Authority notebook

• Have subjects read or write essays about:
  • How smut is inappropriate for children.
  • How smut is inappropriate, period.
  • Some other moral topic, in a similarly condemnatory tone. (An obvious issue for the self-licensing concept which I don't think has been addressed is how domain-specific it is: does being Definitely Not Racist at time A license one to be more sexist at time B?)
• Have subjects read a story in which a child is denied access to smut.
• Have subjects read a story in which a child accesses smut but the consequences for the child are manipulated.
• Give subjects some kind of smut-reviewing task. You might explicitly ask what ages the material is appropriate for, or you could ask less specifically what sorts of audiences the material is appropriate for. (Such an experiment run in former decades would've resulted in a lot of people saying certain materials are inappropriate for women (or even servants), but smut protectionism nowadays seems exclusively age-based, at least in the West.)
• Ask subjects about what movies they would be willing to bring their own (assumedly hypothetical) children to, or what movies they would be willing to show in a classroom.
• Ask subjects to recommend some movies to a child.

• What media the subject selects for themself, for their private enjoyment (perhaps as compensation for participating in the experiment) or ostensibly as part of an unrelated experiment.
• The subject's attitudes about how permissible, safe, or moral consuming or producing smut is for the subject, for adults, or in general.
• Moral attitudes or behavior in a different domain.

Does condemning some sexual behaviors (like homosexuality) license other sexual behaviors?

The five kinds of smut:

• Sex
• Violence
• Drugs
• Profanity
• Gambling

Gambling is often seen as a matter of stupidity rather than morality (right?).

Drugs are tricky at least insofar as attitudes vary widely by drug. Cannabis is very tricky; there's too much disagreement between individuals and between cultures. Tobacco is increasingly seen as just plain stupid (right?). Alcohol of all drugs probably is the best example of an age divide. Drinking is celebrated in adults and condemned in adolescents.

People mostly agree that children need to be protected from media violence, but I think it's less common for adults to like violent media than to like drinking, if only because of the gender gap. People also mostly agree that children need to be protected from sex, but again, there's a gender gap for interest in pornography or weird sex acts or other particularly smutty varieties of sexual behavior.

j1 = go.real(model.ordinal.manip_only,
    ss(sb, tv == 2 & punish.nc != "Relent"))
gelman.diag(j1$samp)

In the below plot, the colored bars show the distribution of responses predicted by the model, and the black bars show the difference between the predicted response and the actual response.

distribution.plot(j1)

I judge this not too bad a fit, especially for the two bumps at $0 and $100. It's nice how the ordinal model has effectively captured the anchoring effect whereby there's a bump at $100 assumedly because 100 is a round number.

mean(j1$jsamp$b.protect_kids > 0)

So under this model, the probability is 87% that the manipulation increased fines.

Let's add a dichotomous predictor for gender, an interaction term for gender and the manipulation, and a linear predictor for (age - 18).

j2 = go.real(model.ordinal.mgai,
    ss(sb, tv == 2 & punish.nc != "Relent"),
    n.adapt = 1500)
gelman.diag(j2$samp)

distribution.plot(j2)

Surprisingly, this fancier model hasn't been able to match the overall proportion any better.

sapply(qw(b.protect_kids, b.female, b.protfem, b.age),
    function (v) mean(j2$jsamp[[v]] > 0))

The probability that b.protect_kids is positive is similar to that in the other model. Now let's try comparing the sizes of this model's different effects.

round(digits = 2, rbind(
    t(sapply(qw(b.protect_kids, b.female, b.protfem),
        function (v) qmean(j2$jsamp[[v]]))),
    "20 * b.age" = qmean(20 * j2$jsamp$b.age)))

The gender effect seems stronger than the manipulation:

mean(abs(j2$jsamp$b.female) > abs(j2$jsamp$b.protect_kids))

The age effect, on the scale of 20 years, seems roughly comparable.

There's other model-checking work I could do (e.g., checking the distribution of responses for women in particular), but this model basically agrees with the former with respect to the effect of the manipulation, which was the real interest, so there doesn't seem much of a point to trying harder to distinguish them.

In this study, the manipulation probably had the opposite of the intended effect. That is, subjects given the opportunity to protect teenagers by punishing businesses who sell tobacco to the underage were subsequently inclined to give a bigger fine to an adult who violated an open-container law. A plausible explanation is that subjects given the manipulation got into a more punitive mindset. In particular, they were reminded of how punishing people for breaking niggling laws can in fact be useful. One could even construe the effect as escalation of commitment.

The third-person effect per se is the phenomenon whereby people expect to be less affected by smut (and be less susceptible to other deleterious media effects) than other people are. A natural prediction is that the greater this gap, the more people will support censorship; the meta-analysis of Feng and Guo (2012) concluded that this prediction is indeed borne out, although the effect is weak. My idea of vicarious restraint is also an explanation for support for censorship. It differs from the third-person effect in that (a) the distinction is between "us" and "them" (adults vs. children) instead of "me" and "everybody else" and (b) there's the added idea that exercising censorship provides moral credentials for oneself or one's own group. In my own theorizing, I haven't emphasized media consumption in particular as much as the literature on the third-person effect has. But that's not too important because there's already literature on how people have a variety of self-serving perceptions compared to other people; it's just that this isn't usually called "the third-person effect".

Given these considerations, perhaps I can make my life easier by focusing on just one of these extensions of the third-person effect:

1. An "us"–"them" difference as opposed to a "me"–"everybody else" difference: do people have greater perceptions of susceptibility for outgroups than ingroups, or is the distinction of importance really just oneself versus others?
2. Censorship providing moral credentials. More trickily, mere reminders of censorship or other kinds of safeguards against harm that don't apply to oneself providing moral credentials. (Perhaps, for example, disallowing felons from buying guns makes nonfelonious people more comfortable buying guns.) Arguably, this prediction better captures the interesting part of vicarious restraint.

The first, I think, is implied by existing literature. The third-person perception gap seems to be enhanced for other people who are of further social distance; for example (quoting from Elder, Douglas, and Sutton (2006)), Cohen, Mutz, Price, and Gunther (1988) "found a linear increase in the TPE [third-person effect] as the comparison other changed from 'other Stanford students' to 'other Californians' and finally the 'public at large'". And Elder et al. (2006) found that for a pro-outgroup message, outgroups were perceived as more influenced than ingroups (and ingroups were perceived as more influenced than oneself).

The second can be further broken down:

1. If I protect some people from deleterious media effects, I get some moral credentials.
2. I can apply moral credentials towards consuming potentially harmful media (guilty pleasures?).
3. I can get moral credentials from other people's actions as opposed to my own. (Kouchaki (2011) showed that this is apparently true when it comes to racism, provided that the other people belong to the ingroup.)

And a lurking background proposition that's relevant (but hopefully not crucial) to the connection between 1 and 2, and which I mentioned earlier:

4. Moral credentials can cross domains. For example, being Definitely Not Racist at time 1 can credential one to be more sexist at time 2. Some support for this is provided by Mazar and Zhong (2010), who found that people given the opportunity to buy "green" products (vs. conventional products) were less generous in a subsequent dictator game and were more likely to lie and steal to get more money. Some less direct support is provided by Sachdeva, Iliev, and Medin (2009) and Jordan, Mullen, and Murnighan (2011), who found that subjects who described or recalled themselves positively or morally (compared to subjects who described themselves negatively) acted less prosocially.

And the extension to 3 that's most interesting to me is:

5. I can get moral credentials from a general arrangement of society (e.g., gun control) rather than an individual's actions (e.g., behaving responsibly with a gun).

Notice that in 5, compared to, say, 1, it is more difficult to describe the effect in terms of moral credits as opposed to moral credentials. If I didn't help enact gun control myself, and I don't particularly identify with whoever did, I don't get to feel I've done anything good. Rather, the existence of gun control suggests my use of guns is less dangerous than it might seem.

Well, I'd really like to run a study from which something interesting could be learned regardless of the outcome. This business of social-psychology experiments that are only interpretable if they turn out as expected is dangerous—it's one reason I've run a lot of subjects without having many manuscripts to show for it. Too often, the design of my studies hasn't allowed me to conclude anything when I've found the manipulation has had little or no effect or an effect in the wrong direction. It seems wiser to choose topics and designs such that a lack of an effect is somehow enlightening.

So let's put 5 aside for now. To me, it's the most interesting of the five hypotheses, but it's hard to falsify (I can't manipulate society; I can only remind people of society's arrangements), and if I did successfully falsify it, only I would care.

I'm leaning towards investigating 1. One trick is that it isn't clear whether a dependent measure regarding media consumption or more straightforwardly antisocial behavior is more likely to work. A media-consumption DV would depend on 2, whereas an unrelated DV would depend on 4.

Some vague ideas:

• The subject imagines they're head of some private transportation company and they get the opportunity to ban cigarette advertising. You could make this feel more altruistic by saying this will result in decrease in revenue. (Or don't, since this would be kind of a confound otherwise.)
• Pornographic moral credits: Do subjects who get to do something feministic (like the manipulation in Study 1 of Monin and Miller (2001)) thereafter admit to more pornography usage?

Now, the moral-credential effect in the context of vicarious restraint for media is that people should feel less susceptible to bad media effects (like racist media making you more racist) after exercising censorship concerning such media.

Perhaps a good general question: How does exercising censorship affect people's attitudes about vulnerability to the material being censored—vulnerability of other people and of oneself?

Even more generally: How does protecting people from some potential threat affect people's attitudes about vulnerability to said threats?

Let's pursue that first general question ("How does exercising censorship affect people's attitudes about vulnerability to the material being censored—vulnerability of other people and of oneself?"). If I get an effect, then I can worry about what follow-up studies to run in order to explain it. Let's not put too many eggs in one theoretical basket at this early stage.

The obvious design here is a study in which subjects do or do not get an opportunity to censor some dangerous media, and then I measure their beliefs about susceptibility to said media, for themselves and for a range of other agents or social groups. As before, I want subjects to be highly likely to take that opportunity if it's offered, so let's take something folks will be eager to censor. I'm inclined to use racist media, but I'm afraid I'll get a floor effect for perceptions of influence on oneself. Everybody will think themselves immune to racist propaganda. Hmm, so then perhaps it shouldn't be propaganda. Consider (a film about?) some kind of scientific study. I could say that various experts have denounced it as inaccurate—providing justification for censorship even among subjects who are inclined not to censor research reports—but also say that these inaccuracies are beyond lay ability to detect and ask the subject to make those vulnerability judgments for people who haven't heard of the experts' objections. Subjects will then be kind of obliged to not jump to the conclusion that they won't be influenced. Still, I'll need to make the scale wide, since I can expect the variability of interest for judgments about oneself to be clustered on the lower end.

One way I could help avoid a floor effect is use something like discrimination against blondes or left-handed people. I could still keep censorship agreement high using the objecting-experts business.

Another possible help: use susceptibility scales asking "How likely do you think it is that such-and-such would be influenced, even if only a little?" instead of "How much do you think such-and-such would be influenced?"

Now, along the lines of what I was saying before, what could I learn from finding no effect in this study? Well, I think the idea that the third-person effect causes censorship would be strengthened, because I would have evidence that the reverse causal relationship doesn't exist.

Okay, here's my idea; and by using deception, I can avoid a purely hypothetical manipulation, too. Subjects are told I'm gonna do this study on a whole bunch of kids that will involved having them consume some media supposedly showing that left-handed people are stupid and evil. I'll say that I'll debrief these kids but I'm conflicted about the whole procedure because I know that it's hard to disbelieve stuff (especially for kids? woo ageism!). The manipulation is that some subjects will also be asked if I should get around this problem by having the kids read analogous vignettes about aliens or something. The DVs will ask subjects how likely each party would be to be influenced by the original version.

On second thought, the subjects of the imaginary experiment should either be teenagers or unspecified people. Otherwise, I'm afraid subjects will infer that the imaginary stimulus is simplistic enough that adults (like the subject) would laugh it off.

Also, let's use albinism instead of left-handedness. The chance is of course much lower that subjects will be albino or know an albino personally than that they'll have analogous experience with left-handedness. And prejudice towards albinos has more face credibility than prejudice towards lefties, since albinos can look so strange.

dbeta.munu = defmacroq(mu, nu, expr =
    dbeta(mu * nu, (1 - mu) * nu))

rbeta.munu = function(n, mu, nu)
    rbeta(n, mu * nu, (1 - mu) * nu)

tv3.model1 = list(

    bugs = as.bugs.model(
       {# Priors

        vuln.s.b0 ~ dnorm(0, 100^-2)
        vuln.adult.b0 ~ dnorm(0, 100^-2)
        vuln.10yo.b0 ~ dnorm(0, 100^-2)
        b.censor.s ~ dnorm(0, 100^-2)
        b.censor.adult ~ dnorm(0, 100^-2)
        b.censor.10yo ~ dnorm(0, 100^-2)

        b.subject.sigma ~ dunif(.1, 100)
        for (s in 1 : N.subjects)
           {b.subject[s] ~ dnorm(0, b.subject.sigma^-2)}

        nu ~ dunif(2, 1000)

        # Likelihood

        for (s in 1 : N.subjects)
           {vuln.s.mu[s] <- ilogit(vuln.s.b0 + b.censor.s * censor[s] +
                b.subject[s])
            vuln.s[s] ~ .(dbeta.munu(vuln.s.mu[s], nu))
            vuln.s.pred[s] ~ .(dbeta.munu(vuln.s.mu[s], nu))

            vuln.adult.mu[s] <- ilogit(vuln.adult.b0 + b.censor.adult * censor[s] +
                b.subject[s])
            vuln.adult[s] ~ .(dbeta.munu(vuln.adult.mu[s], nu))
            vuln.adult.pred[s] ~ .(dbeta.munu(vuln.adult.mu[s], nu))

            vuln.10yo.mu[s] <- ilogit(vuln.10yo.b0 + b.censor.10yo * censor[s] +
               b.subject[s])
            vuln.10yo[s] ~ .(dbeta.munu(vuln.10yo.mu[s], nu))
            vuln.10yo.pred[s] ~ .(dbeta.munu(vuln.10yo.mu[s], nu))}}),

    mk.inits = function() list(
        vuln.s.b0 = runif(1, -2, 2),
        vuln.adult.b0 = runif(1, -2, 2),
        vuln.10yo.b0 = runif(1, -2, 2),
        b.censor.s = runif(1, -2, 2),
        b.censor.adult = runif(1, -2, 2),
        b.censor.10yo = runif(1, -2, 2),
        b.subject.sigma = runif(1, .1, 3),
        nu = exp(runif(1, log(2), log(500)))),
    jags.sample.monitor = qw(vuln.s.pred, vuln.adult.pred, vuln.10yo.pred))

tv3.model1.fake.p = list(
    vuln.s.b0 = -.75,
    vuln.adult.b0 = 0,
    vuln.10yo.b0 = 1,
    b.censor.s = -.25,
    b.censor.adult = .25,
    b.censor.10yo = .25,
    b.subject.sigma = .25,
    nu = 7)

tv3.model1.fake = function() with(tv3.model1.fake.p,
   {set.seed(10)
    N.subjects = 100
    censor = sort(rep(c(0, 1), len = N.subjects))
    b.subject = rnorm(N.subjects, 0, b.subject.sigma)
    vuln.s.mu = ilogit(vuln.s.b0 + b.censor.s * censor +
        b.subject)
    vuln.s = rbeta.munu(N.subjects, vuln.s.mu, nu)
    vuln.adult.mu = ilogit(vuln.adult.b0 + b.censor.adult * censor +
        b.subject)
    vuln.adult = rbeta.munu(N.subjects, vuln.adult.mu, nu)
    vuln.10yo.mu = ilogit(vuln.10yo.b0 + b.censor.10yo * censor +
        b.subject)
    vuln.10yo = rbeta.munu(N.subjects, vuln.10yo.mu, nu)

    data.frame(censor = logi2factor(censor, qw(NoOpp, Censor)),
        vuln.s, vuln.adult, vuln.10yo)})

j.tv3.m1.fake = cached(vulnjags(tv3.model1, tv3.model1.fake(), n.adapt = 2500))

gelman.diag(j.tv3.m1.fake$samp)

The value of b.subject.sigma is a tad high (I aim for 1.1 and below), but I didn't have this problem while fitting the real data below, so no big deal.

coda.qmean(j.tv3.m1.fake$samp, tv3.model1.fake.p)

Generally, things looks good (nu is surprisingly good), although the intervals are wider than I'd like.

j.tv3.m1.real = cached(vulnjags(tv3.model1,
    ss(sb, tv %in% 3:5 & censor != "Refuse"),
    n.adapt = 1000))
j.tv3.m1.real$d = vuln.jags2df(j.tv3.m1.real)

gelman.diag(j.tv3.m1.real$samp)

coda.qmean(j.tv3.m1.real$samp)

We can be reasonably sure the vuln.*.b0s are each on the right side of 0. That is, subjects feel (1) that they themselves are unlikely to be affected by the movie, but (2) typical adults and typical 10-year-olds are likely to be affected. I've replicated the third-person effect. Excellent.

Also, b.subject.sigma is large and nu is small. That is, individual differences and noise in vulnerability judgments are large. So I expected.

Unfortunately, we can't say much about the b.censor.*s yet. These parameters are of most interest, since they determine the effect of the manipulation on each judgment.

sapply(qw(b.censor.s, b.censor.adult, b.censor.10yo), function(v)
    mean(c(recursive = T, j.tv3.m1.real$samp[,v]) > 0))

At least b.censor.s is in the right direction: I want b.censor.s to be negative because exercising censorship should make one feel less vulnerable to media effects. I'm less sure what to predict about the other two.