interpreting fitted bias on drift in an accuracy-coded model

[ari-khoudary]

hi, thanks so much for this great tool. I have embarrassingly simple questions regarding interpretation of fitted values for a biased drift rate in an accuracy coded model.

On the cookbook page, you write "In accuracy coding, if the upper boundary is the correct choice and the lower boundary is the incorrect choice, we need to flip the bias depending on whether the biased choice is correct or incorrect. However, in this case, the motion coherence should always be positive." and use the following example:

m = pyddm.gddm(drift=lambda coh,driftmultiplier,bias,biascorrect : coh*driftmultiplier + bias*(1 if biascorrect else -1),
               parameters={"driftmultiplier": (0, 2), "bias": (0, 2)},
               conditions=["coh", "biascorrect"],
               choice_names=("Correct", "Error"))
pyddm.plot.model_gui(m, conditions={"coh": [0, 1, 2], "biascorrect": [0, 1]})

I am fitting data from an experimental design where motion coherence is constant across trials. However, every trial contains 2 periods of completely uninformative evidence (which we call noise periods). I am interested in testing the hypothesis that, when there is a bias in the environment, drift rates are non-zero during the noise periods. To do this, I wrote the following drift function:

def drift_biased(t, congruent, signal1_onset, noise2_onset, signal2_onset,
                 n1_biased, s1_cong, s1_incong, 
                 n2_cong, n2_incong, s2_cong, s2_incong):
    # drift rate during first noise period
    if t < signal1_onset:
        n1_biased
    # drift rates during first signal period
    elif t >= signal1_onset and t < noise2_onset:
        if congruent == 'congruent':
            return s1_cong
        else:  # incongruent
            return -s1_incong
    # drift rates during the second noise period
    elif t >= noise2_onset and t < signal2_onset:
        if congruent == 'congruent':
            return n2_cong
        else:  # incongruent
            return -n2_incong
    # drift rates during the second signal period
    elif t >= signal2_onset:
        if congruent == 'congruent':
            return s2_cong
        else:  # incongruent
            return -s2_incong

If I fit a model specified as

        model = pyddm.gddm(
            drift = drift_biased,
            starting_position = 0,
            bound="B",
            T_dur = 4.3,
            nondecision=0,
            parameters={'B': (1, 15), 
                        'n1_biased': (-1, 10), 
                        's1_cong': (-1, 10), 's1_incong': (-1, 10),
                        'n2_cong': (-1, 10), 'n2_incong': (-1, 10),
                        's2_cong': (-1, 10), 's2_incong': (-1, 10)},
            conditions = ['congruent', 'signal1_onset', 'noise2_onset', 'signal2_onset']
        )

what is the "right" way to plot / interpret the incongruent drift fits? specifically, if my model returns a positive value for n2_incong, for example, should I interpret that as the model actually estimating -1*n2_incong as the drift? and what if I get a negative value? do the negatives cancel out, such that the fitted value is actually positive?

I know in the cookbook you restricted the values of bias to be strictly positive. Is that necessary for interpretable fits on this approach? Or can I lower-bound the MLE process at -1 and still get interpretable fits?

Thanks so much for the help!

[mwshinn]

I don't notice anything wrong with your implementation. To understand the parameters it might be helpful to visualise it with the model GUI.