Hi Erika,

If you have three populations, can you analyze two at a time? That should give you three different combinations for example (1, 2); (1, 3); (2, 3) and then report the difference between either two.

Here is a technical manual I think will be of your interest. Please look at the part "COMPARISON OF KAPLAN-MEIER ESTIMATES"

http://www.ncbi.nlm.nih.gov/pmc/articles/PMC3932959/

Best regards,

Richard

Do pairwise posthoc tests similar to what you do moving from significant anova to posthoc t-tests. Cheerio

Hi Erika, 

You are precisely correct. When you have more than two groups, you know that are difference exists, but you don't know where. Possibly all combinations are different, or just one pair. You can't know until you test each combination. But visual inspection of the survival curves will often give you a hint as to where to start. Good luck!

If you don't have a precise hypothesis, you can't test one. Does your hypothesis just say that the tree groups are different, or does it make a more specific, testable statement? Without knowing your hypothesis, people recommending tests are guessing. And they are just as likely to be wrong as right.

Erika -

There are tests that look at multiple populations, but as Juehui indicated, it may be more meaningful to look at them in pairs. However, as Ronan said, you need specific hypotheses to address/identify your problem, to even know what it is that you are really trying to learn from your data. You might, for example, want to compare mean survival times, or you might consider another comparison based on order statistics.

The major problem I see with seeking p-values to examine an hypothesis is that they are generally misused. A p-value is a function of sample size. If your sample sizes are large, then a very nearly 'correct' hypothesis will tend to be 'rejected' ('fail to accept'), unless you adjust what you define as an acceptable level accordingly. If a sample size is very small then an hypothesis that is very far from true will be 'accepted' ('fail to reject') unless the level is adjusted in the other direction. (Often 0.05 is used on all cases, and that is not a good idea.) How far to adjust? That depends on the power, which is often used to pick tests rather than analyze the data (though there is a link on my RG page to where I've used it for the latter before). It is all rather nebulous in most applications. (See link below.) That is why I generally recommend confidence intervals over hypothesis tests, even if you have to use the Chebyshev inequality.

If you were to look at, say, estimated mean survival times, for each of the three populations, and estimate a standard error for each, that would be useful. If you compared them two at a time, you could look at confidence intervals around the differences between the estimated means.

At any rate, whatever aspect(s) you chose to examine, it is not really a matter of Are these populations different?, but rather How different are they in ways that might matter to my study?

If you are using JMP, then I suggest some graphical representations of these three populations. (Actually, I recommend graphics, regardless.) Here you could put all three on the same graph. I am a firm believer in the usefulness of graphical displays. If you cannot get what you want to show in JMP, you might still do some fairly decent graphics on a spreadsheet such as Excel, or some other software.

Cheers - Jim

Article Practical Interpretation of Hypothesis Tests - letter to the...

I agree with Jim, focusing on your research objective with meaningful hypothesis tests.

What's your hypothesis? If you haven't got one, then don't bother testing groups against each other. There is otherwise the danger that you will start developing a hypothesis based on your data, rather than testing your hypothesis..

They are called hypothesis tests because you start with a hypothesis, then test it. 

You can easily test statistical differences between three populations using Kaplan-Meier curve or with Cox regression in SigmaPlot or Systat if you have access to those statistical software

Please refer the attached file as an additional information

Article Radiomodulatory role of Rutin and Quercetin in Swiss Albino ...

Here is how to do it on SPSS.

Run the KM analysis, and under "compare factors" select "pairwise over strata". This generates a Chi-sq table with multiple comparisons (kind of like a post-hoc). You might need to use Bonferroni correct the p-values though (not sure about this part).