Populations and subsamples are different things; there is a theoretical population based on the model itself (the “error” distribution) and the model is fit using all the data, not just observations around Trump’s measurement. The number of people around Trump’s measurement depends on what you mean by “around” and is not too relevant to the 84% number (but the graph shows it, and the data are available). The 84% is of a theoretical population estimated using all data.