There are estimation models that take a large number of inputs, and others that rely on only a few inputs. Some of the input data are more objective (e.g., the number of screens in a spec) while other data are more subjective (e.g., a human-assigned "complexity" rating for each of several reports).
One thing that Steve explains is that, when many subjective inputs are present, models with fewer inputs often generate better estimates than models with "lots of knobs to turn."
The reason for this is that human estimators are burdened with cognitive biases that are ridiculously hard to break free of. The more knobs the estimator has to turn, the more opportunities there are for cognitive biases to influence to result, and the worse the overall performance.
It should be reiterated that this same problem does not apply to systems that take many objective measures (such as statistics or budget data from historical projects) as inputs.