Statistics Help @ Talk Stats Forum - Applied Statistics

Newbie to stats - comparing pre & post test results with missing data

miksima — Fri, 17 May 2013 23:15:48 GMT

Hi

I'm a complete beginner to stats and I can't work out how I can analyse some data with SPSS despite it being very simple on the surface.

I have 2 lists of quiz test scores.

List 1 is scores for a group prior to receiving training.
List 2 is scores for the same group after receiving training.

None of the scores have any identifying participant information so I cannot match them against each other.

Also not everyone took part in the post-training test.

So for example: -

Pre-training scores = 14, 1, 13, 10, 3, 13, 13, 11, 14, 5
Post-training scores = 13, 15, 9, 15, 14, 15, 15

That's it, that's all the data I have. I know how to work out the mean but is it possible to analyse this kind of data any further?

I'm sorry for posting such a basic question but I'm hoping that a generous stats-wizard can help out a complete amateur!

Thanks very much!

How would I determine significant factors for a conditional drop?

badmax — Thu, 16 May 2013 23:09:28 GMT

Hi,

I have a financial data set of company-specific financial returns post a recurring event. I am interested in determining which factors would differentiate large negative returns from everything else.

I first thought of LDA but the predictors and responses are very far from normal so I didn't attempt it. I tried using logistic regression. I set my response variable equal to one if the returns were less than -5% and 0 otherwise (arbitrary cut-off, I thought I would iterate through several choices and look for consistency). This model worked terribly. The only time it worked (~25% accurate predictions in-sample) was when I used a huge number of predictors, about 270. The AIC was higher than sparser models but the predictions were much better. Because of this I had no faith in it and I dropped it.

My next idea was to use a much sparser linear regression model, about 20 predictors. First I fit the model and then I added interaction terms for all the coefficients one-by-one (add1 in R). I then made a list of of the coefficients that were statistically significant and, using a different sample, I classified the companies by deciles on each event date according to their standardized values of the significant predictors. I looked at the distribution of returns above and below a certain threshold in the deciles. An ideal outcome is when A) the predictors are statistically significant in the first sample, B) either decile 1 or 10 has more returns below the thresholds than anticipated (ie more than 10%) and C) the deciles that look good in B) also have less returns above a given threshold.

Essentially I am looking for right skew in the tail deciles.

I have tried switching from linear regression to quantile regression, using low values of tau. Both methods seem to work (for now, still working on it) in that the statistically significant predictors translate well into the out of sample decile tests.

I would appreciate comments and suggestions about my method. I can simply skip the whole regression part and apply the decile test to everything but I feel this way I have more confidence in my model, do you agree? Do you have any improvements you can think of?

Thanks a lot.

Econometrics question

Falc — Thu, 16 May 2013 10:02:12 GMT

Econometrics question: http://i.imgur.com/dCHdGnm.png [1]
I would like to check that my answers are correct: for a. I wrote that it would be a DF test (rho - 1) / standard deviation (rho). H0= rho=1 , H1: |rho|<1. The ciritical values would be drawn from a nonstand distribution under a null hypthesis of nonstationarity which is skewed to the right and so the critical values are smaller than that of a standard t test. And the test would be a one sided test.
For part b, would it be a Phillips-Perron test?
For part c I said it would be a standard T-test. (rho-1) / sd(rho). CV from t distribution. H0: rho = 1. H1: rho =/= 1.
Is there anything else I should include in the answer?

One further question: http://i.imgur.com/CeotSNN.png [2] Would I do IV and then do Random effects / GLS?