This would probably depend on whether or not you are interested in the processes occurring during correct vs incorrect responses - and/or the rt differences between the two conditions. If you are only interested in the processes occurring between different prime images then you might not be interested in having incorrect responses as a part of your main analysis. My suggestion would be to exclude them from your main analysis, but to keep hold of them somewhere and to speculate upon them should you find something you would like to investigate. As with non-significant results, incorrect responses do often provide good data - just not everybody is interested. At worst you could use these as a pilot for a follow up....

I agree with Nick, it sounds like your design require response based rejection

the main issue is that your error is probably associated by a classical ERP "error-negativity" and consequently it is probably associated with a specific shape in your EEG.

I think it is better to remove this kind of trials

I also agree that you need to remove this trials especially if you are interested in a comparison between behavioural response and eeg data.

Thank you, Nik. As I understood, some remove trials with incorrect answers, because it is supposed that when subjects answer erroneously it means that they do not pay attention to the task (they are distracted by thoughts or sensations, or they are tired).

If erroneous RT means a non-nonsensical (abnormal) response from the point of view of the objective of the study, then the answer is yes.

Not a problem. I would not simply write them off however. Careful inspection of the eeg waveform could help to reveal whether or not it was a simple mistake in judgement (and hence a cognitive effect worth investigating), or simply down to a poor trial (as you said, tiredness, distractions...), although this will require a fair amount of time. If the latter is the case then maybe you can refigure your design slightly to be provide ample opportunity to remove these effects. If you sacrifice a few minutes of recording by offering less trials or a break here and there then you will ultimately maintain cleaner data and thus less mistakes (hopefully). This might not be feasible, but potentially writing a short questionnaire and asking how the participant felt (ie tired, stimulated, distracted, adapting to the task etc --- rather than what they thought of the experiment) might help to assess potential risks to data cleanliness in follow up studies. The best of luck Vladimir, I look forwards to seeing what comes out of this.

Most researchers would remove erroneous trails from analysis.

You can also do analysis with and without them and see if it makes any difference and then decide and accordingly discuss the results.

I agree with Mante...

An erroneous answer could also be due to the fact that (1) the subject is starting to respond to expectation or (2) is anxious (You will see that often when working with psychiatric patients.).

I would suggest that you compare the results with and without erroneous answer, to explore if there is any significant difference.

Maybe grouping subjects without erroneous and with erroneous answers could yield some differences.

RT filtering of ERP is a complicated question. It depends on what components are you interested in, the more early the component is the less affected is by endogenous processes. Also if you filter ERP based on RT in my opinion it is worth considering too short and long non-error RTs as well. Erroneous trials are informative themselves, e.g. because of prevoiusly mentioned ERN responses.

I'm not completely sure what you mean by "Erroneous RTs", but as a general procedure, either the error-related segments were previously defined as being of interest, or these epochs are excluded from the ERP analysis.

Generally, the ERP analysis includes trials with correct answers. If this is not mentioned, this might be either an omission or really because all trials were included.

kind regards

Asaid

Thanks to everyone. You have helped me a lot. I think, first, I will study only trials with correct answers (as a team did with a very similar design) and if I have time I will play with the data (include incorrect answer trials, study only incorrect answer trials, compare correct and incorrect answer trials, etc.). And as Ágoston said, we always delete very quick and very late answer trials.

Dear Vladimir,

you should definitely remove the trials with erroneous responses since they contain the error (related) negativity which may strongly influence your correct ERP pattern.

Greetings,

Michael falkenstein

Thank you very much, dear Michael. It is an honour to receive comments from a pioneer in the field.

In a sense, it also depends upon what you are trying to measure with your task and the ERPs derived during it. If you are trying to measure a cognitive or affective process that is engaged in a trail condition, than an incorrect response suggests the possibility that the trial in question, for whatever reason, may not have engaged that process and therefore the ERP from it is an invalid measure of what you are tring to get at. Sometimes, as with ERN, the point is to measure what occurs during an incorrect behavioral response trial. In those cases the validity of the measure depends on having only incorrect trials. If incorrect responses possibly lower the validity of your measure and increase error, then I'd say don't include them just as you would not include correct responses when averaging an ERN.

I also think it depends on your hypothesis and aim of research. You can include with or without erroneous RT, but i agree that you at first exclude erroneous RT AND do an analysis then do with erroneous RT too.

As a start, it is always better to begin with trials without bad RT (too early or too late). Nevertheless, I think it is important to explore bad epochs as well, especially when comparing psychiatric/neurologic groups with controls.

My preference is to have the same trials contributing the analyses of behavioral and ERP data. Unless errors were very rare in the task, including the effects in the ERP averages could result in different conclusions for behavioral and ERP data. Also, stimulus locked ERPs for errors and correct trials are often quite different (e.g., remembered vs. forgotten items), in addition to response locked error ERPs (e.g., ERN, Pe).

If people respond too late or too early (the boundary depends on your task) it is clear that they were not doing what they were told to do. That means there can be nothing but noise in the signals you measure there, and those trials should obviously all be removed.

Agreed with Dr West that the behavioral and ERP data need to be based on the same trials, so first apply whatever criteria for trial exclusion you would use for your behavioral data analysis and follow same in your ERP analysis. What is most important here will be the criteria you use to reject trials based on RT - this is more difficult than simply excluding error trials. Depending on where you set a cut point for acceptable RTs, and how many trials this excludes, it may be sensible to analyze rejected trials as a contrasting condition. Along these lines, if RT is the most important behavioral dimension of the task you may instead want to analyze all data but segment according to theoretically relevant cut points in RT; i.e., fast response, intermediate response, slow response. I would not assume that extreme RTs reflect noise in data per se, the introduction of "noise" just depends on the choices you make in averaging the data and whether these averages include trials they should not.

I would also recommend to exclude incorrect trials. However, I would repeat the analysis with correct & incorrect trials to see whether there is a difference. If there is no difference in the pattern of results, it is up to you whether you do it the one way or the other.

cheers, Tobi

I would study separately trials with correct and incorrect answers: they both contain useful information. Cheers

If you are not interested in studying the errors I would also suggest removing trials following an error. In terms of behavior and neural mechanisms, there are robust post-error effects.

A good point by Joseph, however, if your stimulus onset is jittered then you should reduce the need to do this.

All of the prior respondents to make valid points depending on the protocol you're using I would Exclude Error trial for the first pass. It would be helpful to analyze the error trials separately and compare the outcomes of the two. But the final analysis depends on what you're trying to accomplish with the research.

Always remove ERP trials that have been classified incorrect on a behavioural basis to obtain an equal set of trials for the electrophysiological and the behavioural analyses. I guess it is more tricky with trials that are rejected due to their RT (either too fast or too slow) because technically these trials are not incorrect. Very fast reaction times sometimes include technical failures (and the correctness of the response is difficult to judge). Slow correct RTs are not errors - however, if they are excluded for the overall RT analysis, my advice is to exclude them from the ERP analysis as well.

NIels has a point, However, whether there is a difference for correct vs incorrect trials depends on the ERP marker you investigate. If you look at an early component (eg. the somatosensory P50 component) it probably doesn't make a difference, whereas it might look very different for late components such as the decision-related P300 component. I have a couple of studies where there was absolutely no difference in the markers of interest (early components) for analyses with vs without incorrect trials. In fact, excluding incorrect trials reduced the signal-to-noise ratio, so I left them in. It were quite difficult tasks, by the way, in which participants made relatively many errors.

There is a potential problem with excluding incorrect trials: if one experimental condition is more difficult than the other you will end up with unequal numbers of observations. ERP components may tend to become larger with fewer trials and this would be a confound in the study. Hence, I would always do the analyses with and without incorrect trials. If the pattern of results is the same, I would opt for leaving incorrect trials in the sample.

cheers,

Tobi

First of all, what is an erroneous RT? Do you mean an erroneous response has happened when your subjects responded to the wrong stimulus? Or do you consider a response as erroneous when it was too slow or too fast or issued with the wrong finger, hand, foot etc? Besides this, the answer to your question is depending on your research question: in the case you are only interested in the response to any visual stimulus you should include all trials no matter whether the response was correct or not. If your interest is in correct responses to the stimuli, I would only use the trials with correct responses. You might also test whether there are differences in the ERPs as a function of correctness/incorrectness etc. Then also test the ERPs to incorrect trials and compare the two groups of ERPs. Best Wolfgang

Generally, the convention, depending on task and data set, is to organize the data into bins including correct hits, correct rejections, misses and false alarms. What data you analyze and report depends on your study and hypotheses. It is possible to report RT, ERP and percent data for each bin.

Best regards,

Alan

All previous points are well taken. I want to raise a more principled issue. In many cases we want to use ERPs to gain information about the processing flow elicited by an event, often (but not in all experiments - e.g. in "passive" viewing) terminating in a response. This processing flow may take different courses. In one case all necessary processing steps will occur and a correct response will ensue. in other cases the response will be incorrect, e.g. because the correct answer is not known or the stimulus presentation has not been formed or has been forgotten and the participant guesses (slow guess) or because there is high time pressure and the participant guesses on the basis on minimal or no stimulus analysis (fast guess). In all these cases the routines used in processing the stimulus are different from the first case and if errors are mixed up with correct responses, results may be confusing. Hence, I very strongly opt for analyzing correct and incorrect responses separately. Note, however, that guessing may also be lucky and yield a correct response, albeit via a different routine than for other correct responses (Case 1). Although we cannot be certain, many of these lucky guesses, whether fast or slow can be excluded by eliminating very fast and very slow outliers.

In conclusion, if you are interested in "standard" correct information processing, you should eliminate all incorrect responses and fast and slow outliers of correct responses. If this leads to a too low number of trials, the design of the study should be improved and the experiment repeated but you should not mix up trials with different processing routines.

Dear Vladimir, I totally agree with Tobias Katu and Luis Garcia-Larrea. I would also suggest you do not use any software for automatical trials rejection, but rather to inspect visually the trials and separate those with erroneous responses from the others. Finally to compare the results get from the whole set of trials with the "suitably chosen" ones.