On the Class Gradient in Risk-Taking

The notion of measurable general intelligence has been so badly abused by racialists since the birth of scientific racialism at the turn of the century that one is hard-pressed to use it at all. The racialist game is to fetishize raw differences is mean test scores between populations as evidence of differences in general ability between the races in order to explain world order. It is simply not realized in racialist circles that the pattern is no more than an explanandum demanding our attention. What explains population differences in mean test scores are interracial gaps in heath status. Like mean adult height, birth weight, mean adult body weight, and life expectancy, what mean disparities in mean population IQ scores measure is nothing other than heath status — dictated by the feedback loop between disease and nutrition.

The question is not only one of interracial differences in mental ability. After all, Murray and Herrnstein’s argument in The Bell Curve concerned not so much the racial order as the social order. The hegemony of Boasian antiracism became manifest in the reaction to the publication of The Bell Curve in 1994, when a thousand wags went after them for a chapter that digressed into replicating interracial differences in test scores. What they actually argued in the book was that the American class hierarchy was explained by biological differences between the elites and the lower classes. Innate biological differences is probably why, they argued, you’re reading this and that other fellow over there, watching the telly in his pajamas with a beer in his hand, is there, doing that. No one challenged them on this. What is one to make of this silence? Did Murray and Herrnstein’s critics agree with the authors on their accounting of the social order in terms of a biological hierarchy??

Murray and Herrnstein may be mistaken in their conception of the social order, but it is of some importance to control for natural variation in the population. My problem is that since we are proxying class by educational attainment, it may be the case that years of education capture natural differences in ability which in turn explain the class signal in years of schooling. If this were the case then the bias in arrest rates would not be due to class oppression but simply reflects the Darwinian sorting of society on ability. In this picture, the smart guys attain higher status because they behave better. If Murray and Herrnstein are right to paint this picture, then the class gradient of risk-taking propensities should vanish after controlling for cognitive ability. We test this prediction. We use the Armed Services Vocational Aptitude Battery reported by the National Longitudinal Survey to control for cognitive ability. We show that the class gradient is always robust: class is the most important conditioner of risk-taking even after controlling for general mental ability. We document a pattern of correlation between cognitive ability and risk-taking that is the opposite of that predicted by Murray and Herrnstein. We show that cognitive ability is positively associated with smoking, smoking weed, dealing weed, and using hard drugs.

Throughout, we formally report a pair of models that capture slightly different ways of thinking about class in America. One is a ladder: more years of schooling is recorded as higher status. Another is tripartite — as in the working class imaginary — the two diplomas splitting the population into three classes with the middling bulk of everyday people with their high school diplomas bracketed by the college-educated elites from above and the high school-dropouts from below. We will see how keeping both these frames up in the air at the same time will allow us to examine the stability properties of the class gradient.

We start with the probability of ever getting arrested. Only the sex and class gradients survive in the simpler model. In the tripartite class model, the dummy for Hispanic is significant but bears the wrong sign. IQ is not a significant predictor of arrest whether or not we control for race. The class gradient in the probability of ever getting arrested is so large that having no diploma subtracts and having a college-degree adds the equivalent of the sex effect. Being female has as much effect on the odds of getting arrested as 4 additional years of schooling under your belt.

Probability of getting arrested
Model IbseP
Intercept3.140.1920.000
Female-1.060.0650.000
Grade-0.250.0140.000
IQ0.030.0550.301
Black0.050.0830.272
Hispanic-0.140.0880.055
Model IIbseP
Intercept-0.160.0600.004
Female-1.090.0650.000
IQ-0.040.0530.226
No diploma1.100.0870.000
College-0.920.0830.000
Black0.000.0830.480
Hispanic-0.170.0880.027
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of arrest in 1997-2018. N=5,403 12-16 year olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

The puzzle of cognitive ability and risk-taking begins with the probability of dealing weed. We find that more cognitive ability is associated with a greater degree of risk-taking as captured by the actuarial probability of dealing weed. A standard deviation unit of cognitive ability adds as much as 2 years of additional schooling subtracts from the odds of dealing. Being black subtracts 2.6 years. A college degree shaves off 4.5 years.

Probability of dealing weed
bseP
Intercept1.360.2060.000
Female-0.650.0740.000
Grade-0.180.0150.000
IQ0.350.0620.000
Black-0.460.1000.000
Hispanic-0.130.0980.094
Deal
bseP
Intercept-0.940.0680.000
Female-0.660.0740.000
IQ0.330.0600.000
No diploma0.710.0970.000
College-0.810.0950.000
Black-0.490.1000.000
Hispanic-0.150.0980.062
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of dealing marijuana in 1997-2018. N=5,403 12-16 year-olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

Cognitive ability is not a significant risk factor for smoking in either regression model. Each additional year of schooling shaves off a sixth of a percent from the log odds of smoking. Being black shaves off three-fourths, being female or being Hispanic shaves off a quarter of a percentage point. Dropping out of high school adds nearly a whole percent. The underclass smokes like a chimney.

Smoke
bseP
Intercept3.590.1870.000
Sex-0.260.0630.000
Grade-0.160.0130.000
IQ0.050.0530.190
Black-0.740.0790.000
Hispanic-0.260.0870.001
Smoke
bseP
Intercept1.410.0660.000
Sex-0.280.0620.000
IQ0.000.0510.488
No diploma0.930.1100.000
College-0.570.0730.000
Black-0.780.0790.000
Hispanic-0.270.0870.001
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of smoking in 1997-2018. N=5,403 12-16 year olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

Higher cognitive ability is associated with marijuana use. One standard deviation higher cognitive ability adds as much as 3.6 additional years of schooling subtracts from the log odds of doping. How much is that? Being black shaves off the equivalent of 2.1 additional years of schooling; being female shaves off 2.5 years. Having no diploma adds 5.7 years. Having a college degree shaves off 3.4 years.

Use Marijuana
bseP
Intercept1.980.1640.000
Sex-0.250.0570.000
Grade-0.100.0120.000
IQ0.360.0490.000
Black-0.210.0730.002
Hispanic-0.100.0780.111
Use Marijuana
bseP
Intercept0.630.0580.000
Sex-0.260.0570.000
IQ0.340.0470.000
No diploma0.570.0870.000
College-0.340.0690.000
Black-0.230.0730.001
Hispanic-0.100.0780.093
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of using marijuana in 1997-2018. N=5,403 12-16 year olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

One standard deviation higher IQ adds as much as 2 years of additional schooling subtracts from the odds of using hard drugs. Being black shaves off as much as 6 years of school. Being female shaves off 1 year. A college degree naturally adds as much as 4 years worth.

Use hard drugs
bseP
Intercept1.220.1800.000
Sex-0.150.0630.008
Grade-0.140.0130.000
IQ0.300.0540.000
Black-0.850.0900.000
Hispanic-0.070.0830.216
Use hard drugs
bseP
Intercept-0.610.0620.000
Sex-0.160.0630.005
IQ0.270.0520.000
No diploma0.560.0890.000
College-0.610.0790.000
Black-0.880.0900.000
Hispanic-0.080.0830.166
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of using hard drugs in 1997-2018. N=5,403 12-16 year olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

We’ve seen that being black is negatively associated with risk-taking. Blacks are less likely to deal, smoke, dope, snort and shoot. But blacks keep gang. Conditioning on black adds 2 years worth of school to the odds of belonging to a gang. Being Hispanic adds about 3 years. A college degree subtracts nearly 7 years worth. One standard deviation of cognitive ability now subtracts 2 years worth of school.

Gang member
bseP
Intercept-0.410.3300.105
Sex-0.880.1240.000
Grade-0.170.0250.000
IQ-0.330.1070.001
Black0.360.1460.007
Hispanic0.460.1520.001
Gang member
bseP
Intercept-2.450.1120.000
Sex-0.880.1240.000
IQ-0.320.1040.001
No diploma0.480.1330.000
College-1.130.1970.000
Black0.340.1460.009
Hispanic0.430.1520.002
Source: NLSY97 (1997-2018), author’s computations. Response is logit probability of being a member of a gang in 1997-2018. N=5,403 12-16 year olds in 1997. Estimates in bold are significant at the 5 percent level. Response in log odds, Grade in number of years, IQ in standard deviation units, while the rest are zero-one random variables.

We have seen that the stability of the class gradient provides a natural unit of measuring effect sizes across these regressions — in years of school equivalents. The class gradient can be measured internally so to speak: by looking at the variation in the price of a college degree between the paired regressions documented above. Across the regressions, the mean internal price of a college degree is 4.4 years. What is impressive is the stability of the prices of sex and a college degree — always with the same sign as years of school (Grade). With the exception of gang membership, cognitive ability has the opposite sign from that which follows from Murray and Herrnstein’s account. The race dummies also have the same sign as years of school, with the exception of blacks for arrests (the black dummy was insignificant) and blacks and Hispanics for gang membership. The price of college and female is also especially high for the actuarial odds of gang membership: 5 years for female and 7 years for a college degree. Note specially that the odds for arrest for college and female are unexceptional. We find similar gradients for educational class and sex across a range of risks. The class gradient is a much larger story than police bias.

Price of conditioner in years of school units
PriceCollegeFemaleCognitive AbilityBlackHispanic
Arrest3.74.2-0.1-0.20.6
Deal4.63.7-2.02.60.7
Smoke3.51.6-0.34.61.6
Dope3.42.5-3.62.11.0
Drugs4.41.1-2.16.10.5
Gang6.85.32.0-2.2-2.8
Mean4.43.1-1.02.20.3
Ratio of coefficient for conditioner and years of school (Grade). 

Cognitive ability turns out to be positively associated with drug dealing and drug use but negatively associated with gang membership. The race dummies also switch signs when the response is gang membership. Instead of the sign flipping, the effect size of college and female increases for gang membership. Sex and class gradients are the only ones that are robust across the risks examined in this investigation. The class gradient is not even mildly attenuated after we control for cognitive ability, suggesting that perhaps cognitive ability is overrated. What we have found is more like the world of the ‘barefoot hedge-fund managers’ of Duflo and Banerjee. To first order, risk-taking is a monotonic function of status.

The cleaned dataset and replication code can be found on my GitHub.

Leave a Reply

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google photo

You are commenting using your Google account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s