Construction contract auctions are characterised by (1) a heavy emphasis on
the lowest bid, as that is which usually determines the winner of the auct
ion, (2) anticipated high outliers due to the presence of uncompetitive bid
s, (3) very small samples, and (4) uncertainty of the appropriate underlyin
g density function model of the bids. This paper describes a graphical meth
od for simultaneously identifying outliers and density functions by first r
emoving candidate (high) outliers and then examining the goodness-of-lit of
the resulting reduced samples by comparing the reduced sample predictabili
ty (by the expected value of the lowest order statistic) of the lowest bid
with that of the equivalent predictability by Monte Carlo simulations of on
e of the common density functions. When applied to a set of 1073 auctions,
the results indicate the appropriateness of censored and reduced sample log
normal models for a wide range of cut-off values. These are compared with c
ut-off values used in practice and to identify potential improvements.