var_model test_causality vs test_inst_causality redundant code?

[jbrockmendel]

The construction of the restriction matrix is done differently in these two methods, but I suspect they end up equivalnet

[josef-pkt]

@jbrockmendel Add the name of the module, so we don't have to guess.

I never looked at test_inst_causality.
I think those two hypothesis tests are very different.
Test_causality is a Wald test for zero restriction on lag parameters, the instantaneous is some (?) test on the structure of the contemporaneous correlation matrix (I guess).

[jbrockmendel]

I'll have to take a closer look then to be sure. Strong prior for code being duplicated based on a) its in var_model and b) test_inst_causality came in with vecm.

[josef-pkt]

check some references, it's not easy to see from the code what it is supposed to do.

I'm pretty sure that it is a test for the covariance/correlation structure (like a off-diagonal block of zeros)
(maybe similar to #4143 or LR based tests for correlation structures)

[jbrockmendel]

it's not easy to see from the code what it is supposed to do.

You're right, that does sound like a problem in and of itself.

Lines 1760-1770 look analogous to 1633-1653. Then:

       k, p = self.neqs, self.k_ar

        # number of restrictions
        num_restr = len(causing) * len(caused) * p
        num_det_terms = self.k_trend

        # Make restriction matrix
        C = np.zeros((num_restr, k * num_det_terms + k**2 * p), dtype=float)
        cols_det = k * num_det_terms
        row = 0
        for j in range(p):
            for ing_ind in causing_ind:
                for ed_ind in caused_ind:
                    C[row, cols_det + ed_ind + k * ing_ind + k**2 * j] = 1
                    row += 1

vs

        k, t, p = self.neqs, self.nobs, self.k_ar

        num_restr = len(causing) * len(caused)  # called N in Lutkepohl

        sigma_u = self.sigma_u
        vech_sigma_u = util.vech(sigma_u)
        sig_mask = np.zeros(sigma_u.shape)
        # set =1 twice to ensure, that all the ones needed are below the main
        # diagonal:
        sig_mask[causing_ind, caused_ind] = 1
        sig_mask[caused_ind, causing_ind] = 1
        vech_sig_mask = util.vech(sig_mask)
        inds = np.nonzero(vech_sig_mask)[0]

        # Make restriction matrix
        C = np.zeros((num_restr, len(vech_sigma_u)), dtype=float)
        for row in range(num_restr):
            C[row, inds[row]] = 1

... aren't quite doing the same thing, but look related. I'll see if I can sort this out.

[jbrockmendel]

Uh also it looks like VARResults.cov_params may having a wrapping problem

data = sm.datasets.macrodata.load_pandas().data
endog = data[['realgdp', 'realcons', 'realinv']].diff().iloc[1:]
endog.index = pd.date_range('2016-01-01', periods=len(endog))
model = sm.tsa.VAR(endog)
res = model.fit(3)

>>> res.cov_params
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/local/lib/python2.7/site-packages/statsmodels/base/wrapper.py", line 41, in __getattribute__
    obj = data.wrap_output(obj, how=how)
  File "/usr/local/lib/python2.7/site-packages/statsmodels/base/data.py", line 411, in wrap_output
    return self.attach_cov(obj)
  File "/usr/local/lib/python2.7/site-packages/statsmodels/base/data.py", line 531, in attach_cov
    columns=self.param_names)
  File "/usr/local/lib/python2.7/site-packages/pandas/core/frame.py", line 361, in __init__
    copy=copy)
  File "/usr/local/lib/python2.7/site-packages/pandas/core/frame.py", line 533, in _init_ndarray
    return create_block_manager_from_blocks([values], [columns, index])
  File "/usr/local/lib/python2.7/site-packages/pandas/core/internals.py", line 4631, in create_block_manager_from_blocks
    construction_error(tot_items, blocks[0].shape[1:], axes, e)
  File "/usr/local/lib/python2.7/site-packages/pandas/core/internals.py", line 4608, in construction_error
    passed, implied))
ValueError: Shape of passed values is (30, 30), indices imply (10, 10)

[josef-pkt]

open a separate bug issue for VARResults.cov_params
workaround use res._results.cov_params to access the underlying ndarray

I'll let you figure out the two causality tests, but they look very different to me, the first quote is making a restriction matrix for params, the second is making a restriction matrix for vech(sigma_u) (naming cov_resid = sigma_u, IIRC)