It is known that direct-survey estimators of small area parameters, calculated with the data from the given small area, often present large mean squared errors because of small sample sizes in the small areas. Model–based estimators borrow strength from other related areas to avoid this problem. How small should domain sample sizes be to recommend the use of model-based estimators? How robust small area estimators are with respect to the rate sample size/number of domains? To give answers or recommendations about the questions above, a Monte Carlo simulation experiment is carried out. In this simulation study, model-based estimators for small areas are compared with some standard design-based estimators. The simulation study starts with the construction of an artificial population data file, imitating a census file of an Statistical Office. A stratified random design is used to draw samples from the artificial population. Small area estimators of the mean of a continuous variable are calculated for all small areas and compared by using different performance measures. The evolution of this performance measures is studied when increasing the number of small areas, which means to decrease their sizes.