Who has an inner voice?

Young people, effective altruists, mentally ill people, and trans people.

The evidence in this article seems to support the theory that inner speech is related to measurements of neurodivergence, such as alternative sexualities or mental illnesses. Effective altruists have pronounced inner speech, so it is possible that inner speech relates to prosociality (or at least, attempted prosociality).

I tested whether there are non-monotonic effects, where having a null inner voice or a highly prominent one lead to similar outcomes. It appears that the effect of the inner voice is monotonically related to its strength, in other words, the relationship between inner voice strength and outcome variables appears to be monotonic. Scott Alexander’s 2022 survey was used as a dataset.

Here is the classification scheme I used to classify people’s inner speech:

No, I basically never hear words in my head → 0

Yes, I can talk things over to myself if I need to → 1

Yes, it narrates almost everything I do → 2

I then normalized the variable at mean=0, SD=1. Using this classification scheme, this is the response histogram:

Zero-order associations:

Notes:

• ‘inner’ is the inner speech variable standardized at mean=0 | SD=1 - so the differences between groups can be thought of as effect sizes.

• p-values in aggregate tests were calculated with the kruskal-wallis test because it is less sensitive to departures from normality

• One issue with this data is that the measure of inner speech (just asking one question) is probably not very reliable or valid, and can be confounded by social desirability. Because of this, the true relationship between quantity of inner speech and outcomes should be reasonably larger in the real world than in this data.

• This variable measures strength and frequency of inner speech in the same variable, which is a reduction in dimensionality that could be important.

Race (p = .13): (note: this sample is selected for intelligence)

Gender (p < .001, still p < .001 when excluding other):

Moral Views (p < .001):

Depression (p < .001):

Anxiety (p < .001):

OCD (p < .001):

Eating Disorder (p < .001):

PTSD (p < .001):

Alcoholism (p = .36):

Drug Addiction (p < .01):

Borderline (p < .01):

Bipolar (p < .05):

Autism (p < .001):

ADHD (p < .001):

Schizophrenia (p = .26 | though based on my priors this is probably reflective of a true difference):

Mental illness MIRT score (p < .001)

MIRT factor loadings

Mental illness MIRT score by strength of internal monologue

Gender (p < .001):

Sexual orientation (p < .001):

Relationship style (p < .01):

Profession (p < .05):

Political Ideology (p < .001):

Political Spectrum (p < .001, 1→ left | 10→right):

Religious Beliefs: (p < .01):

Having long COVID (p = .3)

Taking the vaccine: p = .38

Correlations:

Open browser tabs: (p = .12)

STEM interest (p = .46)

Age: r = -.21, p < .001

Unhappiness: r = .089, p < .001

Mental illness: r = .12, p < .001

Trust in the media: r = -0.008, p = .49

IQ: r = -0.029, p=0.22 (very selected sample with mean=140)

BMI: r = -0.029, p=0.026

Having migraines: r = 0.03, p = 0.015

Having thought people are out to get you: r = 0.058, p < .001

Feeling bad about being circumcised: r = -0.086, p < .001

Self-reported ability to save money: p = .76

Self-rated episodic memory: r = 0.043, p < .001

Amount of unread emails: p = .36

Controlling for variables:

Age, Gender, and sexual orientation were judged as the most relevant confounders; every other variable that passed significance was tested for whether the association still held up after controlling for those 3. Here are the results:

Passes p < .01:

• Moral Views (lower for no answer)

• Profession (finance and IT people lower in inner speech)

• Unhappiness

• Mental illness

• Self-reported episodic memory

• Feeling bad about being circumcised

• Religious views

• Having thought a group of people are out to get you

• Being EA (relationship reduced by 30-40% or so by covariates, significant even after controlling for political views)

Passes p < .05:

• Political Ideology (p = .03 for NRx being higher)

Doesn’t pass:

• Relationship style

• Political spectrum (linear)

• Political spectrum (as a factor)

• BMI

• Having migraines

Non-linear effects

Axis were inverted to determine whether there is a non-linear effect of inner voice on outcomes, that is, null inner voices and prominent ones lead to the same outcomes. Same methodology used as before. After trying 3 different variables, it appears that the effect is linear.

Group -1.84 is no inner voice, -0.125 is occasional inner voice, while 1.59 is prominent inner voice.

Unhappiness (p < .001, linear):

Hearing random voices (1-5 likert) (p < .001, linear):

Thinking people are plotting to kill you (1-5 likert) (p < .001, linear):

Rcode

Personally, I have a strong, but infrequent inner voice. It spontaneously activates when I am doing nothing, but if I am actively doing something it takes effort to use. I recently tried to read an article by forcing myself to audibly hallucinate every word, but it was exhausting and I gave up. However, while my imagination/inner voice is active, it is highly powerful. Sometimes, I even cry when I imagine sad scenarios, laugh if I imagine something funny, or feel comforted when I picture myself hugging a family member. In summary, you could describe my inner monologue as infrequent, but strong.

I find it incredibly difficult to think in my own voice; I have to speak before doing it. I naturally think in the voices of other people I recently listen to, though I am the one directing the internal monologue. I also rarely think before making normal decisions. I was once interrogated by my old bosses on why I broke some stupid rule; I replied by saying that I put no thought in the action. I must say they were rather surprised by this response.

edit 1 - clarified in the graphs that the prominent inner voice narrates almost everything the person does. graph changed to percent notation for consistency with the title and the axis.