Journal:
Open Differential Psychology
Authors:
Emil O. W. Kirkegaard
Title:
Criminal and antisocial behavior and cognitive ability in a sample of dating site users
Abstract:
The relationship between cognitive ability and a broad variety of criminal and antisocial (CAS) behaviors was examined in a sample of dating site users. Cognitive ability was scored from 14 suitable items using item response theory. 11 of the 13 items were negatively related to cognitive ability with latent correlations ranging from .01 to -.29, and a median of -.14. Correlations for males were stronger with a mean/median male/female ratio of 1.7/1.4.
The CAS items were all positively correlated (mean correlation = .23) and formed a general CAS factor. Attempts were made to use item response theory to score this factor, but results proved unsatisfactory. Instead, a simple score was calculated by summing the items. This score correlated -.16 [-.14 to -.18] with cognitive ability.
To examine the influence of confounders, an OLS model was fit using cognitive ability, gender, age, and cognitive ability x gender as predictors and the simple summed score as the dependent. In this model, the standardized beta cognitive ability was -.19. A series of robustness checks were made using alternative model specifications which provided near identical results.
Key words:
crime, antisocial behavior, violence, cognitive ability, IQ, intelligence, dating site, gender, sex
Length:
~4000 words, ~16 pages.
Files:
https://osf.io/7htpz/
Reviewers:
I will attempt to recruit one of the following: Brian Boutwell, Joseph Schwartz and John Wright.
Back to [Archive] Post-review discussions
One strange thing is that the age controlled (logistic regression) beta for IQ was a smaller (r equivalent .11), but age isn't associated with IQ in this sample. I should think that controlling for age would make the association stronger, not weaker. Any ideas?
Just a few suggestions.
1. You should cite the classical review of the delinquency-IQ relationship by Hirschi and Hindelang: Hirschi, T. and Hindelang, M.J., 1977. Intelligence and delinquency: A revisionist review. American Sociological Review, pp.571-587.
2. Another possible cause of low correlations that you could mention is that those who had serious trouble with the police may be less likely than those with minor offenses to admit to their arrest history on a dating site. Most of those who do most likely had been arrested for very minor offenses or on a false suspicion, under circumstances that they would be prepared to talk about with a date. According to data I have seen, minor delinquency does not have a negative relationship with IQ. Only more serious offenses and convictions do. Minor disrespect for the law and specific types of offenses (e.g., political ones, or blasphemy where this is an offense) may occasionally even have a positive relationship with IQ.
1. You should cite the classical review of the delinquency-IQ relationship by Hirschi and Hindelang: Hirschi, T. and Hindelang, M.J., 1977. Intelligence and delinquency: A revisionist review. American Sociological Review, pp.571-587.
2. Another possible cause of low correlations that you could mention is that those who had serious trouble with the police may be less likely than those with minor offenses to admit to their arrest history on a dating site. Most of those who do most likely had been arrested for very minor offenses or on a false suspicion, under circumstances that they would be prepared to talk about with a date. According to data I have seen, minor delinquency does not have a negative relationship with IQ. Only more serious offenses and convictions do. Minor disrespect for the law and specific types of offenses (e.g., political ones, or blasphemy where this is an offense) may occasionally even have a positive relationship with IQ.
Thanks for your comments, Gerhard.
Today when going thru the dataset for another project, I found another more serious crime outcome variable: "Have you ever been to prison?"
I will expand the paper to also use this outcome.
Today when going thru the dataset for another project, I found another more serious crime outcome variable: "Have you ever been to prison?"
I will expand the paper to also use this outcome.
Gerhard,
I have added a few more references, including the review you mentioned.
I have finished a new draft. This includes an expanded analysis with 5 outcomes, factor analysis, Jensen's method.
Files updated.
https://osf.io/7htpz/files/
I have added a few more references, including the review you mentioned.
I have finished a new draft. This includes an expanded analysis with 5 outcomes, factor analysis, Jensen's method.
Files updated.
https://osf.io/7htpz/files/
Just a few things:
1. In Figure 1, why is prison so high up on the X axis? According to Table 2 its correlation with cognitive ability should be almost as negative as the correlations for arrested and punched. Also, you should mention that this "Jensen effect" is not statistically significant (although suggestive) because of the small sample size of N = 5.
2. Last 2 sentences: If smarter people are more likely to admit to criminal or antisocial behaviors or ideation, this would not necessarily bias the correlations to zero, and the opposite to 1. In the first case, it would make the crime-IQ correlation more positive (less negative), and in the latter case it would make it more negative.
3. Check the spelling throughout. There are multiple corrections needed.
1. In Figure 1, why is prison so high up on the X axis? According to Table 2 its correlation with cognitive ability should be almost as negative as the correlations for arrested and punched. Also, you should mention that this "Jensen effect" is not statistically significant (although suggestive) because of the small sample size of N = 5.
2. Last 2 sentences: If smarter people are more likely to admit to criminal or antisocial behaviors or ideation, this would not necessarily bias the correlations to zero, and the opposite to 1. In the first case, it would make the crime-IQ correlation more positive (less negative), and in the latter case it would make it more negative.
3. Check the spelling throughout. There are multiple corrections needed.
Thanks for the review, Gerhard.
You are right! There is a discrepancy. That's not good. However, I found the bug. In my function to calculate the latent correlations, the method used to handle missing data was rowwise exclusion, while in the other case it was pairwise. I have changed the function to use pairwise as well and now the numbers match. The Jensen coefficient is now -.82.
I added a sentence about the high uncertainty. The confidence interval was already given in the plot, but just to be safe, it's okay to mention in the text as well.
In general, these single study Jensen's method results with few indicators (say, <30) are not to be taken too seriously. Better to do a bunch of studies and then meta-analyze á la te Nijenhuis.
I agree. That's also what I wrote. I wrote "towards 0/1", not "to 0/1".
I have read over the paper and corrected a number of spelling errors. Discussion has been substantially rewritten.
Files updated.
1. In Figure 1, why is prison so high up on the X axis? According to Table 2 its correlation with cognitive ability should be almost as negative as the correlations for arrested and punched. Also, you should mention that this "Jensen effect" is not statistically significant (although suggestive) because of the small sample size of N = 5.
You are right! There is a discrepancy. That's not good. However, I found the bug. In my function to calculate the latent correlations, the method used to handle missing data was rowwise exclusion, while in the other case it was pairwise. I have changed the function to use pairwise as well and now the numbers match. The Jensen coefficient is now -.82.
I added a sentence about the high uncertainty. The confidence interval was already given in the plot, but just to be safe, it's okay to mention in the text as well.
In general, these single study Jensen's method results with few indicators (say, <30) are not to be taken too seriously. Better to do a bunch of studies and then meta-analyze á la te Nijenhuis.
2. Last 2 sentences: If smarter people are more likely to admit to criminal or antisocial behaviors or ideation, this would not necessarily bias the correlations to zero, and the opposite to 1. In the first case, it would make the crime-IQ correlation more positive (less negative), and in the latter case it would make it more negative.
I agree. That's also what I wrote. I wrote "towards 0/1", not "to 0/1".
3. Check the spelling throughout. There are multiple corrections needed.
I have read over the paper and corrected a number of spelling errors. Discussion has been substantially rewritten.
Files updated.
Low cognitive ability has repeatedly been linked with antisocial and criminal conduct. The relationship appears to be robust and not conditioned by the measurement of cognitive ability or problem behavior. In this sense, the findings presented here converge with other studies.
The measurement of cognitive ability remains unclear. It would be helpful to know the actual 14-items used to create the scale, their content, and distribution. As the core variable in the analysis more detail is necessary.
The criminal behavior measure is, of course, a bit limited but prior research shows that even content limited measures usually have good statistical and content validity. Given the items included, which range form minor forms of deviance (cheating on tests), to more serious indicators of criminal behavior (arrest/prison), the expectation that a uniform factor would be extracted seems misplaced. Yes, criminals are generalists and this is confirmed with the inter-item correlations, but you really do not have a sufficient number of items to produce a reliable, single factor.
I would suggest creating a summated index and comparing the results against the results of found using only arrest. The additional variance in the index would allow you to more closely examine how cognitive scores perform across the index. Is a score of zero on the crime index associated with a cognitive score that is significantly higher than ascending index scores?
Prior research seemingly converges to show that average IQ differences between delinquents and non-delinquents is around .5sd. For adult offenders, average differences around 1sd are common. How much difference are we seeing with this sample of older, probably better educated and less criminal, men and women?
I’m not concerned with the performance of the age variable. Age, as you point out, can be thought of as an exposure period, or depending on the items used to construct the cognitive ability scale, it could reflect domain specific knowledge that is age-graded. Who knows.
Lastly, the sample is clearly highly selected but I don’t see this as a problem. Yes, the sample is composed of older people who use the internet and a case can be made that subjects may be a bit brighter than comparatively situated others......but the effect of this type of selection would be to minimize variance on crime and cognitive measures. The results would thus be more conservative compared to samples that contain more variance in these measures.
The measurement of cognitive ability remains unclear. It would be helpful to know the actual 14-items used to create the scale, their content, and distribution. As the core variable in the analysis more detail is necessary.
The criminal behavior measure is, of course, a bit limited but prior research shows that even content limited measures usually have good statistical and content validity. Given the items included, which range form minor forms of deviance (cheating on tests), to more serious indicators of criminal behavior (arrest/prison), the expectation that a uniform factor would be extracted seems misplaced. Yes, criminals are generalists and this is confirmed with the inter-item correlations, but you really do not have a sufficient number of items to produce a reliable, single factor.
I would suggest creating a summated index and comparing the results against the results of found using only arrest. The additional variance in the index would allow you to more closely examine how cognitive scores perform across the index. Is a score of zero on the crime index associated with a cognitive score that is significantly higher than ascending index scores?
Prior research seemingly converges to show that average IQ differences between delinquents and non-delinquents is around .5sd. For adult offenders, average differences around 1sd are common. How much difference are we seeing with this sample of older, probably better educated and less criminal, men and women?
I’m not concerned with the performance of the age variable. Age, as you point out, can be thought of as an exposure period, or depending on the items used to construct the cognitive ability scale, it could reflect domain specific knowledge that is age-graded. Who knows.
Lastly, the sample is clearly highly selected but I don’t see this as a problem. Yes, the sample is composed of older people who use the internet and a case can be made that subjects may be a bit brighter than comparatively situated others......but the effect of this type of selection would be to minimize variance on crime and cognitive measures. The results would thus be more conservative compared to samples that contain more variance in these measures.
John,
Thanks for reviewing.
The item questions, answer options, sample sizes and pass rates has been added in a table in the appendix (after references). Furthermore, a plot of the distribution of scores has been added to the main text.
A factor is really just a common source of variance. In this case, all items share variance, so it is possible to estimate a common cause of this shared variance, i.e. a common or general factor.
A summed score would in fact be very similar to the factor score. A factor score is a kind of optimally weighted mean. Since standardized measures such as correlations are not concerned with mean levels, the distinction between sums and means is irrelevant. When all correlations are positive, the correlation between a weighted mean and an unweighted mean score would be high. The difference is that the weighted score weighs having been arrested or been to prison higher than having cheated on an exam, as common sense would have it. In general, the rarer crime, the more serious it is and the higher factor loading.
I calculated a summed score. The correlation to the crime score, r=.75, is shown in the attached plot. Note that this value is not suited for analysis because the sample size is reduced to a mere n=2240. This is because a sum like this cannot take into account missing data, so one has to use only complete cases. If this wasn't the case, one could have used this variable with something like ordinal logistic regression. Nonetheless, I attach some plots of interest: 1) cognitive ability by crime sum and 2) distribution of crime sum scores. As can be seen, there is also a fairly linear negative relationship to see here. The correlation is -.13. I have added a discussion of this alternative approach to the paper.
Given all this, I think it is best to choose a single variable that best measures this underlying criminal factor that we are unable to measure well given the nature of the data. Since arrest history has a loading of ~1, this variable is the most suited.
I computed the group means for each of the crime variables as well as the group difference, shown below.
(The cognitive data are standardized, the above 0 values are due to selection. No everybody answers all the questions, smarter people tend to answer more.)
The gaps are not large, but they are roughly in the size order one would expect: Largest for prison, next largest for arrest (often means violence) and violence. Low for non-violence.
I have added this table to the paper.
I agree. The point with the study was to see whether the relationship could be seen i yet another context and to add 1 (or 5) more datapoints towards an eventual meta-analysis.
---
The files have been updated.
Thanks for reviewing.
The measurement of cognitive ability remains unclear. It would be helpful to know the actual 14-items used to create the scale, their content, and distribution. As the core variable in the analysis more detail is necessary.
The item questions, answer options, sample sizes and pass rates has been added in a table in the appendix (after references). Furthermore, a plot of the distribution of scores has been added to the main text.
The criminal behavior measure is, of course, a bit limited but prior research shows that even content limited measures usually have good statistical and content validity. Given the items included, which range form minor forms of deviance (cheating on tests), to more serious indicators of criminal behavior (arrest/prison), the expectation that a uniform factor would be extracted seems misplaced. Yes, criminals are generalists and this is confirmed with the inter-item correlations, but you really do not have a sufficient number of items to produce a reliable, single factor.
I would suggest creating a summated index and comparing the results against the results of found using only arrest. The additional variance in the index would allow you to more closely examine how cognitive scores perform across the index. Is a score of zero on the crime index associated with a cognitive score that is significantly higher than ascending index scores?
A factor is really just a common source of variance. In this case, all items share variance, so it is possible to estimate a common cause of this shared variance, i.e. a common or general factor.
A summed score would in fact be very similar to the factor score. A factor score is a kind of optimally weighted mean. Since standardized measures such as correlations are not concerned with mean levels, the distinction between sums and means is irrelevant. When all correlations are positive, the correlation between a weighted mean and an unweighted mean score would be high. The difference is that the weighted score weighs having been arrested or been to prison higher than having cheated on an exam, as common sense would have it. In general, the rarer crime, the more serious it is and the higher factor loading.
I calculated a summed score. The correlation to the crime score, r=.75, is shown in the attached plot. Note that this value is not suited for analysis because the sample size is reduced to a mere n=2240. This is because a sum like this cannot take into account missing data, so one has to use only complete cases. If this wasn't the case, one could have used this variable with something like ordinal logistic regression. Nonetheless, I attach some plots of interest: 1) cognitive ability by crime sum and 2) distribution of crime sum scores. As can be seen, there is also a fairly linear negative relationship to see here. The correlation is -.13. I have added a discussion of this alternative approach to the paper.
Given all this, I think it is best to choose a single variable that best measures this underlying criminal factor that we are unable to measure well given the nature of the data. Since arrest history has a loading of ~1, this variable is the most suited.
Prior research seemingly converges to show that average IQ differences between delinquents and non-delinquents is around .5sd. For adult offenders, average differences around 1sd are common. How much difference are we seeing with this sample of older, probably better educated and less criminal, men and women?
I computed the group means for each of the crime variables as well as the group difference, shown below.
non offender offender d
0.34 0.04 0.29
0.44 0.06 0.38
0.31 0.06 0.25
0.26 0.16 0.10
0.20 0.08 0.12(The cognitive data are standardized, the above 0 values are due to selection. No everybody answers all the questions, smarter people tend to answer more.)
The gaps are not large, but they are roughly in the size order one would expect: Largest for prison, next largest for arrest (often means violence) and violence. Low for non-violence.
I have added this table to the paper.
Low cognitive ability has repeatedly been linked with antisocial and criminal conduct. The relationship appears to be robust and not conditioned by the measurement of cognitive ability or problem behavior. In this sense, the findings presented here converge with other studies.
Lastly, the sample is clearly highly selected but I don’t see this as a problem. Yes, the sample is composed of older people who use the internet and a case can be made that subjects may be a bit brighter than comparatively situated others......but the effect of this type of selection would be to minimize variance on crime and cognitive measures. The results would thus be more conservative compared to samples that contain more variance in these measures.
I agree. The point with the study was to see whether the relationship could be seen i yet another context and to add 1 (or 5) more datapoints towards an eventual meta-analysis.
---
The files have been updated.
Some additional items I've come across:
q59919
Have you ever stolen a glass from a bar?
Yes.
No.
16747
q22569
Have you ever used a fake ID to do or acquire something you were legally barred from as a result of your age?
Yes
No
14067
q19928
Honestly, did you ever torture a cat, dog, or any other furry animal for pleasure?
Yes, but I regret it.
NO WAY!
Yeah, that's fun.
No, but I would do it.
10725
q39226
You stop to pick up a newspaper and notice that the coin-operated dispenser was not completely closed. No one is around so you have the opportunity to take a paper without paying. Which of the following would you do?
Pay for a paper and close the dispenser.
Steal a paper and close the dispenser.
Steal a paper and leave the dispenser open.
Steal all of the remaining papers.
2111
It's very difficult to find them because they are not properly tagged and there are many ways to be antisocial or criminal.
There are also a number of items related to littering such as:
q17017
Do you litter?
Often
Rarely
Never
35221
---
Thoughts on expanding items to cover more broadly of antisocial behavior, or keep them relatively focused on crime? The good thing about more items is that Jensen's method works better with more items and also when there are more diverse loadings.
q59919
Have you ever stolen a glass from a bar?
Yes.
No.
16747
q22569
Have you ever used a fake ID to do or acquire something you were legally barred from as a result of your age?
Yes
No
14067
q19928
Honestly, did you ever torture a cat, dog, or any other furry animal for pleasure?
Yes, but I regret it.
NO WAY!
Yeah, that's fun.
No, but I would do it.
10725
q39226
You stop to pick up a newspaper and notice that the coin-operated dispenser was not completely closed. No one is around so you have the opportunity to take a paper without paying. Which of the following would you do?
Pay for a paper and close the dispenser.
Steal a paper and close the dispenser.
Steal a paper and leave the dispenser open.
Steal all of the remaining papers.
2111
It's very difficult to find them because they are not properly tagged and there are many ways to be antisocial or criminal.
There are also a number of items related to littering such as:
q17017
Do you litter?
Often
Rarely
Never
35221
---
Thoughts on expanding items to cover more broadly of antisocial behavior, or keep them relatively focused on crime? The good thing about more items is that Jensen's method works better with more items and also when there are more diverse loadings.
Emil,
What strikes me about your paper is that even with less than perfect measures of cognitive ability and antisocial behavior, the resulting gaps/distributions are almost exactly what would be predicted. I suspect that if you increased the number of items in the measure of antisocial behavior, especially if those items tapped into more serious forms of conduct, the results would converge entirely with those found in other studies. Cognitive ability, regardless of how it is measured, is modestly correlated with and predictive of antisocial behavior.
Excellent job.
John
What strikes me about your paper is that even with less than perfect measures of cognitive ability and antisocial behavior, the resulting gaps/distributions are almost exactly what would be predicted. I suspect that if you increased the number of items in the measure of antisocial behavior, especially if those items tapped into more serious forms of conduct, the results would converge entirely with those found in other studies. Cognitive ability, regardless of how it is measured, is modestly correlated with and predictive of antisocial behavior.
Excellent job.
John
I'll see what I can find. There are some 2500 items to semi-manually look over, so one either has to read every single item (takes hours), or use search strategies that may not find all relevant items.
After looking over the items again, I will add all the criminal, antisocial and rule-breaking items I found, except for the drug related ones. These will be treated in a separate paper and they are special due to non-linear relationships where the occasional drug users have the highest IQs.
After that, re-do factor analysis. Probably with more items, can get a factor score with better distribution, so suitable for analysis.
I'll let you know when I've come up with a revision.
After looking over the items again, I will add all the criminal, antisocial and rule-breaking items I found, except for the drug related ones. These will be treated in a separate paper and they are special due to non-linear relationships where the occasional drug users have the highest IQs.
After that, re-do factor analysis. Probably with more items, can get a factor score with better distribution, so suitable for analysis.
I'll let you know when I've come up with a revision.
I have substantially revised the paper. It is now about twice as long but much better I think.
Let me know what you think.
Let me know what you think.
Again, I substantially revised the paper. The methods are now much better. I will solicit a new reply from the reviewers.
https://osf.io/qm8se/
https://osf.io/qm8se/
Again, I substantially revised the paper. The methods are now much better. I will solicit a new reply from the reviewers.
https://osf.io/qm8se/
Emil,
I've read the revised version and agree that it is much better. The results appear robust and insensitive to various specifications. Indeed, it is striking just how similar your results are to other studies that include much more detailed measures of cognitive ability. The effect sizes and correlational patterns are very similar to those produced in other studies.
John
Forgot to post earlier, but: Meisenberg indicated to me in email that he's ok with the version. Thus it has been published.
© Copyright 2024 OpenPsych - All Rights Reserved