The question of how to interpret IAT scores comes up a lot, so I'd post an explanation here to help demystify this subject.

As a participant completes the IAT, Inquisit keeps a running trial-by-trial tally of d, which is the standard metric used to interpret IAT results. The final d score for a given participant thus appears on the last row of data for that participant. Inquisit reports 3 d scores, which appear in the last 3 columns labeled "expressions.da", "expressions.db", and "expression.d".

Expressions.da is the d score from just the practice (i.e., first) blocks for both pairings.

Expressions.db is the d score from just the test (i.e. second) blocks for both pairings.

Expressions.d is the d score for both practice and test blocks. This is the score that is typically reported as the measure of association.

D scores can be positive or negative. A positive score indicates an association of targetA with attributeA  and targetB with attributeB.  A negative score indicates an association of targetA with attributeB and targetB with attributeA. Translating the score into a preference or attitude thus depends on how you have assigned your real world categories to these 4 groups. You can determine the mappings by looking at the topmost section of your IAT script, where you'll see the <item> definitions for each category.

For example, let's say you are studying implicit attitudes towards neckties vs ascots, where neckties are targetA, ascots are targetB, pleasant words are attributeA, and unpleasant words are attributeB. A positive d score would indicate the participant is more of a necktie sort of person. A negative score would indicate an affinity for the classic and elegant ascot look. A score of zero would indicate no preference.

Inquisit also records all of the raw responses and response latencies for every trial so that you have the option of running post hoc analyses on the data (e.g., to experiment with different methods of handling outliers). We offer an SPSS command script that provides a good starting point for this, which you can download from our IAT page:

http://www.millisecond.com/download/samples/v3/IAT/default.aspx

If you run the script as is, you'll get exactly the same d scores as those described above, but the script can be easily edited to do custom analyses.

-Sean

One addition on my part, if I may, since this also a common source of confusion: IAT D-scores may vary between -2 and +2, *not* -1 and +1 as is often erroneously stated (Sriram, Greenwald, & Nosek (2006), p. 57, footnote 2; Nosek & Sriram (2007), p. 396, footnote 2)

~Dave

Hi,

When I look at my data, I do not see the rows labeled with "expressions" at all. I see all the individual latency scores for each trial, though. What does this mean? Can I run the SPSS script and get the D scores that way? I also have my survey data to deal with. What steps do you recommend so that I can get my D score and then correlate to data?

Thanks,

Jamie

This probably means you did not use any of the templates available from http://www.millisecond.com/download/samples/v3/IAT/default.aspx (all of which compute D) but instead used an outdated script from a different source (see http://www.millisecond.com/forums/Topic3889.aspx#3889).

Once again #2 in http://www.millisecond.com/community/forums/p/1456/4893.aspx#4893.

This sounds like a topic that should be discussed between you and your advisor.

Regards,

~Dave

Hi Sean,

First I want to say that I really appreciate this overview!  It has been helpful to me as I have begun my data analysis.

One question I have regarding how the data are calculated is whether errors and latency were considered when calculating d scores.  In other words, were Greenwald et al.'s (2003) "improved scoring algorithm" recommendations for calculating d scores followed?  If not, would I need to modify and use the syntax that has been developed in order to examine errors and latencies?

Thanks for the help!

Laura

Laura,

the IAT templates scripts available at http://www.millisecond.com/download/samples/v3/IAT compute "D with built in error penalty" (d_biep) according to Greenwald et al. (2003). This is the recommend measure for the specfic IAT procedure implemented by these scripts. The only difference to Greenwald et al.'s improved scoring algorithm is that the Inquisit scripts do not discard data or stop the IAT if subjects have < 300ms latencies on 10+% of the trials. You would have to do that later during data preparation. You can compute other variants of D based on the raw data with your preferred stats package.

Regards,

~Dave

HI Dave,

As always, thanks for the quick response!  I should have been more specific.  Are the "expressions.d" scores calculated using Greenwald's improved scoring algorithm?  How do the "expressions.d" scores differ from "d_biep" scores?  Basically I am wondering if I can use the scores that were automatically calculated for me or if I need to run the template script/syntax in order to follow Greenwald's guidelines.  Thanks!

Laura

As I already said in my previous response: 'expressions.d' is 'd_biep' which is computed according to the improved scoring algorithm, the only difference being that subjects with latencies below 300ms on 10+% of the trials are not automatically discarded.

~Dave

Hi Dave,

Thanks for being so helpful. I have downloaded a sample script from the millisec's site, but i couldn't find "d_biep" in the IAT script, nor is it in the .dat file. I'm wondering does that mean that the when an error is made, does the script simply compute the latency after a correct response is made, or does it add 600 ms to the block mean?

See my previous reply to this thread. The IATs are set up to only move on after the correct key has been pressed, thus D includes a "built in error penalty" (biep).

Regards,

~Dave

Hi,

I am having trouble figuring out which one goes with which in my case (How to pair them and make them A or B). I am working with the attributes good and bad and the targets Latino and White. IN my script I have them as Attribute_Left: good, Attribute_right: bad, target 1: white, target 2: Latino. I will attach a screen shot of my results after I conduct a frequency test in spss. I would appreciate any help. If more information is required please do not hesitate to let me know.  

Elizabeth

For the IAT scripts available from millisecond.com (to which this thread applies), the denominations and pairings are clear: They are always attrinbuteA, attributeB, targetA and targetB. For scripts coming from other sources, please refer to the documentation provided by the respective scripts' authors.

Regards,

~Dave

I used the script available from prof. Greenwald's website. I will attach it to see if it helps.[View:http://www.millisecond.com/community/cfs-file.ashx/__key/CommunityServer.Discussions.Components.Files/12/2553.LATINOIAT.exp.txt:550:0]

Then you should refer to the documentation provided by Prof. Greenwald re. interpreting the data generated with that script and how his accompanying SPSS syntax works.

Hi,

I'm doing a research paper on self esteem, and I'm using the Greenwald self esteem IAT and the Rosenberg explicit self esteem test. My question is, how to correlate the two mesasures, or how to make them "compatible" for each other. I have the raw data (the d scores from the IAT, and the points from the rosenberg scale).

Thanks!

This is a general methodological / data analysis question and is as such completely unrelated to (using) Inquisit. I thus recommend you review the available literature dealing w/ relations between implicit and explicit measures and take this up w/ your colleagues and/or advisors.

Regards,

~Dave

Hey Dave,

thank you for your explanations which are always very helpful.

I understood that expressions.d is the IAT-sore calculated after the improved scroing algorithm according Greenwald et al. 2003. This means I can cite it this way, right?

I used one of the scripts of your Task library. I´m not sure which part of the script calculates the expressions.d-score. Maybe you could upload an example of the accordant command? This would help me to control my data.

One last question: Greenwald et al. (2003) used the pooled SD for their calculations. Which formula did you use for computing the pooled SD?

Thank you very much in advance!

Kind regards, Magdalena

Yes, review the previous replies throughout this thread for details.

Simply look at the <expressions> element in the script you used and find '/ d'. Additionally look at the /ontrialend attributes in the script which are involved the necessary computations.

You'll also find this information directly by looking at the script. Again, refer to the <expressions> element in the script and look at the various entries starting with 'sd'.

Regards,

~Dave

Hey Dave,

thank you for your quick reply. You already helped us very much. But we are still wondering about the resulting score (expressions.d). We tried to compute the score for several subjects manually and got different values than the program calculated (differences about 0.04). So we looked at the formula in the script for understanding which mistake we made. But we couldn´t comprehend on part of the SD-formula, what means values.ssXX?

Thanks a lot and best wishes, Magdalena

'ss' would indicate a sum of squares. Minor differences a la 0.0x may very well be due to rounding error.

Hi Dave, I have only one last question.

The IAT is making compatible and incompatible matching but not in the same order every time. My question is, that this is calculated in the last d score, or i have to make some changes with the reversed ones.

Thanks in advance,

Martin

The order of compatible / incompatible blocks is counterbalanced between subjects based on the numerical subject id -- see the <variables> element for details. This does not affect the computation of d in any way.

Hey Dave,

I've read in another thread that having a D score that is higher than 1.0 is rare, however, I seem to have 4 instances of it in my data. 

Any ideas why this might be?

Thanks

Rare does not mean impossible. D may vary between -2 and +2 in theory. Other than that, I have nothing to offer on this subject, I'm afraid. Might be your particular subject matter, etc. Those are theoretical and methodological questions, so you should be better equipped than anybody else to answer them (since you know your particular area of research best).

Cool. I just wanted to check that it wasn't a technical issue. 

Thanks 

Dear Sean,

Can I just check that from having read the above (very helpful) description and having translated that to the Single Target IAT that I have understood the expression.d score for the Single Target IAT correctly:

A positive score would mean there was an association between Target A and Attribute A

A negative score would mean there was an association between Target A and Attribute B

So with you necktie example. Lets say Target A was neckties, Attribute A was pleasant and Attribute B was unpleasant. A positive score would mean the participant liked neckties, a negative score would mean that they did not like neckties? With my study Target A is 'disabled', attribute A 'pleasant' and attribute b 'unpleasant'. So a positive score would mean that participants associated 'disabled' with 'pleasant' (thus positive implicit attitudes), and a negative score would mean that participants associated 'disabled' with 'unpleasant' (thus negative implicit attitudes) right?

Many thanks,

Michelle

Yes.

Hello, I have received conflicting reviews on cutoff scores for the IAT. In one article it is said that any value above +0.65 or below -0.65 indicates a preference one way or another. Some other areas say that any value positive or negative indicates a preference whereas 0 does not. In my data the value I receive (mean value of 150 participants) was 0.46, standard deviation of 0.31. Does this indicate a preference? I would be happy to talk more if you should have the time.

Thank you for your help,