Why are observationally-based estimates of the cloud feedback more uncertain than model-based estimates?

The experimental data is sparse and hence the result has a lot of uncertainty. Dessler did the best he could with the available data. With more years of data, the error bars might shrink and observation will provide a better test of the models. If the error bars in 50 years are comparable to the model error bars, then the models include all of the major inputs. If the error bars remain large, then nature is more complex than the models.