# Thread: Help with choosing statistics

1. ## Re: Help with choosing statistics

to add my two cents here:

The main difference between parametric and non-parametric statistics is that with non-parametric, you cannot define your distribution with a finite number of parameters. Note that this applies to the statistical models you use as well (statistical models are defined as the set of distributions that are assumed).

To continue on to the OP's question, I would like to be sure...are you just asking if the ranked position is different among the four figures, or are you asking if one is statistically higher than another? If it's the first, one option is to set up a 4x4 table (rank x figures) and use a chi-square test. If it's the second, we'll need more of a description.

2. ## Re: Help with choosing statistics

In the texts I have seen it tends to talk of normally distributed data not normally distributed residuals. But that raises another question, why would the residuals be normal if the data is not?

I got this off topic from the original poster so I will leave it at that

3. ## Re: Help with choosing statistics

Think about a paired t-test. In that case our data is really the both sets of pairs. The "data" can look really non-normal. It could be anything in this situation. But as long as the differences look approximately normal then we're good. Regression is another example. We don't care what the distribution of Y looks like. We care about what the distribution of Y|X looks like. Another way to say that is to just write your assumption with respect to the error terms.

4. ## Re: Help with choosing statistics

ok...thank you

5. ## Re: Help with choosing statistics

Originally Posted by noetsi
In the texts I have seen it tends to talk of normally distributed data not normally distributed residuals. But that raises another question, why would the residuals be normal if the data is not?
You bring up a very good question. Maybe I can hopefully aid in clarifying (beyond what Dason's kindly stated already). I'm assuming that you're referring to regression since you're referring to residuals (if not, feel free to clarify and we can branch off from there).

First off, what you're doing when you set up a (linear) regression model is assuming that the data follows this linear relationship. A lot of people I come across get lost here in somehow coming to the impression that the dependent variable has to be normally distributed. This is not true (in the majority of cases)! Also, something to remember is that most software packages use maximum likelihood (MLE) to model the data. In MLE, the model makes the assumption that the errors are normally distributed with mean 0 and standard deviation . So it'll fit the data under this assumption and try to find the beta coefficients that maximize the likelihood under this assumption. Hence, you will tend to see that the residuals are normally distributed.

Note that there's a big difference between the model actually following a residual that's normally distributed and merely assuming that it does. MLE will do the best that it can at making them normal regardless of which it is. It's up to you to verify the model.

Hope this helps.

6. ## Re: Help with choosing statistics

Originally Posted by Dave4

There are 3 figures made out of the fourth. An I'm asking them to rank which one is most different, but they don't know which one is the 4th, so they just rank from least to most. (Does it matter what I expect them to answer for these figures? like in a frequency table/chi-square?) then I ask them to associate this figure with something specific, and then let them to the ranking once more. I'm pretty sure (just looking at the data) that there is a difference in their ranking. But I want to show this statisticly and which one of the figures that they think is the most different. Does this make things clearer?

One dude I talked to said I had to use non-parametric tests on this, but with you discussing this above I'm not sure anymore
Out of curiosity, there are two things I'm wondering. I hope you don't mind answering.
1) Lets say that we go through the long process of teasing out what you need to do and figure out the methods you have to implement to complete it. Are you sure that you'll be willing to and competent enough to carry it out? Or are you just trying to get ideas to pass onward and seek further help?
2) Is there someone on campus maybe or a TA that can also help you in person?

I only ask these questions because I'm concerned that we won't get anywhere otherwise. I'm definitely willing to help. I just don't want to waste my time.

7. ## Re: Help with choosing statistics

Link, yes getting help from a TA would probably be easier, but I want to try work this one out myself. About competence I'm not sure, I've had two statistic courses at the uni, but that's over 5 years ago and all is bit of a blur... I started reading up again, but I had a hard time using it on my data. So yea, if you want to and can spare the time I would really appreciate it.

8. ## Re: Help with choosing statistics

I'm going in circles here... I found somewhere that with big samples you should use parametric tests after all, so, since my data will be around 100 samples I guess I should be starting over!