Estimates of drift rate 'v' varying with a categorical variable change when the reference level of the variable is changed

[Lipika-T]

Hello,

I am trying to fit a drift diffusion model to a task where the participants have to identify a target symbol in each trial and there are 2 possible targets/responses (2 choice). The 2 target symbols appear in equal number of trials. Additionally, each trial belongs to an Emotion condition which has 3 levels, and all trials are intermixed. I have tried to model this in a stimCoding way with the 2 targets as the bounds and varying the drift rate 'v' with Emotion. The target symbols are coded as '1' and '-1' in the column 'stim' in the dataset and is taken as a categorical variable. By default, the Emotion level 'Negative' was taken as reference level. But when I changed the reference level to 'Neutral' the estimates change, even relative to each other. For instance, the drift rate for stim=1 is larger for Neutral than Negative when Negative is the reference and is larger for Negative compared to Neutral when Neutral is the reference, as given by the 'v_C(stim):C(Emotion)[1,Neutral]' and 'v_C(stim):C(Emotion)[1,Negative]' terms respectively. Is there something wrong with the way this is coded or interpreted?
From what I understand the values are computed in this way, for example when Negative is the reference :
v(stim=1,Emotion=Negative) = v_Intercept + v_C(stim)[1]
v(stim=1,Emotion=Neutral) = v_Intercept + v_C(stim)[1] + v_C(Emotion)[Neutral] + v_C(stim):C(Emotion)[1,Neutral]

HSSM version
'0.2.3'

Screenshots

Example dataset from one participant

Default model with Negative as reference level

Model with Neutral as reference level

[frankmj]

Hi there, I believe this relates to this point below https://patsy.readthedocs.io/en/latest/formulas.html#redundancy-and-categorical-factors you might try running a simpler model of just v ~ 0 + C(stim):C(Emotion) and see what you get

…

On Wed, Jan 8, 2025 at 1:07 PM Lipika-T ***@***.***> wrote: Hello, I am trying to fit a drift diffusion model to a task where the participants have to identify a target symbol in each trial and there are 2 possible targets/responses (2 choice). The 2 target symbols appear in equal number of trials. Additionally, each trial belongs to an Emotion condition which has 3 levels, and all trials are intermixed. I have tried to model this in a stimCoding way with the 2 targets as the bounds and varying the drift rate 'v' with Emotion. The target symbols are coded as '1' and '-1' in the column 'stim' in the dataset and is taken as a categorical variable. By default, the Emotion level 'Negative' was taken as reference level. But when I changed the reference level to 'Neutral' the estimates change, even relative to each other. For instance, the drift rate for stim=1 is larger for Neutral than Negative when Negative is the reference and is larger for Negative compared to Neutral when Neutral is the reference, as given by the 'v_C(stim):C(Emotion)[1,Neutral]' and 'v_C(stim):C(Emotion)[1,Negative]' terms respectively. Is there something wrong with the way this is coded or interpreted? From what I understand the values are computed in this way, for example when Negative is the reference : v(stim=1,Emotion=Negative) = v_Intercept + v_C(stim)[1] v(stim=1,Emotion=Neutral) = v_Intercept + v_C(stim)[1] + v_C(Emotion)[Neutral] + v_C(stim):C(Emotion)[1,Neutral] HSSM version '0.2.3' Screenshots Example dataset from one participant Screenshot.from.2025-01-08.19-50-05.png (view on web) <https://github.com/user-attachments/assets/1c1dffe2-5b7c-4d9a-9f10-338458fafb8c> Default model with Negative as reference level Screenshot.from.2025-01-08.19-49-31.png (view on web) <https://github.com/user-attachments/assets/021ccaa6-c68a-4a50-9c9a-2f760f9ed578> Screenshot.from.2025-01-08.18-53-08.png (view on web) <https://github.com/user-attachments/assets/733f3e57-a2e9-4cb8-83d6-c1cc4e0dd883> Model with Neutral as reference level Screenshot.from.2025-01-08.19-48-52.png (view on web) <https://github.com/user-attachments/assets/9cae5540-fe95-4855-aae1-75a477206735> Screenshot.from.2025-01-08.18-52-48.png (view on web) <https://github.com/user-attachments/assets/72d3edb8-90c3-44cf-8e8a-32264b259bd6> — Reply to this email directly, view it on GitHub <#623>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAG7TFDFML4DA6ONBUW4HXD2JVSPNAVCNFSM6AAAAABU2PRZAGVHI2DSMVQWIX3LMV43ERDJONRXK43TNFXW4OZXG44TQOBZGU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[AlexanderFengler]

@Lipika-T did @frankmj's advice yield here?

[Lipika-T]

Thank you for your response. I tried the above model and got 6 drift rate values so there is no reference level here.

In that case do we take an average of the absolute values for the 2 drift rates (for stim 1 and -1) for each emotion if we want to report a single drift rate value?