Parallel regression assumption survey data

[Oskar Lindgren]

I have a quite large survey data set (N=12,053) and trying to check the parallel regression assumption of my tentative model. However, it seems that the brant test always generates significant results, even when removing independent variables from the model, especially for the full model. Does Brant's test usually report significant results for larger datasets (mine is not huge but you get the point). Should I just relax the assumption and go with gologit2?
My depvar is categorical (five levels). Here's the output (variable names removed) generated when running Brant test:

 Brant test of parallel regression assumption

                 |       chi2     p>chi2      df
 ---------------+------------------------------
            All |     364.68      0.000      48
 ---------------+------------------------------
  Var        |      32.71      0.000       3
  Var        |      24.31      0.000       3
  Var        |      17.36      0.001       3
  Var        |       4.63       0.201       3
  Var        |      13.81      0.003       3
  Var        |      25.79      0.000       3
  Var        |      40.68      0.000       3
  Var        |      32.77      0.000       3
  Var        |       4.09       0.252       3
  Var        |      13.66      0.003       3
  Var        |       6.98       0.072       3
  Var        |       9.95       0.019       3
  Var        |       6.04       0.110       3
  Var        |      11.61      0.009       3
  Var        |       8.56       0.036       3
  Var        |      14.23      0.003       3

Thanks,
Oskar

[Maarten Buis]

A statistical test is better at detecting deviations from the bull hypothesis as the sample becomes bigger. This is intentional and correct: if you look harder you are more likely to find something. However, it also means that statistical tests become basically meaningless in large samples. The null hypothesis is always (a bit) false, so in large samples you will always detect these small and meaningless deviations. What you should do is add the detail option to brant and look at coefficients and see if they are close enough to be the same to justify the assumption that they are the same. This is subjective, but all model choices are at their core subjective.

[Oskar Lindgren]

Thanks Maarten Buis, that's what I figured!