Statistics and stats software phear and handholding thread

Me first. I'm embarrassingly confused about whether it's possible to do a two-tailed binomial test on pass-fail categorical data, in SPSS or anywhere. SPSS only seems to give me the option of a one-tailed test. Plus if something is either greater than chance or less than chance, that's yer lot, isn't it? But clearly it isn't, because my advisor is asking me to do a two-tailed binomial test.

I have not yet started my stats class, but I will be haunting this thread in T-minus two weeks!

Via hoos on FB - looks fantastic but too advanced for me: http://analyticsmadeskeezy.com/

Got 100% in A-level Statistics but could be damned if I know what a two-tailed binomial test is now. Something to do with interquartile ranges? Standard deviations?

No means, ranges or standard deviations for binomial tests, since you've just got pass/fail to deal with. But I can't work out where a second tail would come from! If we're betting the farm on something happening above chance level, then either it is or it ain't, in my tiny brain.

From what I gather, you're looking for data that falls in the zones of excess passes (tail 1) and excess fails (tail 2) as compared to expected pass/fail ratios, but this is BY NO MEANS certain

relevant Wiki quote:

However, if testing for whether the coin is biased towards heads or tails, a two-tailed test would be used, and a data set of five heads (sample mean 1) is as extreme as a data set of five tails (sample mean 0), so the p-value would be 2/32 = 0.0625 \approx 0.06 and this would not be significant (not rejecting the null hypothesis) if using 0.05 as the cutoff.

sounds like this is analogous to your situation ljubljana http://onlinestatbook.com/2/logic_of_hypothesis_testing/tails.html

say you do 100 trials, and you get 40 passes.

the one tailed probability is the probability of getting 40 or fewer passes, i.e. getting a result that extreme or more extreme in that direction.

the two tailed probability is the probability of getting 40 or fewer passes + the probability of getting 60 or more passes, i.e. i.e. getting a result that extreme or more extreme in either direction.

if the two tailed calculation isn't built in to SPSS (which i have never used, maybe it is) the trick is noting that the binomial distribution is symmetric about its mean. this means P(N < 40) = P(N > 60). so you can just do the one tailed probability and double it.

more or less what I said. not bad for a lit graduate EH CAEK

i didn't understand what you said tbh

Thanks, that's an excellent site, caek, I will go back to it for other stuff. And not only do I now understand why it's perfectly possible to do a two tailed test, I now understand why I'm supposed to be doing one, which I definitely didn't before.

imago, I sort of did understand what you said and I agree that you and caek are concurring!

All I know is the three of you are going to be my heroes. I expect my stuff will be much more basic than ljubljana's, actually. But I have low tolerance for learning new software packages if/when they piss me off! I am normally a pretty easygoing person but in these cases I get THE RAGE and want to kill things.

yeah ljags was right, i just don't think that wikipedia quote is clearly expressed (like basically every wikipedia page about a technical subject)

i get the feeling that most of them are written by graduate students and then lightly edited by undergrads, probably a lethal combo

xp to q - what I'm doing is super basic. It all just sounds hard, and I'm very slow at it. I get there in the end, it's just that often I'm the only one who can see the end in sight, while teachers roll their eyes.

Embarking on test theory, along with latent class analysis, DIF analysis, EFA/CFA, etc.

Ugh, sorry about that. SPSS and I have ended our relationship, yay! That class was actually not so hard, and I slacked BIG TIME and still got an A.

Does anyone know anything about modeling in Mplus? I can't decide whether I want to constrain my parameters (fnnnnarrr). (this is for a latent class analysis).

Who amongst us knows just how non-normally distributed residuals have to be before they are too laughably non-normal for a 3-way repeated measures ANOVA? I keep reading stuff about the robustness of ANOVA to non-normal residuals, but can't find a guide to how far that stretches. And the culprits are 0s in my data, which I think complicates things further. This is for an analysis of looking time to various images on a screen.

i don't know anything about that stuff. i did a phd heavy on stats and first heard of anova during my second post doc. weird how stats is so domain specific.

my holiday project is to understand latent dirichlet allocation, lol

I still ride for LDA fwiw screw you deep learning

Gpower tells me that for a one-way ANOVA with two groups I need 128 participants (64 per group)

How come people seem to routinely do these tests with a smaller N? I mean, how come the extremely serious and often po-faced contributors at cross-validated don't pick up this guy on that problem?