Long RTs/SSDs lead to error

[Gilles86]

Hi again,

I have a task with relatively long RTs. For some subjects and some conditions the RTs/SSDs are more than 1000ms. It seems that for these subjects the stopsignal-library breaks down:

After 7.000000 retries, still no good fit found.

---------------------------------------------------------------------------
ZeroProbability                           Traceback (most recent call last)
<ipython-input-2-63e49e851d1c> in <module>()
     26 
     27 
---> 28         ss = stopsignal.StopSignal(ds, depends_on={'mu_go':['difficulty', 'color']})

/usr/local/lib/python2.7/dist-packages/stopsignal/stopsignal.pyc in __init__(self, data, **kwargs)
     73         self.std_depends = kwargs.get('std_depends', False)
     74 
---> 75         super(StopSignal, self).__init__(data, **kwargs)
     76 
     77     def create_ss_knode(self, knodes):

/usr/local/lib/python2.7/dist-packages/kabuki/hierarchical.pyc in __init__(self, data, is_group_model, depends_on, trace_subjs, plot_subjs, plot_var, group_only_nodes)
    346         self.db = None
    347 
--> 348         self._setup_model()
    349 
    350     def _setup_model(self):

/usr/local/lib/python2.7/dist-packages/kabuki/hierarchical.pyc in _setup_model(self)
    357 
    358         # constructs pymc nodes etc and connects them appropriately
--> 359         self.create_model()
    360 
    361     def __getstate__(self):

/usr/local/lib/python2.7/dist-packages/kabuki/hierarchical.pyc in create_model(self, max_retries)
    437         else:
    438             print "After %f retries, still no good fit found." %(tries)
--> 439             _create()
    440 
    441         # create node container

/usr/local/lib/python2.7/dist-packages/kabuki/hierarchical.pyc in _create()
    427         def _create():
    428             for knode in self.knodes:
--> 429                 knode.create()
    430 
    431         for tries in range(max_retries):

/usr/local/lib/python2.7/dist-packages/kabuki/hierarchical.pyc in create(self)
    166                 kwargs['doc'] = node_name
    167 
--> 168             node = self.create_node(node_name, kwargs, grouped_data)
    169 
    170             if node is not None:

/usr/local/lib/python2.7/dist-packages/stopsignal/stopsignal.pyc in create_node(self, node_name, kwargs, data)
     54         kwargs['value'] = new_data['rt']
     55         kwargs['issd'] = np.array(new_data['ssd'], dtype=np.int32)
---> 56         return self.pymc_node(name=node_name, **kwargs)
     57 
     58 class KnodeInhibitions(Knode):

/usr/local/lib/python2.7/dist-packages/pymc/distributions.pyc in __init__(self, *args, **kwds)
    269                 random = debug_wrapper(random)
    270             else:
--> 271                 Stochastic.__init__(self, logp=logp, random=random, logp_partial_gradients = logp_partial_gradients, dtype=dtype, **arg_dict_out)
    272 
    273     new_class.__name__ = name

/usr/local/lib/python2.7/dist-packages/pymc/PyMCObjects.pyc in __init__(self, logp, doc, name, parents, random, trace, value, dtype, rseed, observed, cache_depth, plot, verbose, isdata, check_logp, logp_partial_gradients)
    757         if check_logp:
    758             # Check initial value
--> 759             if not isinstance(self.logp, float):
    760                 raise ValueError(
    761                     "Stochastic " +

/usr/local/lib/python2.7/dist-packages/pymc/PyMCObjects.pyc in get_logp(self)
    914                     (self._value, self._parents.value))
    915             else:
--> 916                 raise ZeroProbability(self.errmsg)
    917 
    918         return logp

ZeroProbability: Stochastic srrt_like(green.hard)'s value is outside its support,
 or it forbids its parents' current values.

I tried upping the limit of mu_go in stopsignal.py, line 94 to 2e3, but that doesn't seem to help.

Any thoughts?

Thank you so much!

[twiecki]

Hmmm, sounds like outliers. Have you tried excluding very long and very short RTs?

[Gilles86]

Hey,

This helps for some subjects. But then, when fitting the data in a hierarchal model, I now get an integration error after taking a lot of samples.

[---------------- 44% ] 8811 of 20000 complete in 1403.7 secgsl: qag.c:261: ERROR: could not integrate function Default GSL error handler invoked. Aborted (core dumped)

Can this exception be caught for problematic samples, so these can be skipped?

[twiecki]

Sorry for the slow response. I would figure out what parameter values cause it to crash (which would require some digging into the code).