We present techniques for bridging the gap between idealized inverse covariance weighted quadratic estimation of 21 cm power spectra and the real-world challenges presented universally by interferometric observation. By carefully evaluating various estimators and adapting our techniques for large but incomplete data sets, we develop a robust power spectrum estimation framework that preserves the so-called EoR window and keeps track of estimator errors and covariances. We apply our method to observations from the 32-tile prototype of the Murchinson Widefield Array to demonstrate the importance of a judicious analysis technique. Lastly, we apply our method to investigate the dependence of the clean EoR window on frequency--especially the frequency dependence of the so-called wedge feature--and establish upper limits on the power spectrum from z = 6.2 to z = 11.7. Our lowest limit is Delta(k) < 0.3 Kelvin at 95% confidence at a comoving scale k = 0.046 Mpc^-1 and z = 9.5.