The difference in regression sums of squares, however, cannot be used to
compute Q in the second example where β_0 = 20 is included in the null
hypothesis. In this case, SS(Total) for the reduced model is based on Y_i^*
and hence it is different from SS(Total) for the full model. Consequently, it
is important to develop the habit of either always using the residual sums of
squares, since that procedure always gives the correct answer, or being very
cautious in the use of differences in regression sums of squares to compute
Q.