Achieving precise and accurate estimation of carbon (C) in forest soils is a challenge especially because of high natural variability in the forest floor. Although most researchers have used a simple random sampling (SRS) design for within-plot soil sampling, double sampling for stratification (DSS) can be used to decrease costs, increase precision and increase power. We compared estimates of C and N stocks based on DSS to those estimated by SRS, in the humus forms (O plus A horizons) of Douglas-fir (Pseudotsuga menziesii (Mirb.) Franco) stands in the Cascade Mountains in Oregon, USA. The DSS was 5 to 66 times more efficient than SRS for estimating soil C stocks. Coefficients of variation estimated from DSS were about one quarter of those estimated by SRS and reported elsewhere in the literature. To achieve the same level of precision with 6 to 16 samples per plot using DSS, 43 to 863 samples per plot for would be required using SRS. The DSS design is more powerful than SRS and therefore permits a scientist to detect smaller changes than SRS with the same number of samples. The estimated costs of sampling for this study using DSS were about one-twentieth of those if SRS had been used to obtain the same precision. These results show that large gains in efficiency were realized by using a more complex within-plot sampling design, i.e. double sampling for stratification, compared with simple random sampling.