distribution (of a variable) needs to be transformed.

I know it's a longshot, but I'm desperate.

 

[U2Man]

 

[snowbunny00774]

Oh VP, that sucks

 

[Vincent Vega]

Can you ask me in half a year again?
I'm just starting it.

 

[Angela Harlem]

do you mean where the whatsit is 0.05 or under, meaning no correlation? honestly, i just finished a unit on research methods (in justice, but it's not unique to the field, obviously), and i have momentarily blanked everything out. which is a shit because i need to remember it all for another unit coming up next.

 

[U2Man]

I need information/advice regarding the skewness of a distribution, and what level of skewness is acceptable before the distribution (of a variable) needs to be transformed.

i dont think that there's a single definitive answer to this question. it depends on the model and parameters etc.

 

[VintagePunk]

do you mean where the whatsit is 0.05 or under, meaning no correlation? honestly, i just finished a unit on research methods (in justice, but it's not unique to the field, obviously), and i have momentarily blanked everything out. which is a shit because i need to remember it all for another unit coming up next.

Yes, it does have to do with correlations, or in my case, lack thereof.

If your sample is skewed (ie. not a normal distribution), there is a good chance that when performing correlations, you will get a false negative - the two variables will appear to be not correlated, when indeed they are. To fix this, you transform the variable to make the distribution normal. I'm working with two variables that should be correlating, but they're not.

Ideal skewness is zero. As I'm sure you know, rarely in research is anything ever ideal. So, I need to know what level of skewness is acceptable, before you do a transformation.

What I've been told is this - you eyeball the distribution, and if it's not normal, do a transformation. I've done histograms with a normal distribution laid over top of them, and while mine aren't perfectly normal, they're not nearly as bad as some of the examples of positive and negative skewness I've seen. The two distributions I'm concerned with right now, the skewness of them are -.233 and .222. I want to know if that's within the bounds of being "normal enough."


U2Man, what do you mean by the model and the parameters?


Does anyone know how to interpret the numbers of either the Kolmogorov-Smirnov or the Shapiro-Wilks tests of normality?

 

[Angela Harlem]

Is your significance level set at 0.05? It's a standard which you can use if you haven't been given one, but is going to affect the rejection of your null hypothesis, as you already know. 5 and 1 percent levels are the most common. Do you have to do t-tests and all that, too? We never did Kolmogorov-Smirnov or the Shapiro-Wilks tests of normality, unfortunately.

 

[VintagePunk]

Is your significance level set at 0.05? It's a standard which you can use if you haven't been given one, but is going to affect the rejection of your null hypothesis, as you already know. 5 and 1 percent levels are the most common. Do you have to do t-tests and all that, too? We never did Kolmogorov-Smirnov or the Shapiro-Wilks tests of normality, unfortunately.

Yeah, SPSS flags all significant correlations, and also gives their significance levels.

We covered t-tests and z-tests in my intro classes many moons ago, but no, I don't need to do any for my research. I'm looking at correlations at the moment, and depending on what comes of that, will probably do multiple regression and/or structural equation modeling.

Thanks for your help, though.



ETA - I looked up the two normality tests I mentioned earlier and I now know what the numbers mean, but I still don't know how normal is normal enough.

 

[Angela Harlem]

so you have one positively and one negatively skewed? and they are 'within' normal? Or rather not correlated?.... (trying to guess here from your numbers )

SPSS is satan spawn I love statistics, and don't mind the programme, but something about this two-tailed testing and pearson chi-square bizo etc makes me want to kill people. I severely understimated statistics and research. I'm fine with data that are already worked out and I just need to interpret. But not all this stuff.