We investigate how sampling of parents or children based on their extreme phenotypic values selected from clinical databases would affect the power of identification of quantitative trait loci (QTL) by a transmission disequilibrium test (TDT). We consider three selective sampling schemes based on the selection of phenotypic values of parents or children in nuclear families: (1) two children, one of extreme value, the other random; (2) two children extremely discordant; (3) one parent of extreme value. Other family members not specified will be recruited randomly with regard to phenotypic values. Our study shows that the second sampling scheme can always enhance the power for QTL identification, sometimes dramatically so. The increase in the statistical power of the TDT is particularly dramatic when h2 at the QTL under test is small or intermediate (e.g. 0.05 or 0.10). For the other two sampling schemes, under dominant effects at the QTL, the power is always increased relative to random sampling; however, under recessive or additive genetic effects, the power gain is generally minor or even decreased a little sometimes. Allele frequencies at the QTL and the selection stringency are important for determining the effect of selective sampling on the power of QTL identification. Our study is useful as a practical guideline on how to perform the TDT efficiently in practice by taking advantage of the extensive databases accumulated that are enriched with people of extreme phenotypic values.