Experimentally
The variation in results that are caused by confounding factors can be eliminated by making the confounding factors the same for all test conditions. For example, the same soil is used for all plants, all plants are exposed to the same schedule of light, herbivores are excluded, and the same type of plant is tested. Note that, since you have not varied those factors, your results strictly apply only to the set of conditions under which the plants were grown.

Statistically
Knowing that there are other relevant factors, you record them as you test for the effect of water. For example, you water plants in the wild and don't water an equal number, making sure that the water you provide is the only water each plant receives during the experiment. In addition, you measure the amount of light each receives, you test the soil in which each is growing, you measure herbivory on each, and you record the species of plant. Naturally, the number of plants tested would have to be quite large to provide sufficient numbers for each factor alone and in combination.

Experimentally and Statistically
It is possible to expand your experiment to apply more than one level of the other relevant factors and measure the response of the plants under variations of all factors. Statistics are used to separate the individual effects of each factor and any effects of their combinations so that the main effect of water can be calculated. An experiment that includes the simultaneous effects of more than one factor is called a factorial experiment. Obviously, the number of plants tested would be greater than those needed when using only one level of the other relevant factors, but you would gain insight into plant response when other factors change.

Number of Subjects
Statistical comparisons often compare the average group response relative to variability within the groups. Even when averages are not the same, a large amount of variabilty may prevent you from concluding that the difference in the averages was great enough to show that the treatments had the effect you were trying to prove.

When all test subjects are carefully picked to conform to the same characteristics, then the variability in response to the treatment from subjects should be very low and few subjects would be needed. On the other hand, you generally have much less control over many factors when picking volunteers to participate in an experiment, and responses may show variability from the effects of many factors you were unable to control. Even if you used a questionnaire to impose uniformity or by excluding certain categories altogether, you would need a relatively large number of subjects to reduce the variabilty in the statistics so that a difference between treatments would be statistically significant.