Abstract : A robust confidence interval using biweights for the case of five observations is proposed when the underlying distribution has somewhat heavier tails than the Gaussian. The distribution of a t-like statistic is approximated by a Student's t on the nominal four degrees of freedom using different scale factors which depend upon the value of the biweight weights. Results given by Monte Carlo simulations indicate that, even for very high coverage probabilities, the intervals proposed are highly efficient, in terms of the expected length of the confidence interval. (Author)