MINIMUM SAMPLE SIZE: In non-parametric case, it has been suggested that the minimum sample size is n = 30 as a rule of thumb. See link below:

http://www.reconstrue.co.nz/IPCT-J%20Vol%206%20Robin%20hill%20SampleSize.pdf

Anderson & Darling, for instance, suggests that for purposes of distribution type verification, the minimum sample size is n > 5 or at least 6.

TESTING THE POPULATION: If resources permit, one might want to sample the population by taking a test sample for several test sample. Try to determine the minimum sample by:

(1)   nmin = (Z2σ2) / E2

... where Z = standard score; σ = estimated standard deviation of the unknown population; and E = σ / ntest.

KNOWN POPULATION: If the population is known, one can use the Yamane method to determine minimum sample size:

(2)   nY = N / (1 + n(e2)

... where N = population size; e = error level which is equal to 1 - CI. If using 95% confidence interval then 1 - 0.95 = 0.05 error level. This method tends to produce quite large minimum sample size. If you are doing sample size estimation prior to data collection, this approach may be appropriate if the population size is known.

PARAMETRIC: in case of structural equation modeling, Nunally suggests that each parameter should have at least 10 observations, i.e. GDP equation has 4 parameter: GDP = C + I + G + T^ would mean 4(10) = 40 with 10 counts for each parameter.

REFERENCES: I attached 2 articles here that may be helpful. Agresti article suggests n = 30 in order to see full manifest of Central Limit theorem (CLT). The Gou article tackle's the issue from extreme value perspective (Weibull distribution). 

Other relevant articles:

[1] Westland, J. Christopher (2010). "Lower bounds on sample size in structural equation modeling". Electron. Comm. Res. Appl. 9 (6): 476–487.

[2] Nunnally, J. C. (1967). "Psychometric Theory". McGraw-Hill, New York: 355.

[3] Yamane, Taro. 1967. Statistics: An Introductory Analysis, 2nd Ed., New York: Harper and Row.

[4] Stephens, M. A. (1974). "EDF Statistics for Goodness of Fit and Some Comparisons". Journal of the American Statistical Association 69: 730–737.

[5] Anderson, T.W. and Darling, D.A. (1954). "A Test of Goodness-of-Fit". Journal of the American Statistical Association 49: 765–769.

Thank you very much.

In the French school of prospective, it is to identify the key variables indicative of trends or disruptive (technology, political environment, social....) . Normally no more than 20 variables and after collecting your data in relation to these key variables of evolution