Hello Thyago,

You raise a good point. Usual parametric tests assert that error terms are independent of one another, which most likely will not be the case in snowball sampling. That is, if you have some quantitative outcome, the values would most likely show non-zero correlation / covariance across cases.

In regression software, the Durbin-Watson test (if available) is a check on case independence (absence of autocorrelation/autocovariance).

The result is that any analytic method would have to take into account this lack of independence in order to yield suitable estimates of both effect and error risk. ARIMA or multilevel models might work. The principal concern I would have is whether the snowball sample included sufficient cases to make such methods feasible.

Good luck with your work.

Great, thank's for response.

A practical answer must really depend on how much detail of the sampling is retained in the data records that eventually exist, but also on the overall planning of the study.

There are two obvious possibilities:

(1) on the assumption that the data-records enable you to group together that originating from the same seed, and that there are enough such independent groups, you can use this structure in the analysis, possibly by employing a block-jackknife technique on some sensible test statistic.

(2) design the study so that you get statistically independent groups of subjects, but using snowball sampling within each group, designed so as to get effecting the same sample-size in each group, so that you can assume an equal variance for test-statistics evaluated for each group. You could then proceed with the usual procedure for independent-equal-variance data on these test-statistics .

thanks for your answer