Technotheoria

Most of the negative health effects of smoking do not come from nicotine itself. Very few people get addicted to nicotine in the form of gum. As such, nicotine could potentially be used as a stimulant or nootropic, the same way people use caffeine or amphetamines.

In the scientific literature, what caught my attention was a large meta-analysis on the effect of nicotine on cognitive performance. It was incorrectly executed, so I decided to run on my own. Several of the studies reported multiple effect sizes from multiple groups, so I tracked them. I divided the sample sizes by the number of effect sizes that were reported for each individual group, to get an estimate of the effective sample size of each study.

The aggregate effect size is about 0.38 SD, translating to 6 IQ points.

There is no dose-response relationship.

The quick-and-dirty meta-analytic effect is 0.33. The existence of it was extremely unambiguous (p < .0001).

Random-Effects Model (k = 84; tau^2 estimator: REML)
  logLik  deviance       AIC       BIC      AICc   
-53.6688  107.3375  111.3375  116.1752  111.4875   
tau^2 (estimated amount of total heterogeneity): 0.1178 (SE = 0.0320)
tau (square root of estimated tau^2 value):      0.3432
I^2 (total heterogeneity / total variability):   62.10%
H^2 (total variability / sampling variability):  2.64
Test for Heterogeneity:
Q(df = 83) = 255.1649, p-val < .0001
Model Results:
estimate      se    zval    pval   ci.lb   ci.ub      
  0.3250  0.0508  6.3939  <.0001  0.2254  0.4246  * 
---
Signif. codes:  0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1

I tested for moderators; none of them were significant in the aggregate: dose, smoking status, or domain of cognition.

Mixed-Effects Model (k = 84; tau^2 estimator: REML)
  logLik  deviance       AIC       BIC      AICc   
-46.0473   92.0946  116.0946  143.5801  121.2946   
tau^2 (estimated amount of residual heterogeneity):     0.1155 (SE = 0.0342)
tau (square root of estimated tau^2 value):             0.3399
I^2 (residual heterogeneity / unaccounted variability): 60.32%
H^2 (unaccounted variability / sampling variability):   2.52
R^2 (amount of heterogeneity accounted for):            1.91%
Test for Residual Heterogeneity:
QE(df = 73) = 216.0037, p-val < .0001
Test of Moderators (coefficients 2:11):
QM(df = 10) = 12.2553, p-val = 0.2683
Model Results:
                                           estimate      se     zval    pval    ci.lb   ci.ub    
intrcpt                                      0.3758  0.1788   2.1018  0.0356   0.0254  0.7262  * 
smoking_statusSmoker                         0.2308  0.1140   2.0247  0.0429   0.0074  0.4542  * 
dose_mg                                     -0.0150  0.0099  -1.5103  0.1310  -0.0344  0.0045    
domainAlerting attention-RT                  0.0017  0.2076   0.0082  0.9935  -0.4052  0.4085    
domainFine motor                            -0.1482  0.2320  -0.6390  0.5228  -0.6029  0.3065    
domainLong-term episodic memory-accuracy    -0.1982  0.2106  -0.9411  0.3467  -0.6109  0.2146    
domainOrienting attention-accuracy          -0.0517  0.2740  -0.1887  0.8503  -0.5888  0.4854    
domainOrienting attention-RT                 0.0655  0.2159   0.3034  0.7616  -0.3576  0.4887    
domainShort-term episodic memory-accuracy    0.0902  0.2133   0.4231  0.6722  -0.3278  0.5083    
domainWorking memory-accuracy               -0.3580  0.2436  -1.4694  0.1417  -0.8355  0.1195    
domainWorking memory-RT                     -0.0367  0.2328  -0.1577  0.8747  -0.4930  0.4196    
---
Signif. codes:  0 ‘’ 0.001 ‘’ 0.01 ‘’ 0.05 ‘.’ 0.1 ‘ ’ 1

I averaged all of the effect sizes of the individual publications to estimate publication bias. The regression test suggested that if there was any publication bias, it was likely negative — people wanted to report nulls more than positive effects.

#Meta-analysis with aggregated effect sizes
Random-Effects Model (k = 41; tau^2 estimator: REML)
  logLik  deviance       AIC       BIC      AICc   
-18.3017   36.6034   40.6034   43.9812   40.9278   
tau^2 (estimated amount of total heterogeneity): 0.1088 (SE = 0.0344)
tau (square root of estimated tau^2 value):      0.3299
I^2 (total heterogeneity / total variability):   75.42%
H^2 (total variability / sampling variability):  4.07
Test for Heterogeneity:
Q(df = 40) = 182.8701, p-val < .0001
Model Results:
estimate      se    zval    pval   ci.lb   ci.ub      
  0.3769  0.0616  6.1155  <.0001  0.2561  0.4978  * 
##### Regression test for publication bias
Regression Test for Funnel Plot Asymmetry
Model:     mixed-effects meta-regression model
Predictor: standard error
Test for Funnel Plot Asymmetry: z = -0.5439, p = 0.5865
Limit Estimate (as sei -> 0):   b =  0.4960 (CI: 0.0497, 0.9422

So, it does seem that the effect of nicotine on intelligence is real, just not exciting — maybe an increase of 3-8 IQ points. I should note that, if the effect is indeed generalised. The effects of it on real world outcomes and broader cognitive function haven’t been studied very well, so it’s difficult to tell how much the cognitive enhancement generalises.

• Nicotine, through smoking, increased performance on the Raven’s test by 6 IQ points, but the effect barely passed significance testing.
• Nicotine reduces insepction time by about 0.5 SD (good), and the study had low p-values in the 0.001-0.01 range. Seems legit.
• One study claims nicotine increases mental speed, but not performance on intelligence tests. I’m not inclined to put much weight into this finnding, as they had just 55 subjects. The effects of amphetamines on IQ/SAT type tests, for example, are very small (d = 0.05-0.15), but probably real.