Statistical analysis is just normalized measurement. A measurement is always modified by calibration (effects accuracy) and sampling (effects precision). See “Relational Measurements and Uncertainty” (in Measurement Nov 2016) for a formal development. The statistical tolerances, as you explain very clearly, define the sampling effects (precision) and the other errors cases you note are where the polling response is not calibrated to the actual voting response, this is the accuracy. The vector sum of precision and accuracy (which cannot be zero, see reference) equals the indeterminacy of a measurement.