Field Notes

concept

Pornography substitution hypothesis

StatusWorkingReviewdue UpdatedStands on6 sourcesSupportsnotes

The pornography substitution hypothesis holds that access to pornography reduces sexual offending by supplying an alternative outlet for people who might otherwise offend. It is the third of three possible readings of the association between pornography and sexual aggression, and the one most often left out of the debate.

The three are not rivals in the usual sense, because they can be true of different people or at different levels of measurement at the same time:

Reading Claim Where its evidence comes from
Socialization Consumption produces aggression Laboratory and correlational studies
Selection Aggression-prone people seek the content Inference from correlational designs; untested directly
Substitution Consumption displaces offending Aggregate and quasi-experimental crime data

Three mechanisms, often conflated

“Substitution” covers three different claims, and keeping them apart matters because they predict different things and fail differently.

Satiation is drive reduction: consumption temporarily lowers the value of the reinforcer, so the urge is discharged rather than acted on. It is temporary and returns with the drive. It should not be confused with habituation, which reduces response to a repeated stimulus, lasts longer, and is what the tolerance-escalation worry invokes. The substitution argument needs satiation; its critics usually argue from habituation.

Voluntary incapacitation requires no psychology at all. Time spent consuming media is time not spent offending, which is routine-activities theory rather than drive theory. It survives even if satiation is entirely false.

Catharsis, the hydraulic model in which aggression accumulates and must be vented, is the weakest of the three and the one usually attacked. Neither of the others depends on it, so refuting catharsis — which the literature has largely done — leaves the substitution case untouched.

Why the substitution literature exists at all

The economic logic is straightforward and does not require anyone to think pornography is good. Pornography is a complement to masturbation and consensual sex, and those are substitutes for coercive sex. If the second relation is stronger than any direct incitement effect, cheaper pornography reduces offending on net.

That prediction has been tested repeatedly by exploiting sudden changes in availability.

Legalization studies. Kutchinsky examined Denmark’s 1965 liberalization and found registered sex offences falling, most clearly child molestation. He later compared the United States, Denmark, Sweden, and West Germany from 1964 to 1984, a period in which pornography went from scarce to abundant, and found that in none did rape rise faster than non-sexual violent crime. Diamond, Jozifkova, and Weiss found the same pattern in the Czech Republic after 1989, and argued explicitly for a displacement function.1

Technology-shock studies. Kendall used the arrival of the internet across United States states from 1998 to 2003 and found a ten-percentage-point rise in access associated with roughly a 7.3 percent fall in reported rape, with no comparable effect on other violent or property crime.2 The effect concentrated in men aged 15 to 19 — those living at home, for whom the non-pecuniary price of pornography fell furthest. A German study using broadband roll-out found a substitution effect on overall sex crime, driven by child sexual abuse, with rape estimates negative but not significant.3

Why it is not settled

A contrary quasi-experiment exists, and it is itself contested. Bhuller, Havnes, Leuven, and Mogstad used a Norwegian broadband roll-out with limited funding as an instrument and found internet use associated with a substantial increase in rape and other sex crimes — in reports, and in the published version in charges and convictions.4 They set out the same three mechanisms this note distinguishes, report that none of their analyses suggests the relationship is driven by reporting behaviour, and conclude for a direct effect on propensity, “possibly as a result of increased consumption of pornography.”

That conclusion should not be taken at face value, for three reasons that have nothing to do with disliking it.

The implied magnitude is not credible. Gelman pointed out in 2011 that at the reported coefficients — a one-point rise in internet use associated with 0.14 more rapes and 0.16 more child sex abuse cases per 100,000, against base rates of roughly 15 and 25 — the Norwegian internet expansion would account for something like half of all rapes and a third of child sexual abuse in the country.5 An estimate that large from one instrument is more likely to indicate a specification or identification problem than a discovered cause of half a nation’s sexual violence.

Nordic reporting is a documented special case. The Nordic paradox is the finding that the most gender-equal countries record Europe’s highest reported intimate-partner violence, attributed to lower tolerance and greater willingness to report. Norway is therefore the setting where a reporting confound should be considered most seriously, not least, and an author’s statement that they checked for it is not the same as a reader having seen the check.

The comparable German study found reporting effects real. Diegmann, using broadband roll-out in Germany, found substitution rather than facilitation and found evidence of a reporting effect, concluding that the true consumption effect was stronger than the raw estimate suggested. Two studies of the same treatment reached opposite substantive conclusions and opposite conclusions about whether reporting mattered.

Set against that, the finding is not nothing. It survived refereeing at a leading economics journal, and the instrument is plausibly exogenous.

The natural remaining defence is that the effect appears in charges and convictions and not only in reports, which a pure reporting artefact should not produce. That defence is weaker in this jurisdiction than it looks, and the vault’s own criminal-justice work is what weakens it.

A conviction is an institutional output, not a measurement of what happened, which is the general property of a Correction channel: a system’s own products cannot establish its own error rate. For Nordic sexual offences there is more than the general worry. Evidence patterns in rape and unlawful-threat cases found that in the Swedish judgments examined, the same evidentiary pattern — a complainant account plus testimony about their later statements or emotional state — was followed by conviction in 90 percent of rape cases and 18 percent of unlawful-threat cases, which the authors could not easily justify and considered might reflect a lower operative threshold in rape. Fällda för sexövergrepp documents the same concern case by case.

If the threshold for conviction in sexual-offence cases can move with social attitude, then reports, charges, and convictions are not three independent measurements. They are three outputs of one system whose sensitivity was changing across exactly the years Bhuller studied, in the region where that change was most pronounced. Rising convictions are then consistent with rising true incidence and with a system convicting more readily on contested evidence, and the study’s design does not separate them.

The Swedish material is not Norwegian, and the two systems are adjacent rather than identical. But it is evidence that the assumption doing the work here — that a conviction count tracks underlying incidence — is one the vault has specific reason to doubt in this setting.

The honest position is that the internet-shock literature is mixed rather than lopsided. On rape specifically: Kendall finds a decrease in the United States, Diegmann finds no significant effect in Germany, Bhuller finds an increase in Norway. The legalization studies are a different treatment and cannot be added to that tally.

One frame that might reconcile the divergence is that the internet does different things depending on what it is used for: substitution where it supplies non-contact material, facilitation where it supplies contact, by lowering an offender’s search costs for victims. That is a hypothesis worth testing and not a finding, and it cannot be supported by citing Bhuller and colleagues, who examined that mechanism and preferred another.

The ecological fallacy cuts both ways. Ferguson and Hartley’s critique of the aggregate literature is the sharpest statement of the problem, and it is notable that it comes from authors otherwise sceptical of pornography harms: a correlation between population-level availability and population-level crime cannot establish what happens inside any individual.6 Aggregate substitution findings are subject to exactly the objection that defeats aggregate harm findings. Neither licenses an individual-level claim.

Reported crime is a contaminated outcome. Every study here measures reported offences, over decades in which reporting propensity, legal definitions, and policing all changed considerably, and often in the same direction as the treatment.

How the levels might reconcile

The Confluence model of sexual aggression offers a reconciliation that requires no side to be wrong. If pornography raises risk only for men in whom other risk factors already converge, and functions as an outlet for the large majority in whom they do not, then individual-level studies could correctly find a positive association within the high-risk subgroup while population-level studies correctly find a net reduction, because the high-risk group is small.

On that account the question “does pornography cause or prevent sexual aggression” has no population-level answer, and the policy-relevant quantity is the interaction rather than the main effect. That is a harder thing to measure and a less satisfying thing to legislate from, which may be why the debate keeps returning to a main effect that the evidence does not support in either direction.

Bearing on the vault

Does pornography use cause adolescent sexual aggression records that the individual-level association survives only for violent content and only weakly. This note supplies the third arrow that question needs, and the combination narrows what any intervention can claim: restricting access is not obviously protective even on the harm literature’s own terms, because the same restriction removes whatever substitution exists.

  • Read Bhuller and colleagues’ reporting analysis in the full text rather than its abstract, and record what individual reporting data it actually used. The vault currently has the authors’ summary of their own check.
  • Find whether anyone has answered Gelman’s magnitude objection or shown the scatterplot he asked for, since an effect accounting for half of Norwegian rapes is the single largest reason to doubt the estimate.
  • Determine whether any study separates violent from nonviolent pornography in an availability-shock design, since the individual-level evidence says that is the distinction that carries the association.

  1. Milton Diamond, Eva Jozifkova, and Petr Weiss, “Pornography and Sex Crimes in the Czech Republic”, Archives of Sexual Behavior 40 (2011). ↩

  2. Todd Kendall, “Pornography, Rape, and the Internet”, Toulouse School of Economics conference paper, 2007. ↩

  3. “The internet effects on sex crime offenses — evidence from the German broadband internet expansion”, International Journal of Law and Psychiatry (2019). ↩

  4. Manudeep Bhuller, Tarjei Havnes, Edwin Leuven, and Magne Mogstad, “Broadband Internet: An Information Superhighway to Sex Crime?”, Review of Economic Studies 80, no. 4 (2013): 1237–1266. ↩

  5. Andrew Gelman, “Is the internet causing half the rapes in Norway? I wanna see the scatterplot”, 17 May 2011, working the reported coefficients against Norwegian base rates. ↩

  6. Ferguson and Hartley, “Problems with Aggregate Data and the Importance of Individual Differences in the Study of Pornography and Sexual Aggression,” comment on Diamond, Jozifkova, and Weiss (2010). The same authors’ later meta-analysis is Ferguson and Hartley pornography aggression meta-analysis. ↩

Built on 6 sources (2 archived here, 4 external).

Working out connections…