anyone any good at statistical tests.

[HauntedLittleLunatic]

If I have a survey as follows:

The answer to q1 can be male or female.
The answer to question 2 can be yes or no.

My hypothesis would be that if you answer female to q1 you are more likely to answer yes to q2 (i know that's not a null hypothesis but hey ho!).

How can test this statistically with 2 sets of categorical data?

Hypothetical sample size approx 1000.
Hypothetical 50/50 male female response to q1
Hypothetical response rate of 10% yes overall to q2.

Hope that isn't too cryptic (not intended to be...but it is a hypothetical research proposal I'm working on and I have to say how I will analyse the data....)

Hope you can see why its in geeky stuff...

See, this is why i did science - too many ifs and buts in social sciences

(yeah I did science and now I'm pretending I can convert to social science.....and I don't like it...I want my numbers back....and my hypotheses..and my assays which give me a straight number.....I don't like words and questionaires.....)

You can have null hypotheses in the behavioural/ social sciences too, if that's the paradigm you're working in.

Petsonally I never got the hang of the far reaches of the qualitative methods, mind?.

Chi squared should work as you have known 'expected' values (50% for sex, 10% for yes), so can establish if you have more difference than expected in your 'observeds'.

I don't know what paradigm I'm working in....actually I do...I am trying to stay firmly in the positivist so I can at least keep some numbers...words scare me .....I plan to steer clear of the other paradigm thingy....sounds even more scary...

Yup. I could never work out how "grounded theory", for example, was different to "talking to people", so I stuck with numbers too!

chi2 is based on obs - expd. you calculate the expd assuming your null holds, ie expd "yes, female" is 10% x 50% x 1000 , but use the total and the response rates to each question you observe in your data. the null assumption you are testing is that q1 and q2 are independent, ie that the joint prob can be expressed as the product of the two marginal probs. does that make any sense?

I wouldn't both with yates correction when you have such large numbers, it will barely change anything.

chi2 is based on obs - expd. you calculate the expd assuming your null holds, ie expd "yes, female" is 10% x 50% x 1000 , but use the total and the response rates to each question you observe in your data. the null assumption you are testing is that q1 and q2 are independent, ie that the joint prob can be expressed as the product of the two marginal probs. does that make any sense?

I wouldn't both with yates correction when you have such large numbers, it will barely change anything.

I am nowhere as competant as Springforward, but might I suggest you take a step back from the tools for a moment and make sure you have a full understanding of the formulation of the question?

As I understand it, you have a fairly large sample which includes male and female respondents, and you want to identify if the answer to a single yes/no question, Question 2, is dependent on the sex of the respondent (dependency is the correct term rather than correlation as Springforward pointed out).

You have discounted the application of a null hypothesis on the basis that it is not positivist, but you have not proposed a positive hypothesis to test. An example would be "Females are x percentage points more likely than males to answer yes to Q2". We could test our sample against this hypothesis but the answer would not be very meaningful, and to answer the question 'is there a correlation' we would have to look at all possible values of x. So instead, we have a null hypothesis "females and males are equally likely to answer yes to Q2". If we show that the null hypothesis is very unlikely to be true, we have shown that the opposite, i.e. "females are more likely than males to answer yes to Q2" is very likely to be true - which is a positive statement, and exactly what we want.

Fishers Exact Test may be the most exact (duh!), but given the large sample size and bivariate states of both variables, χ² (chi-squared in case the encoding breaks) is an approriate approximation. You will be performing an "independence test", anyone reading this that wants to know more there are even YouTube videos!

Thanks Mr anchovy.

Unfortunately the scenario being tested is quite different from that which you describe (that has been assumed by others on the thread and I let it stand cos I don't think that actually affects the stats I need).

I take on.board what you are saying about framing the research question which has actually been reworded and.tightened up today independently of your comments.

I am not sure that I am looking at a dependence if such a small proportion of those that answer females to q1 answer yes to q2. Would it be more of an 'influence'.

I am a little confused about the need for a hypothesis....we were told that in social science you have a research question INSTEAD of a hypothesis...

Yes, the fact that Q1 relates to gender is a distraction, but it doesn't affect the statistics - what we are testing is whether the answers to Q1 and Q2 are independent. Do not impart a colloquial meaning to the term 'dependent', as a technical term it is simply the opposite of 'independent'; the closest common language term is probably 'related', and if there is any relation, no matter how slight, between the answers then they are said to be dependent. Influence implies causality which has nothing to do with statistics

Fair point! Hadn't thought about it as the opposite of independent in that way...

Oh give me back my proper science with proper cause and effect and numbers. I don't like this social science malarky.... ignores the fact that I thought this masters would be a good idea