Pollster Trend Model with House Effects

141 Likely Voter Polls from Pollster (7/1-10/25) 
Dependent variable: Obama Lead over Romney
Ordinary Least Squares

                 coefficient   std. error   t-ratio   p-value 
  ------------------------------------------------------------
  const           5.58618      0.59754       9.349    4.21e-16 ***
  Debate1        −5.58195      1.19655      −4.665    7.76e-06 ***

  Trends 
  Pre-Debate     −0.05086      0.00956      −5.321    4.56e-07 ***
  Post-Debate    −0.00685      0.04817      −0.142    0.8871  

  House Effects
  Rasmussen      −2.24266      0.46214      −4.853    3.53e-06 ***
  Gallup         −4.20741      1.20723      −3.485    0.0007   ***
  ReasonRupe      3.95669      2.02626       1.953    0.0531   *
  AllstateNJ      3.85498      2.02674       1.902    0.0594   *
  UTechNJ        −3.70445      2.03258      −1.823    0.0707   *
  NPR             2.82984      1.46391       1.933    0.0555   *
  DemCorps        2.26101      1.05003       2.153    0.0332   **
  WashTimesJZA    2.05687      1.04592       1.967    0.0514   *
  ARG            −1.82185      0.85877      −2.121    0.0358   **

  Methods
  NetNotIpsos     2.67967      0.80298       3.337    0.0011   ***
  IpsosReuters    0.63428      0.65526       0.968    0.3349

Mean dependent var   1.120567   S.D. dependent var   3.078671
Sum squared resid    506.0489   S.E. of regression   2.004061
R-squared            0.618638   Adjusted R-squared   0.576264
F(14, 126)           14.59960   P-value(F)           3.00e-20

Results for polls with Rasmussen excluded (N=103)

                 coefficient   std. error   t-ratio   p-value 
  ------------------------------------------------------------
  const           6.21481      0.789507      7.872    7.96e-12 ***
  Debate1        −5.58066      1.33422      −4.183    6.73e-05 ***

  Trends
  DaysBefore     −0.0606532    0.0140282    −4.324    3.99e-05 ***
  DayB4*Deb1      0.0141113    0.0520640     0.2710   0.7870  

  House Effects
  Gallup         −4.08319      1.16342      −3.510    0.0007   ***
  UTechNJ        −3.97064      1.95781      −2.028    0.0455   **
  ReasonRupe      3.81785      1.94636       1.962    0.0529   *
  AllstateNJ      3.69654      1.94702       1.899    0.0609   *
  NPR             2.95549      1.43444       2.060    0.0423   **
  DemCorps        2.35408      1.02893       2.288    0.0245   **
  WashTimesJZA    2.28335      1.01499       2.250    0.0269   **
  ARG            −1.76455      0.825079     −2.139    0.0352   **

  Methods
  NetNotIpsos     2.75094      0.772406      3.562    0.0006   ***
  IpsosReuters    0.611675     0.629106      0.9723   0.3335  

Mean dependent var   1.883495   S.D. dependent var   3.000983
Sum squared resid    329.3809   S.E. of regression   1.923775
R-squared            0.641432   Adjusted R-squared   0.589057
F(13, 89)            12.24691   P-value(F)           8.35e-15

The method for measuring pre- and post-debate trends differs in the two models. For polling with Rasmussen included, I created separate pre- and post-debate trends and tested whether each significantly differs from zero. In the model without Rasmussen, I measured whether the post-debate trend differed from the pre-debate trend by including an overall trend measure and an interaction term that tests whether the post-debate slope differs from the slope estimated using all the observations. Mathematically the tests are identical, but the estimated coefficients tell different stories. The first model shows that the trend before the Denver debate was significant but shows no trend since then. The second model shows that, in the sample without Rasmussen polls, the remaining pollsters all report results that follow the same common trend line. There is no difference between pre-debate and post-debate trends in the sample without Rasmussen polls.