Calculate your sample size needed. Sample size calculator are available

http://www.calculator.net/sample-size-calculator.html

Good luck

Béatrice

Calculate your sample size needed. Sample size calculator are available

http://www.calculator.net/sample-size-calculator.html

Good luck

Béatrice

Hi Joanna,

whether your results verify your hypotheses or not says nothing about the actual test power of your study. You might have just been lucky with your current observation. It is better to make a plan how much participants needed in order to have a decent test power BEFORE running the experimental sessions.

Well, since test power depends on several factors (you should read about it!), it is difficult to give an accurate answer to your question. An important question is: Is it a one-shot experiment or a repeated game with feedback after rounds? In the former case, the individual is the independent observation and you can easily use some of the standard power calculation programs as suggested by Beatrice. If you have a (nested) repeated design, the group becomes the independent observation and you need to consider this when doing the power calculation (if you only make analyses on this level, for instance with non-parametric comparisons. If you analyse the impact of feedback on individual behavior in a repeated public goods, power calculations are quite complicated because you need to address the nested data structure with clustured standard error regression or mixed effects models. But in this case you've certainly not enough participants yet...

Not sure whether this was really helpful.

Best,

Robert

Thank you Béatrice! It is a very good calculator and I will share it with my friends!

Dear Robert,

Thank you for your idea, they are very helpful to me! My experiment is unluckily a nested repeated design, and I am trying to use HLM to deal with the data.  I see what you mean and it really makes sense, so I will try to collecte more data in the next few days. Thank you again!

Joanna

Hi Joanna,

Unless your study involves within-subjects manipulation or a very strong effect, your sample sizes are too small. I suggest you read a bit about power and sample size planning as already suggested. I good way to start and a real eye-opener is the presentation from Simmons that I linked here. It suggests to have at least 50 participants per condition unless  you think you measure a very strong effect (stronger than "men weigh more than women"). 

Now, you are talking about experimental groups and control groups. Does that mean that you have different experimental and control conditions, or does it simply mean you sample per group of particpants? And do you measure the outcome at the group level or at the individual level? If you measure it at the group level, than you should focus on sampling enough groups, not on sampling individuals.

Good luck with your research,

Tim

Forgot the link

http://papers.ssrn.com/sol3/papers.cfm?abstract_id=2205186

Hi Tim! That is really useful advice to me. My study involves between-subjects manipulation so I have two conditions as experimental and control conditions. As you pointed out, I think I should of course add more samples. As far as both group level and individual level are concerned, I feel it is really a very big work for me to finish.

Thank you a lot for the link, I am sure I can learn a lot from that, please keep in touch!

Joanna

Hi Joanna,

you may also want to consider sequential analyses. This is very time and resource efficient, because you can decide based on the data you have so far whether it is necessary, or even worth it to continue collecting more participants. You do of course have to correct for Type 1 errors. Here is a very helpful article that came out recently:

Lakens, D. (2013). On the benefits of adaptive designs and sequential analyses for psychological science. Available at SSRN 2333729.

Best, Anouk

Hi Anouk !

Thank you for the sharing, I have heard of sequential analyses but never used it. Your idea means a lot for me, thanks! I will read it as soon as possible.

Joanna

I found a software to test sample size needed for an experiment, it seems very useful.

G*Power: Statistical Power Analyses for Windows and Mac

http://www.gpower.hhu.de/

G*Power is a tool to compute statistical power analyses for many different t tests, F tests, χ2 tests, z tests and some exact tests. G*Power can also be used to compute effect sizes and to display graphically the results of power analyses.

Others have shared excellent, insightful ideas already. But, I'd just note that the exact minimum sample size acceptable for research varies quite a bit depending on which statistics reference you consult and by specific sub-field. For example, in brain-imaging research, which is expensive and can be difficult to obtain, sample sizes below 10 are fairly common. In contrast, some social psychology research involves sample sizes well into the 100's and occasionally 1,000's.

From the various statistics texts I've used over the years, authors suggest a minimum of 15 or around 30 from what I've read. So, I've used "n > 30" as a rule of thumb over the years. Of course, power analyses is always preferential.

One more thought - the minimum sample size needed for good data analysis is also dependent on the statistical procedures you'll employ. For example, many of the non-parametric inferential statistics will work quite fine with fairly small samples. Procedures involving modeling (e.g. SEM), however, benefit from large samples.

Hello,

I got a question too. When I am using this calculator for the sample size does the calculated sample size stand for the whole sample? Or just one group?

For example Joanna said she got 8 groups. Does she simply need to split the sample? Or do you think one has to consider that the more groups the higher n?

Hi, Jay! That's a very good conclusion and your ideas worth a lot for me. Thank you!

Hi, Carmen! I don't get your point very well. I think maybe my statement about groups is misleading, so let me explain my groups first. A group has 4 subjects and they will play a social dilemma game together with their group members. So for my experimental treatment, I got 8 groups' data for analysis. For my control treatment, I got 7 groups' data for analysis. 

Am I make it clear? Hope my terrible English does not make you misunderstand something. :-D

I have changed the statement in my question either. 

Joanna

Hallo Joanna,

maybe I thought in a different direction. I was thinking about experiments in which I got a group and a controll group. Let's say the calculator sais I need 100 probands. Do I split those 100 probands into two groups? Or do I double the size according to design?

Never mind about your english. Mine is terrible too as you can see :S

Hi Carmen! I think  it depends on whether your experimental design is between-subjects or within-subjects. If yours is between-subjects design, I think it is better to double it or at least add more sample to each group. If yours is within-subjects, there is no need to split because all the 100 subjects will go through both your experiment group and control group.

As you see, I am also a freshmen in the field of sample size. So I don't have faith in this answer:-D Hope other followers can give a more powerful answer.

As Jay points out old textbooks used non-parametric methods - when the sample size was   n< 30 ;  when sample size was n= or > 30 the sample was considered normally distributed and parametric tests were applied.

Hello Joanna, today I asked our stat teacher at university. And she suggested G*Power too. She told me that with G*Power it is possible to calculate the group size of one or more groups. As far as I understood she said one does not simply double the size (but I could have misunderstood her here). So she suggested G*Power and fill in how many groups one wants to use. She also said a huge n-size can be a problem as it manipulates the power function in such a way that H0 could be discarded.

Hi Carmen! Sorry to reply it so late. Thank you for sharing! Then I can have less groups than expected. Haha! 

Just to follow up on Carmen's comment, I don't see how a huge n-size is ever a problem. it's all a matter of interpretation. Of course, if you would test thousands of participants you have so much power that every little (and possibly uninteresting) effect would turn out to be significant. This is why it's generally much more interesting to look at effect sizes and confidence intervals rather than p-values. I can highly recommend Geoff Cumming's book 'The New Statistics' if you want to know more about testing and interpreting effect sizes and confidence intervals. 

Sounds good, Anouk. Googeling the web I found an article dealing with what Anouk is talking about. This may be of interest for you too, Joanna:

https://www.apa.org/education/ce/confidence-intervals.pdf