Why does Duolicious ask me irrelevant questions?

Gold diggers are less likely to fold their toilet paper before wiping, weebs are more open to dating outside their ethnicity, and home owners fall in love less easily.

How do we know all this? Because we ask irrelevant questions, of course!—And lots of them. Right now we have 24,534,316 answers from our 98,406 members, who all answered part of our personality quiz. Surprisingly, even answers to our sillier questions hint at answers to serious ones. If you ask someone an irrelevant question like "Do you think it's acceptable to wear pajamas to the grocery store?" then you also have a surprisingly strong hint at their answer to: "Do you think the police discriminate against Black people or other minorities?"

In this blog post, we'll list some of our sillier personality questions and the (at times deeper) answers they're correlated with.

Silly questions, serious business

Before we get to the questions, we should explain how we analyzed our users' responses. We wanted to know: If someone gives an answer to one question, how predictable are their answers to the other questions? To figure that out, we computed the Pearson correlation coefficient (PCC) for each pair of answers. The PCC is a number between -1 and 1. If a pair of questions has a PCC near 1, that means everyone on Duolicious always answers those questions the same way; a PCC of -1 means everyone always answers them the opposite ways (i.e. people who answered "yes" on one question always answer "no" on the other); and a PCC of 0 means there's no relationship between the two questions. Here's what the correlation coefficients look like when plotted for the first 300 or 400 questions.

You can see a red line which runs diagonally—That just says the obvious: That the correlation between a question and itself is always 1—But then there's also vertical and horizontal bands of intensity to the colors. What they roughly show is that some questions are more revealing than others, and therefore give a stronger hint as to how someone will answer the other questions. But even the lightly colored bands (i.e. the "irrelevant") questions tell us something, which you can see because there's no vertical or horizontal band which is purely one color.

Just get to the questions already!

Alright, so we cherry-picked some questions from our question bank. For each question, we show the 5 questions which are most likely to be answered the same way, and the 5 questions which are most likely to be answered differently. The Pearson correlation coefficient (r) is given for each question.

We'll start with a couple of predictable questions to illustrate how this works.

Do you pray?

Alright, so people who pray are more likely to be religious and less likely to support gay marriage. No surprises so far. How about another predictable question?

Could you date an OnlyFans model?

Someone who'd date an OnlyFans model doesn't find virgins attractive in particular? Who knew! Okay, here's a more interesting one:

Is it acceptable to put ketchup on eggs?

So, if you've got progressive views about ketchup and eggs, you're probably also pretty open-minded about french fry and pizza consumption techniques. But odds are you're also less selfish and more open to public displays of affection than people who believe ketchup and eggs should be kept separate. Surprising, no?

Would you date a robot if they had a great personality?

In other words, people who'd date a robot are less likely to believe in traditional gender roles. You know who else doesn't believe in gender roles? People who are into other people who talk to their plants:

Would you date someone who talks to their plants?

Are most people idiots?

Do you fold your toilet paper before wiping?

People who fold their toilet paper are slightly more likely to believe in separation of church and state and are a bit less likely to be gold diggers.

Do you like visiting art museums?

Virgins don't hang out in art museums.

Would you date someone who insists on celebrating obscure holidays, like "International Talk Like a Pirate Day" or "International Tuba Day"?

Do you think having a pet turtle could be exciting?

Wannabe pet turtle owners are the voice of reason at parties, apparently.

If you discovered a friend's significant other on a dating app, would you stay out of it rather than telling your friend?

Do you catch feels too easily?

People who catch feels would still choose their partner over their job even though they're typically dissatisfied with how much money they make. Cute!

Are you a weeb?

Have you ever tried to make an ex jealous?

Don't tell anyone they have spinach in their teeth if you hear they're trying to get back at their ex.

Do you have any younger siblings?

Do you have any older siblings?

Based on this, younger siblings appear to be less progressive than older siblings. We looked this up to see if social science studies agreed with us on this, and it turns out that the studies are mixed. Our best guess for the reason is that this is one of those associations that changes over time. A (somewhat) well-known example of another association which changed over time is anti-vaccine sentiment. Around the 1990s, it was mostly leftists who disliked vaccines. Though the association reversed over time, and now it's the rightists who dislike them.

Are you eccentric?

Do you think it's okay to tell a small lie in order to win a board game?

Don't trust people who lie when playing board games.

Were your parents overprotective?

Have you ever been gaslighted?

Have you ever cried during a particularly touching commercial?

Should all traffic lights play music to make waiting more enjoyable?

Could you fall for someone who uses Yoda quotes to give life advice?

Should superheroes be held accountable for collateral damage caused during their heroics?

Is it acceptable to have an imaginary friend as an adult?

Was your mom a nice lady?

Do you think Bigfoot is just a bear who really loves to play hide and seek?

Could you be wooed by someone who belts out pop ballads in karaoke without a hint of shame?

Have you ever left a mean comment on social media?

Do you like it when store attendants ask, "do you need help with anything? 😊"

Does the location of a tattoo affect its attractiveness?

Should flag burning be illegal?

Should long-term partners know each others' phone passwords?

Imagine you only have 7 white socks and 9 black socks in a drawer, and you're completely in the dark. If you only pull out 3 socks, are you sure to have a pair whose colors match?

What's "spurious correlation?"

Shhh... Don't worry about that.

But seriously, it wouldn't be surprising if some of the associations in this article really are spurious, especially because lots of the correlation coefficients aren't very big. That said, even spurious relationships can be useful to observe. In any case, hopefully this post's inspired you to figure out your next date's political stance by asking them how they wear their pajamas.

Appendix

#!/usr/bin/env python3

import pandas as pd
import psycopg
import seaborn as sns
import matplotlib.pyplot as plt
import json
import os
import numpy as np

# Database connection parameters
db_params = {
    'dbname': 'duo_api',
    'user': 'postgres',
    'password': 'password',
    'host': 'localhost',
    'port': '5433',
}

# File paths
correlation_matrix_file = 'correlation_matrix.csv'
correlation_dict_file = 'question_correlations.json'

# Check if the correlation matrix already exists
if os.path.exists(correlation_matrix_file):
    # Load the saved correlation matrix
    correlation_matrix = pd.read_csv(correlation_matrix_file, index_col=0)

    # Convert the index and columns to integers to ensure consistency
    correlation_matrix.index = correlation_matrix.index.astype(int)
    correlation_matrix.columns = correlation_matrix.columns.astype(int)
else:
    # Connect to your PostgreSQL database
    conn = psycopg.connect(**db_params)

    # Sample the data
    query = """
    SELECT
        answer.person_id,
        answer.question_id,
        answer.answer::int * 2.0 - 1.0 AS answer -- Will be -1 or +1
    FROM
        answer
    WHERE
        answer.answer IS NOT NULL
    """

    df = pd.read_sql_query(query, conn)
    conn.close()

    # Pivot the DataFrame
    pivot_df = df.pivot(
        index='person_id',
        columns='question_id',
        values='answer')

    # Compute the correlation matrix between questions
    correlation_matrix = pivot_df.corr(method='pearson')

    # Save the correlation matrix to a CSV file for future use
    correlation_matrix.to_csv(correlation_matrix_file)

# Initialize a dictionary to store the top 10 most and least correlated
# questions for each question
correlation_dict = {}

# Iterate through each question to find the top 10 most and least correlated
# questions
for question in correlation_matrix.columns:
    # Drop the question itself to avoid self-correlation
    correlations = correlation_matrix[question].drop(labels=[question])

    # Find the top 10 most correlated questions with their correlation
    # coefficients
    top_10_most = correlations.nlargest(10).items()

    # Find the top 10 least correlated questions (most negative correlations)
    # with their correlation coefficients
    top_10_least = correlations.nsmallest(10).items()

    # Store the results in the dictionary with correlation coefficients
    correlation_dict[question] = {
        'most_correlated': [
            {'question_id': qid, 'correlation': corr}
            for qid, corr in top_10_most],
        'least_correlated': [
            {'question_id': qid, 'correlation': corr}
            for qid, corr in top_10_least]
    }

# Write the correlation_dict to a JSON file
with open(correlation_dict_file, 'w') as json_file:
    json.dump(correlation_dict, json_file, indent=4)

# Optional: Display a heatmap of the correlation matrix
plt.figure(figsize=(20, 16))
sns.heatmap(
    np.sign(correlation_matrix) * np.sqrt(np.abs(correlation_matrix)),
    cmap='coolwarm',
    annot=False,
    fmt=".2f",
    linewidths=.0)

# Set the title and axis labels
plt.title('Question Correlation Matrix', fontsize=20)
plt.xlabel('Questions')
plt.ylabel('Questions')

# Get the number of questions (i.e., columns/rows in the correlation matrix)
num_questions = correlation_matrix.shape[0]

# Set the ticks for x and y axes to show only every 100th label
tick_positions = range(0, num_questions, 100)
labels = [str(i) for i in tick_positions]
plt.xticks(tick_positions, labels=labels, rotation=90)
plt.yticks(tick_positions, labels=labels)

# Show the plot
plt.show()