“You” speaks to me: Effects of generic-you in creating resonance between people and ideas

Novel selection.

Our sample consisted of all books that were selected for Oprah’s Book Club between the years 1996 and 2018, so long as they were: adult fiction, written after 1900, originally written in English, and had “Popular Highlights” available. Books also needed to be available through the Amazon Kindle application. We reasoned that these books would provide a sample of fiction that had been widely read by the general public. This yielded a sample of 56 books written by 46 authors (SI Appendix, Table S3).*

Selection of control passages.

To test our hypothesis that generic person indicators (i.e., generic “you,” “we,” “one,” and “people”) would be more likely to appear in highlighted passages compared to other parts of the book that readers had not highlighted, we identified a set of control passages, which readers had not highlighted while reading. In selecting the control passages, we took several considerations into account: First, we wanted to ensure independence between highlighted and control passages (that is, writers may have been building toward a generalization or insight, and thus the sentence immediately preceding a highlighted passage may have been less likely to contain a generic person reference). Second, given that some readers may have stopped reading a book at any point after starting, we wanted to ensure to the extent possible that people had read the selected control passages (which would mean that they would have had equal opportunity to highlight them). To meet these considerations, we selected yoked control passages (method for selection described below). Control passages were matched to highlighted passages in number of sentences; word count was included as a fixed effect in all analyses, as sentence length varied.

Data collection.

Data were collected from the Amazon Kindle desktop application using downloaded e-book editions of the selected novels. The method for selecting highlighted and control passages was devised a priori and was preregistered (24). Trained research assistants first synchronized their viewing settings through the Amazon Kindle desktop application to portrait mode; font size was left as the default setting upon the e-book download. Research assistants then enabled the Popular Highlights feature, allowing them to view highlighted passages.

Each highlighted passage was copied and pasted into a spreadsheet for subsequent use. In addition to the highlighted passage, all other text on the screen was copied and pasted so that human coders had additional contextual information to use when determining whether a given indicator was generic. If the highlighted passages spanned two pages, approximately half the text on each page preceding and following the passage was selected for context.

To select yoked control passages, research assistants navigated from the Popular Highlights window to each highlighted passage in the text. They then navigated two screens back in the text and found the start of the third complete sentence on the screen to mark the beginning of the control passage. If the third available sentence was also highlighted, research assistants began with the start of the next sentence. They then selected control text that was matched in the number of sentences to the yoked highlighted passage. The third sentence on the screen was chosen as a starting point to allow for sufficient surrounding context, and they navigated two screens back to ensure that the control content had been read, and thus had an equal opportunity to be highlighted.

Several exceptions were built into the instructions to account for unusual scenarios. First, if a highlighted passage included an incomplete sentence, research assistants were instructed to treat it as a full sentence when selecting the yoked control passage. For example, the following highlighted passage led to the selection of a two-sentence-long control passage: “... on the bright and sunny day. They hadn’t enjoyed themselves like this for many years.” Second, if there were two highlighted passages on a given page, the control passages had the same number of sentences separating them as the two highlighted passages did. Third, if the highlighted passage was located within the first two screens of the book, the control passage was selected by navigating two screens later; this only occurred with four sets of highlighted and control passages, all from the same book.

Researchers collected additional context for each control passage following the same method described for the highlighted passages.

In addition to collecting the passages and their surrounding context, the following information was recorded for each passage: 1) Date the e-book was downloaded to the Kindle library and date that highlights were recorded (given that the Popular Highlights may change over time depending on Amazon’s method for selection); 2) Passage location provided by Amazon (for control and highlighted passages); 3) Number of highlights (for highlighted passages only); 4) Number of sentences; 5) Number of words [determined by Linguistic Inquiry and Word Count (2015; LIWC) for control and highlighted passages; 34].

The full set of passages included in this study is available via the Open Science Framework (https://osf.io/6J2ZC/).

Text analysis and coding.

All passages were run through LIWC 2015. Given the high volume of passages, we created four subsets of passages for human coding: 1) Passages identified as containing the word “you” (including all variants such as “your,” “you’ll,” etc.); 2) Passages identified as containing the word “we” (including all variants such as “our,” “us,” etc.); 3) Passages identified as containing the word “one” (including variants such as “one’s”); 4) Passages identified as containing the word “people” (including “person” and “persons”; identified through a custom dictionary). Coders used these subsets of data when determining which uses of each indicator were generic.^†

Two condition-blind coders coded all passages that contained the word “you” to determine whether each use of “you” in a given passage was generic or not, following prior methods (18). Given the restricted range of generic-yous in each passage (M = 0.37, range: 0–11, mode: 1), we created a categorical variable indicating presence vs. absence of generic-you in a passage. Reliability between coders was excellent, K = 0.89. Discrepancies were resolved by an independent coder.

A different set of two condition-blind coders identified all generic instances of “we,” “people,” and “one.” These coders trained on 20% of the data, given that separate practice datasets with generic “we,” “people,” and “one” were not available. Similar to the frequency of generic-you, the ranges of generic “we” (M = 0.11, range: 0–7, mode: 1), “one” (M = 0.01, range: 0–2, mode: 1), and “people” (M = 0.05, range: 0–4, mode: 1) were restricted, so we created categorical variables indicating presence vs. absence of each of these indicators. Reliability for the 80% of the data that the coders independently coded was good (overall K = 0.77; K_We = 0.75, K_One, = 0.94, K_people = 0.61). Disagreements were resolved through discussion.

Analytic approach.

Our general approach for all studies was to use multilevel models, which consider hierarchically structured data (e.g., passages nested within books) and random effects. All analyses were conducted using R’s lme4 package (35). We began by attempting to run the model proposed in our preregistration document. If models did not converge, we removed random effect terms that contributed little variance to simplify the models and improve likelihood of convergence (for discussion, see 36).

Primary analyses.

We preregistered two primary analyses for study 1. First, we planned to examine the effect of condition (highlighted passages vs. nonhighlighted control passages) on the presence (vs. absence) of generic-you. Second, we planned to examine the effect of condition (highlighted passages vs. nonhighlighted control passages) on the presence (vs. absence) of at least one of the indices of generality that we coded for (i.e., generic-you, generic-we, generic-people, and generic-one).

To examine these questions, we conducted multilevel logistic regression models. In both models, condition (highlighted passages vs. nonhighlighted control passages) was entered as a fixed effect. The number of words in each passage (i.e., word count, mean centered) was also entered as a fixed effect to control for the role of passage length on the presence or absence of the linguistic indicators of interest (although yoked passages were matched on sentence length, word count could still vary between the yoked control and highlighted passages). The models also included the interaction between condition and word count. Both models also included “book” as a random effect with random intercepts. Although our preregistration additionally specified including “passage” as a random effect, model fit was singular, so passage was dropped as a random effect (additionally, inspection of a model that only included passage as a random effect revealed that it contributed little variance—i.e., 0.08).

We report the fixed effect of condition from the models in the main text; below in Table 1, we report all fixed and random effects.

Participants.

Participants were 382 individuals recruited from TurkPrime (37). Seven individuals were screened out for not being native English speakers. Nine individuals dropped out nearly immediately or did not provide consent, and an additional three participants were excluded for completing fewer than 90% of the trials. This left a sample of 363 participants (M_age = 38.36, SD = 11.28; 155 female, 70% White, 15% Black, 6% Asian, 5% Hispanic, 4% identifying with other races/ethnicities, or preferring not to respond).^‡

Stimuli.

Stimuli consisted of all passages from study 1 that contained generic-you and were one sentence in length (n = 55) as well as their yoked control passages (n = 55). See SI Appendix, Table S4 for all passages. We restricted our stimuli to passages that contained generic-you because it was by far the most commonly observed generic person reference in study 1. We restricted the stimuli to passages that were one sentence in length to reduce the burden on participants who would be reading each passage.

Design.

We used a within-subjects, repeated-measures design. Each participant read 11 highlighted passages and 11 nonhighlighted control passages. As a between-subjects factor, participants were randomly assigned to receive one of five batches of passages. Batches were constructed by randomly dividing the 110 passages into five groups, each containing an equal number of control and highlighted passages. Highlighted and yoked control passages could be spread across different batches, and all passages were presented in random order.

Procedure.

After providing consent and indicating that they were native English speakers, participants began the main task. They were presented with the following text to introduce the purpose of the study and the concept of resonance: “We’re interested in gathering a database of meaningful quotes. When reading each passage, rate to what extent it resonates or has an impact on you.” Participants were then presented with 22 highlighted and nonhighlighted control passages, in random order. Only one passage was presented per screen. For each passage, participants responded to the question: “How much does this passage resonate with you?” (1, Not at all; 2, A little; 3, A moderate amount; 4, A lot; 5, A great deal; M = 2.50, SD = 1.37). After completing the main task, participants answered demographics questions.

Primary analysis.

Multilevel analyses were conducted using R’s lme4 package to examine the effect of condition (highlighted vs. nonhighlighted control passages) on resonance ratings. Our preregistration specified a complex model that accounted for all potential random effects; however, initial attempts to fit this model revealed that fit was singular. Removing batch and retaining book allowed for model convergence, but inspection of the random effects revealed that book was contributing very little variance (0.04); hence, we removed both batch and book as random effects to allow for the most parsimonious model. Notably, however, all model variations revealed a significant main effect of condition.

The final model included condition (highlighted vs. nonhighlighted control), the number of words in each passage (i.e., word count, to control for the role of passage length on resonance ratings), and the interaction between condition and word count. We included passage as a random effect with random intercepts; participant was also included as a random effect, and initial tests confirmed that allowing for random slopes at the participant level for the effect of condition improved model fit (χ² = 329.69, P < 0.001). The fixed effect of condition from this model is reported in the main text; all additional fixed and random effects are reported below, in Table 2. As indicated below, the interaction between condition and word count was significant, such that resonance ratings decreased slightly for highlighted passages as the number of words increased.

Participants.

Participants were 318 individuals recruited from TurkPrime. Three individuals were screened out for not being native English speakers. Eight individuals dropped out nearly immediately or were excluded on the basis of having duplicate internet protocol (IP) addresses and suspicious responses. An additional seven participants were excluded who completed fewer than 90% of the trials. This left a sample of 300 participants (M_age = 35.61, SD = 10.72; 118 female, 70% White, 11% Black, 5% Asian, 5% Hispanic, 9% identifying with other races/ethnicities, or preferring not to respond).

Stimuli.

Stimuli consisted of the one-sentence highlighted, generic-you passages used in study 2, except that we also created an alternative version of each passage that contained first-person singular pronouns instead of generic-you pronouns.^§ Fifteen (i.e., 28%) of the passages were altered to ensure that the first-person singular passages were clear and to make the manipulation maximally clean. See SI Appendix, Table S5 for all passages. A detailed description of situations that precipitated altering the passages is contained in the preregistration document (26).

Design.

The 54 passages were divided into three batches of 18 that were roughly matched on the resonance ratings collected in study 2. That is, we created three tertiles from the 54 passages that ordered them in terms of how resonant they were judged to be by participants in study 2 (lowest tertile: M = 2.43, SD = 0.19; middle tertile: M = 2.84, SD = 0.11; highest tertile: M = 3.31, SD = 0.27). We then randomly assigned passages within each tertile to one of three batches; this ensured that passages across the three batches were matched in terms of how resonant they were. Finally, within each batch, we created an “A” version and a “B” version, which counterbalanced whether a given passage was presented with generic-you vs. first-person singular pronouns (hereafter, sometimes referred to as “I" versions, for short). This ensured that generic-you and I versions of the passages were presented equally across participants. Altogether, then, there were six sets of passages.

As a between-subjects factor, we randomly assigned participants to receive one of the six batches of passages. Each participant rated 18 passages total (half containing generic-you and half containing first-person singular pronouns) on resonance using the same measure described in study 2. Passages were presented in random order.

Preliminary analyses.

Multilevel analyses, which allowed us to consider the hierarchical structure of the data, were conducted using R’s lme4 package to examine the effect of condition (generic-you passages vs. “I” passages) on resonance ratings. Our preregistration proposed a preliminary examination of whether the counterbalanced version of passages (i.e., whether a given passage was presented with first-person singular pronouns vs. generic-you pronouns) affected resonance ratings. However, we subsequently realized that this counterbalancing was nested within each of the three batches of passages. Thus, we examined the counterbalanced factor of passage version as a random effect instead, nested within the three distinct batches of passages. The fullest model with all random effects included yielded singular fit. Subsequent inspection of the random effects revealed that book was explaining very little variance, as were batch passage version. Thus, these random effect terms were removed. Given that our design ensured that each batch of passages had an equal distribution of passages that varied in resonance ratings and that our model still included the random effect of passage (to account for stimulus-level variability), we deemed this appropriate both theoretically and analytically.

Next, we examined whether including random slopes for the effect of condition at the participant level improved model fit; it did not (χ² = 1.03, P = 0.597), so only intercepts were allowed to vary. Thus, our final model included condition, word count, and their interaction as fixed effects. Passage number and participant were entered as random effects with random intercepts.

Once this model was determined, we proceeded to conduct one preliminary analysis (as outlined in our preregistration) to examine whether altering a given passage was associated with increased or decreased resonance ratings. Condition (generic-you passage vs. “I” passage) and whether a passage was altered were entered as fixed effects; the interaction between these two terms was also included in the model. As outlined in the paragraph above, passage number and participant were entered as random effects with random intercepts. When including whether the passage was altered in the model, the effect of condition was in the expected direction, b = 0.06, P = 0.055, but whether a passage was altered did not affect resonance ratings, b = −0.17, P = 0.164. Importantly, condition did not interact with whether a passage was altered to affect resonance ratings, b = −0.11, P = 0.096; thus, this factor was not considered in subsequent models.

Primary analysis.

The fixed effect of condition on resonance ratings from this model is reported in the main text; Table 3 reports all fixed and random effects from the model.

Participants.

Participants were 206 individuals recruited from TurkPrime. One individual was screened out for not being a native English speaker. Five individuals dropped out nearly immediately, and an additional participant was excluded for responding to fewer than 90% of the trials. This left a sample of 199 participants (M_age = 35.77, SD = 11.30; 90 female, 76% White, 9% Black, 6% Asian, 5% Hispanic, 4% identifying with other races/ethnicities, or preferring not to respond).

Stimuli.

Stimuli consisted of 20 one-sentence statements that we created. Each statement contained a generic-you version and a first-person singular pronoun version. See SI Appendix, Table S6 for all passages.

Design.

As a between-subjects factor, we counterbalanced which statements were presented with generic-you vs. “I” (by creating a set A and a set B version of the stimuli); thus, as in study 3, each passage was equally presented with both linguistic frames across participants. Each participant saw 10 passages expressed with generic-you and 10 expressed with “I,” in random order. After each passage was presented, participants were asked to rate the extent to which it resonated with them, with the same question used in studies 2 and 3, except rather than asking participants to rate each passage we asked them to rate each “statement.”

Preliminary analyses.

Multilevel analyses, which allowed us to consider the hierarchical structure of the data, were conducted using R’s lme4 package to examine the effect of condition (generic-you passage vs. first-person singular pronoun passage) on resonance ratings. We conducted one preliminary analysis, as outlined in our preregistration. Specifically, we tested whether set (i.e., which particular passages were assigned to generic-you vs. “I”) affected resonance ratings. It did not (b = −0.08, P = 0.345), so this term was not considered further.

Including random slopes for the effect of condition at the participant level did not improve model fit (χ² = 4.27, P = 0.118), so only intercepts were allowed to vary. The final model thus included the fixed effect of condition and random intercepts for passage and participant.

Primary analysis.

The fixed effect of condition on resonance ratings from this model is reported in the main text; Table 3 reports all fixed and random effects from the model.

Participants.

Participants were 1,074 individuals recruited from TurkPrime. Sixteen individuals were screened out for not being native English speakers. Thirteen individuals dropped out nearly immediately, an additional six participants were excluded for responding to fewer than 90% of the trials, and one was excluded because of a duplicate IP address. This left a sample of 1,038 participants (M_age = 38.99, SD = 12.67; 519 female, 74% White, 9% Black, 7% Asian, 5% Hispanic, 5% identifying with other races/ethnicities, or preferring not to respond).

Stimuli.

Stimuli consisted of the 20 one-sentence statements used in study 4, with slight alterations to make them parallel across conditions. Each statement contained a generic-you version, a generic-people version, and a first-person singular pronoun version ("I" version). See SI Appendix, Table S6 for all passages.

Design.

The design was identical to study 4 with one exception: Participants were randomly assigned to receive statements with one of three sets of linguistic contrasts: generic-you vs. generic-people (n = 345), generic-you vs. “I” (n = 346), or generic-people vs. “I” (n = 347).

Preliminary analyses.

Multilevel analyses, which allowed us to consider the hierarchical structure of the data, were conducted using R’s lme4 package to examine the effect of condition (i.e., each linguistic contrast) on resonance ratings. In all models, passage and participant were entered as random effects. For the models testing the contrast between generic-you vs. “I” and generic-you vs. generic-people, including random slopes for condition at the participant level significantly improved model fit (generic-you vs. generic-people: χ² = 6.36, P = 0.041; generic-you vs. I: χ² = 22.22, P < 0.001); for the model testing the contrast between generic-people and “I,” including random slopes for condition at the participant level did not significantly improve model fit (χ² = 2.66, P = 0.265), so only random intercepts for participant were included.

Next, we conducted one preliminary analysis for each linguistic contrast, as outlined in our preregistration. Specifically, we tested whether set (i.e., which passages were assigned to be presented with certain wording, e.g., generic-you vs. “I”) affected resonance ratings. It did not for any of the linguistic contrasts (generic-you vs. generic-people effect of set: b = −0.07, P = 0.339; generic-you vs. “I” effect of set: b = 0.02, P = 0.774; generic-people vs. “I” effect of set: b = 0.01, P = .847), so this term was not considered further in any of the models.

Primary analyses.

The fixed effect of condition on resonance ratings from these models is reported in the main text; Table 3 reports all fixed and random effects from the model.