Prediction of Dichotomous Variables.

Fig. 2 shows the prediction accuracy of dichotomous variables expressed in terms of the area under the receiver-operating characteristic curve (AUC), which is equivalent to the probability of correctly classifying two randomly selected users one from each class (e.g., male and female). The highest accuracy was achieved for ethnic origin and gender. African Americans and Caucasian Americans were correctly classified in 95% of cases, and males and females were correctly classified in 93% of cases, suggesting that patterns of online behavior as expressed by Likes significantly differ between those groups allowing for nearly perfect classification.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 2.

Prediction accuracy of classification for dichotomous/dichotomized attributes expressed by the AUC.

Christians and Muslims were correctly classified in 82% of cases, and similar results were achieved for Democrats and Republicans (85%). Sexual orientation was easier to distinguish among males (88%) than females (75%), which may suggest a wider behavioral divide (as observed from online behavior) between hetero- and homosexual males.

Good prediction accuracy was achieved for relationship status and substance use (between 65% and 73%). The relatively lower accuracy for relationship status may be explained by its temporal variability compared with other dichotomous variables (e.g., gender or sexual orientation).

The model’s accuracy was lowest (60%) when inferring whether users’ parents stayed together or separated before users were 21 y old. Although it is known that parental divorce does have long-term effects on young adults’ well-being (28), it is remarkable that this is detectable through their Facebook Likes. Individuals with parents who separated have a higher probability of liking statements preoccupied with relationships, such as “If I’m with you then I’m with you I don’t want anybody else” (Table S1).

Prediction of Numeric Variables.

Fig. 3 presents the accuracy of predicting numeric variables as expressed by the Pearson product–moment correlation coefficient between the actual and predicted values. The highest correlation was obtained for age (r = 0.75), followed by density (r = 0.52) and size (r = 0.47) of the Facebook friendship network. Closely following were the personality traits of “Openness” (r = 0.43), “Extraversion” (r = 0.40), and “Intelligence” (r = 0.39). The remaining personality traits and SWL were predicted with somewhat lower accuracy (r = 0.17 to 0.30).

• Download figure
• Open in new tab
• Download powerpoint

Fig. 3.

Prediction accuracy of regression for numeric attributes and traits expressed by the Pearson correlation coefficient between predicted and actual attribute values; all correlations are significant at the P < 0.001 level. The transparent bars indicate the questionnaire’s baseline accuracy, expressed in terms of test–retest reliability.

Psychological traits are examples of latent traits (i.e., traits that cannot be measured directly). As a consequence, their values can only be measured approximately, for example, by evaluating responses to questionnaires. The transparent bars presented in Fig. 3 indicate the accuracy of the questionnaires used as expressed by their test-retest reliabilities (Pearson product–moment correlation between the questionnaire scores obtained by the same respondent at two points in time). The correlation between the predicted and actual Openness score (r = 0.43) was very close to the test–retest reliability for Openness (r = 0.50). This indicates that for the Openness trait, observation of the user’s Likes is roughly as informative as using their personality test score itself. For the remaining traits, prediction accuracies correspond to roughly half the questionnaire’s test-retest reliabilities.

The relatively lower prediction accuracy for SWL (r = 0.17) may be attributable to the difficulty of separating long-term happiness (29) from mood swings, which vary over time. Thus, although the SWL score includes variability attributable to mood, users’ Likes accrue over a longer period and, so, may be suitable only for predicting long-term happiness.

Amount of Data Available and Prediction Accuracy.

The results presented so far rely on individuals for which between one and 700 Likes were available. The median number of Likes was 68 per individual (IQR, 152). Therefore, what is the expected accuracy given a random individual and how does prediction accuracy change with the number of observed Likes? Using a subsample (n = 500) of users for whom at least 300 Likes were available, we ran predictive models based on randomly selected subsets of n = 1, 2, …, 300 Likes. The results presented in Fig. 4 show that even knowing a single random Like for a given user can result in nonnegligible prediction accuracy. Knowing further Likes increases the accuracy but with diminishing returns from each additional piece of information.

• Download figure
• Open in new tab
• Download powerpoint

Fig. 4.

Accuracy of selected predictions as a function of the number of available Likes. Accuracy is expressed as AUC (gender) and Pearson’s correlation coefficient (age and Openness). About 50% of users in this sample had at least 100 Likes and about 20% had at least 250 Likes. Note, that for gender (dichotomous variable) the random guessing baseline corresponds to an AUC = 0.50.

Predictive Power of Likes.

Individual traits and attributes can be predicted to a high degree of accuracy based on records of users’ Likes. Table S1 presents a sample of highly predictive Likes related to each of the attributes. For example, the best predictors of high intelligence include “Thunderstorms,” “The Colbert Report,” “Science,” and “Curly Fries,” whereas low intelligence was indicated by “Sephora,” “I Love Being A Mom,” “Harley Davidson,” and “Lady Antebellum.” Good predictors of male homosexuality included “No H8 Campaign,” “Mac Cosmetics,” and “Wicked The Musical,” whereas strong predictors of male heterosexuality included “Wu-Tang Clan,” “Shaq,” and “Being Confused After Waking Up From Naps.” Although some of the Likes clearly relate to their predicted attribute, as in the case of No H8 Campaign and homosexuality, other pairs are more elusive; there is no obvious connection between Curly Fries and high intelligence.

Moreover, note that few users were associated with Likes explicitly revealing their attributes. For example, less than 5% of users labeled as gay were connected with explicitly gay groups, such as No H8 Campaign, “Being Gay,” “Gay Marriage,” “I love Being Gay,” “We Didn’t Choose To Be Gay We Were Chosen.” Consequently, predictions rely on less informative but more popular Likes, such as “Britney Spears” or “Desperate Housewives” (both moderately indicative of being gay).

This is further illustrated in Fig. S1, which shows the average levels of personality traits and age for several popular Likes. Each Like attracts users with a different average personality and demographic profile and, thus, can be used to predict those attributes. For example, users who liked the “Hello Kitty” brand tended to be high on Openness and low on “Conscientiousness,” “Agreeableness,” and “Emotional Stability.” They were also more likely to have Democratic political views and to be of African-American origin, predominantly Christian, and slightly below average age. The same Likes were used to create Fig. S2, presenting their relative popularity in four groups: Democrats, Christians, Homosexuals, and African-American individuals. For example, although liking “Barack Obama” is clearly related to being a Democrat, it is also relatively popular among Christians, African Americans, and Homosexual individuals.