Light detection and ranging (LiDAR) is commonly used to predict forest inventory attributes of interest across large regions. Most studies utilize model-derived estimators whose performances are affected by training data, and give less attention to using design-derived estimators. The influence of sample design and estimation method is an important consideration for determining sample sizes or calibration plot densities; however, this has not been systematically explored, particularly in mixed-species forests. In this study, 10 sample selection methods (four equal probability and six variable probability selection methods) and six estimation techniques (two model-derived and four sample-derived estimators) across a range of sample sizes are evaluated using LiDAR-derived predictions of volume per ha. Results show that the use of variable probability selection methods combined with sample-derived estimation techniques are more efficient than using model-derived estimates. Estimation technique had a greater effect on sample efficiency than did selection method, though specific combinations were more efficient than others. For example, random forest imputation was the most efficient at the lowest sample sizes (n < 50); however, significant biases were obtained when used with variable probability selection methods. The required plot densities across the combinations of selection methods and estimation techniques used in this study ranged from one plot per 15.7–32.6 ha. Use of a variable probability selection method based on attributes derived directly from LiDAR point clouds coupled with a ratio or regression estimator was a very efficient LiDAR-assisted sampling design that should be considered more in the future.