The FastDay Forum

the mean of the 25 lowest losers is 0.63 kg and that of the 25 biggest losers is 4.49 kg. Variance 0.4 and 1.4 resp. The difference in the groups is statistically significant P<0.001

[carorees]

the mean of the 25 lowest losers is 0.63 kg and that of the 25 biggest losers is 4.49 kg. Variance 0.4 and 1.4 resp. The difference in the groups is statistically significant P<0.001

That's a good start anyway! What test did you use?

I think we need to go to 27 users from the lower group due to the top 5 of these having the same weight loss (so that is everyone up to Bee on the excel spreadsheet). I am also omitting the 4:3 people from the top group (so that takes us to Kate_in_Kemble as the lowest of the top group to give us 25 users). Do you agree?

I've tweaked the questionnaire to add a Q on alcohol and on diet/sugary drinks. Any ideas on the best way of administering the questionnaire? I was wondering if there is a way of making it available to be filled in anonymously online by the selected users and uploaded to where we can crunch the numbers. I have written it in Word but should it be put into excel or something to make the data easier to manipulate? I'm out of my depth when it comes to that kind of thing.

Student's t-test built into spreadsheet package.

The free surveymonkey doesn't give us an excel download. Let me think...

Edit: Google docs has a Forms feature, starting from a spreadsheet.

[carorees]

Hi Phil

Unfortunately because the data is skewed I think the t test is not the most appropriate...perhaps should be Mann-Whitney? I'm not too brilliant with stats so may be wrong, but anyway, it looks hopeful even bearing in mind that the t test is not quite right for the data.

I have no idea about Google docs etc...would you be willing to take on the job? Or shall I just PM each user and attach the questionnaire. I could then construct a spreadsheet and strip out the user names as I input the answers to the questions. Then, once anonymised, it could be opened for statistical analysis by those who can do that kind of thing....you?? Who??

if you have it in Excel I'l have a go at it on Google

I think the data should be stuck somewhere (Google again) for anyone to have a hack at it.

The t-test says they're two populations, as the mean is statistically different, so does the skew matter ? The probability was 10E-7 or something

[carorees]

Well the questionnaire is in Word, so should I forward that to the prospective respondents and then input their responses into excel or try to convert it into excel? I'm not all that familiar with excel so am not sure how to convert the questionnaire into a document that is easy for people to read and fill in.

I agree that the responses data should be held in the cloud for anyone to manipulate (once it has been anonymised).

My stats understanding is too basic to comment on anything other than I had heard that Mann-Whitney should be used for data that is not normally distributed. As far as analysing the output from the questionnaire we need to think about the best way of comparing the two groups considering the sample size is still fairly small.

Mann Whitney says P<0.0001 too. Ua = 625 and z= -6.05

given we took out the middle of the range it's not a surprise.

If the data is going to be numbers or logicals (yes/no) then it's better in a spreadsheet. If there are going to be a lot of words then a document is better.

If you put it somewhere I'll take a look.

The t-test says they're two populations, as the mean is statistically different, so does the skew matter ? The probability was 10E-7 or something

Yes, in a sample this small it can matter.

As far as analysing the output from the questionnaire we need to think about the best way of comparing the two groups considering the sample size is still fairly small.

Logistic regression would be ideal, since it allows you to combine several explanatory variables, but the sample is quite small - in multiple linear regression you have a rule of thumb of n=10 per variable, more if the data is skewed.
Combining variables would be of interest, but at the moment I can't think of a good way to do it. Maybe someone else on the forum does logistic regression.

So what you would be doing is looking at one explanatory variable at a time, using t-test or MWU (depending on the skewness of the variable) to compare the two groups (the Biggest vs the Baddest Losers)? It is a good place to start anyway, if you find something interesting you can always try to explore that further.

If possible, aim for binary coding of the variables, otherwise you get even smaller subgroups and lose statistical power (use Kruskal-Wallis in that case).

[carorees]

I've split out the nerdy stats stuff from the original thread about the monthly weigh in results so we can focus on the future analysis of the questionnaire results. I think we need to discuss the analysis before finalizing the questionnaire and sending it out.

To this end can all who wish to contribute to questionnaire and stats methodology please check out the proposed questionnaire: http://dl.dropbox.com/u/32116230/52fast ... naire.docx
and the excel sheet I've created to input the results: http://dl.dropbox.com/u/32116230/52fast ... onses.xlsx

If we need to reduce the variables to two options we would need to group the responses where I have currently given several options. Do we think it is best to leave the questionnaire with several answers being possible to some questions and then group them afterwards or limit the responses further?

I'm not particularly adept with excel so I may have created a cumbersome beast for data entry that one of you could improve on...please feel free to criticize.

And here's the link to the monthly weigh-in results for ease of access: http://dl.dropbox.com/u/32116230/52fast ... 32013.xlsx

If possible, aim for binary coding of the variables, otherwise you get even smaller subgroups and lose statistical power (use Kruskal-Wallis in that case).

Are we not falling at the first fence by picking off two groups - upper and lower quartile - when Kruskal-Wallis (which I am just reading about) seems to like you to have a random selection of subjects from the population ?

ETA: I suppose where we're heading is the ability to say something like..

"80% of the biggest loser group didn't count calories on feed days, compared to 30% of the smallest losers. Statistically significant"

whereas if we wanted to correlate weight loss against say height we'll have two results as we have two groups. We might then find that weight loss is related to height in the biggest losers but not relevant in the smallest losers"

[carorees]

If we can go for binary coding and MWU at this stage, though are we more likely to spot a significant difference?

[carorees]

So thinking some more about this, should we send the questionnaire to everyone who indicated that they would be willing to participate (and even tout for more) and then do the Kruskal-Wallis test (or possibly Chi-squared test for independence?) for each variable?

Urrgh...statistics puts my head in a spin...please someone just tell me what to do with the questionnaire and offer to do the statistics!!! *buries head in sand*

If it's any consolation, I'm supposed to be fairly bright and I can't make head nor tail of stats...

You're being very brave!

To this end can all who wish to contribute to questionnaire and stats methodology please check out the proposed questionnaire: http://dl.dropbox.com/u/32116230/52fast ... naire.docx

Hi,
I've had a quick look at the questionnaire. I think it is generally a good questionnaire. it includes everything I can think of. Thank you for putting it together. I've made a few suggestions regarding the format of the items. If you're interested, where would you like me to send the reviewed file?

Generally, in questionnaires and data analysis, when there is perfectly a continuous variable such as age, time on diet etc, it is best to keep it that way, especially during the survey. Mostly because it is very easy to recode a continuous variable into a categorical variable. Similarly, taking arbitrary cut of score or number of participants is not recommended. Instead it is best to benefit from a continuous variable such as weight loss. Logistic regression would be very helpful to investigate the differences between higher and lower weight-loss groups but in this case I don't think it is necessary. A simple multiple regression would do just fine to understand what predicts higher weight loss. But then, as Coffeecat mentions the sample size vs number of variables in the analysis will have to be taken into consideration.

Running a series of individual mean comparisons(t-tests, anovas etc...) on different variables between any groups increases the error rate (i.e. p values) on aggregate and results in misleading findings and false positives and thereby is not very useful.

thanks for reading.
best,
e.