When would you want to reduce variance? In a sampling-estimation context, low variance of the estimate is a goal.  Several things I've read suggest (though I can't quite connect the dots) that lowering variance in the data will improve the estimate and that one should design the sampling structure toward that goal.  That variance there is the one among the sampling units rather than in their constituent populations (say plants on plots)?
On the other hand, in a regression context, I have read that variance in the data, at least in the independent variables, contributes something to the quality of the model and that smoothing, at least via aggregation, is something to be avoided.  So if one were designing an experiment using regression, higher between-unit variance would be no problem or even beneficial?  
Two different strategies, two different approaches to variance in the data?   
 A: Paradoxically, a large spread in the independent variables makes a regression slope easier to estimate. Here's why. 
Ordinary least squares regression (OLS) assumes constant error variability, so I'll assume that in my explanation. Given the squared errors are on average $\sigma^2$, it will be very difficult to detect a difference of $0.01\sigma$ units between two different conditions. It will be very easy to detect a difference of $100\sigma$ units. Imagine the true underlying slope is 1/10. Then, x-values separated by a distance of $0.1\sigma$ will differ in their responses by $1/10 \times 0.1\sigma = 0.01\sigma$. You could easily mistake that by a factor of 2 or even end up with a negative number. When you went to estimate the slope, it would be off by the same factor (or sign flip).
If the x values had been placed farther apart, at an interval of length $1000\sigma$, then the gap in responses will be about $1/10 \times 1000\sigma = 100\sigma$. You might estimate it at 99.5 or 100.5 if you only have a few samples. The slope estimate ends up erring only by half a percent of its true value.
The second setup, with lots of variability in the independent variable, makes it much easier to peg the coefficient.
