Exploratory Data Analysis
Exploratory Data Analysis
Introduction
This report initiates an exploratory investigation into the role of speech characteristics in evaluating interaction quality. The experiment consists of four parts featuring semi-structured dialogues conducted via Zoom, with questions that progressively increase in intimacy. After the conversation, participants completed the Goodness of Interaction (GOI) Questionnaire, which assesses their interaction experience. Our analysis focuses on individual speech patterns and dyadic synchrony, examining metrics such as word count similarity and speaking time ratio within dyads. To ensure robustness, we arbitrarily divided our 120 dyads into two datasets of 60 dyads each, conducting the initial analysis on the first dataset and planning to replicate our findings using the second dataset.
Pre Process
Load Libraries
library(tidyverse)
library(reshape2)
library(ggcorrplot)
library(lmerTest)
library(performance)
library(parameters)
library(corrplot)Load Data
data <- read.csv("Data_with_Gender.csv")
head(data)## iDyad iSubject iPartner TW TW_partner ST ST_partner WPM
## 1 101_102 101 102 2039 1846 846.088 891.503 147.2107
## 2 101_102 102 101 1846 2039 891.503 846.088 123.9325
## 3 103_104 103 104 1713 2551 766.419 956.859 135.2987
## 4 103_104 104 103 2551 1713 956.859 766.419 159.7310
## 5 105_106 105 106 1060 1042 451.392 423.877 140.0591
## 6 105_106 106 105 1042 1060 423.877 451.392 148.3425
## WPM_partner GOI Gender
## 1 123.9325 81.8 1
## 2 147.2107 55.8 1
## 3 159.7310 76.5 1
## 4 135.2987 67.1 1
## 5 148.3425 84.5 1
## 6 140.0591 58.2 1Speech Attributes Predictors
Variables
Predictive:
GOI - Goodness of Interaction questionnaire score for self-assessment of interaction quality.
Predictors:
TW - The total number of words spoken by each participant during the conversation.
ST - The total amount of time, measured in seconds, during which a participant actively spoke, excluding any periods of silence, pauses, or non-verbal communication. These exclusions were determined using Audacity software for precise audio processing and silence removal.
WPM - Words per minute, calculated as TW / (ST / 60).
Partner - _partner suffix indicating
that the variable is measured for the second participant within the
dyad.
Descriptive Statistics
GOI
GOI_mean <- mean(data$GOI)
GOI_med <- median(data$GOI)
GOI_SD <- sd(data$GOI)
cat(c("Mean:", GOI_mean, "\nMedian:", GOI_med, "\nSD:", GOI_SD))## Mean: 68.4285449735
## Median: 72.05
## SD: 19.2574304302369ggplot(data, aes(x = GOI)) +
geom_density(fill = "#FBB4AE") +
geom_vline(xintercept = GOI_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Total Words
TW_mean <- mean(data$TW)
TW_med <- median(data$TW)
TW_SD <- sd(data$TW)
cat(c("Mean:", TW_mean, "\nMedian:", TW_med, "\nSD:", TW_SD))## Mean: 1460.30833333333
## Median: 1299.5
## SD: 695.25686923301ggplot(data, aes(x = TW)) +
geom_density(fill = "#B3CDE3") +
geom_vline(xintercept = TW_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Speaking Time
The total amount of time, measured in seconds, during which a participant actively spoke, excluding any periods of silence, pauses, or non-verbal communication.
ST_mean <- mean(data$ST)
ST_med <- median(data$ST)
ST_SD <- sd(data$ST)
cat(c("Mean:", ST_mean, "\nMedian:", ST_med, "\nSD:", ST_SD))## Mean: 641.629175
## Median: 574.6865
## SD: 272.839385756241ggplot(data, aes(x = ST)) +
geom_density(fill = "#CCEBC5") +
geom_vline(xintercept = ST_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Speech Rate
- Words Per Minute
WPM_mean <- mean(data$WPM)
WPM_med <- median(data$WPM)
WPM_SD <- sd(data$WPM)
cat(c("Mean:", WPM_mean, "\nMedian:", WPM_med, "\nSD:", WPM_SD))## Mean: 135.145140308667
## Median: 134.5623547
## SD: 20.2735024769839ggplot(data, aes(x = WPM, color = factor(Gender))) +
geom_density(fill = "#DECBE4", alpha = 0.4) +
geom_vline(xintercept = WPM_mean, linetype = "dashed", color = "gray40") +
theme_classic()
GOI VS Speech Attributes: Correlation Matrix
# Remove rows with NA values in selected columns
#clean_data <- data |> select(GOI, TW, TW_partner, ST, ST_partner, WPM, WPM_partner) |> na.omit()
# Calculate the correlation matrix and p-values
#corr_results <- corr(as.matrix(clean_data))
#corr_matrix <- corr_results$r
#p_matrix <- corr_results$P
# Plot the heatmap with correlation coefficients and significance markers
#corrplot(corr_matrix, method = "color", col = colorRampPalette(brewer.pal(n = 8, name = "RdYlBu"))(200),
# type = "upper", p.mat = p_matrix, sig.level = c(0.001, 0.01, 0.05),
# insig = "label_sig", pch.cex = 1.2, number.cex = 0.9,
# tl.col = "black", tl.srt = 45, addCoef.col = "black")Report:
The correlation matrix reveals that longer speaking duration, measured by spoken words (TW, r = 0.266, p = 0.003) and speaking time (ST, r = 0.270, p = 0.003), is positively associated with higher interaction quality evaluations (GOI), as anticipated. Notably, the speaking duration of the partner (TW_partner, r = 0.315, p < 0.001; ST_partner, r = 0.347, p < 0.001) shows an even stronger correlation with GOI. This suggests that the more one person spoke during the dialogue, the higher their partner perceived the quality of the interaction. Although a higher speech rate (WPM) was also positively correlated with GOI (r = 0.097, p = 0.289), these associations were not statistically significant for either the individual or their partner (WPM_partner, r = 0.054, p = 0.561).
Synchrony
In this section, we explore the role of synchrony in dyadic interactions and its impact on perceived interaction quality (GOI). Synchrony, in this context, refers to the alignment or discrepancy in speaking behaviors between partners within a dyad. To quantify this, we calculated both absolute differences and relative ratios between dyad members across three key speech characteristics: total words spoken (TW), speaking time (ST), and words per minute (WPM).
Initially, we derived the absolute differences between each participant’s speech metrics and those of their partner (TW_diff, ST_diff, WPM_diff). These measures provide a straightforward indicator of speech asymmetry, where higher values represent greater divergence between dyad members.
Next, we computed the relative ratios for each speech characteristic (TW_ratio, ST_ratio, WPM_ratio), offering a perspective on the proportional contribution of each dyad member to the interaction. Ratios closer to 0.5 indicate a balanced interaction, while deviations suggest dominance by one partner.
-To further explore these relationships, we applied log transformations to the difference and ratio measures (e.g., TW_diff_log, TW_ratio_log) to address potential skewness in the data. We then examined the influence of these synchrony measures on interaction quality (GOI) using linear mixed-effects models and linear regression.
Similarity
Absolute Distance Between Dyad’s Participants
data_diff <- data |>
mutate(
TW_diff = as.numeric(abs(TW - TW_partner)),
ST_diff = abs(ST - ST_partner),
WPM_diff = abs(WPM - WPM_partner))
data_diff |>
select(iSubject, TW_diff, ST_diff, WPM_diff) |>
head()## iSubject TW_diff ST_diff WPM_diff
## 1 101 193 45.415 23.278206
## 2 102 193 45.415 23.278206
## 3 103 838 190.440 24.432296
## 4 104 838 190.440 24.432296
## 5 105 18 27.515 8.283413
## 6 106 18 27.515 8.283413Ratio
data_diff_ratio <- data_diff |>
mutate(
TW_ratio = TW / (TW + TW_partner),
ST_ratio = ST / (ST + ST_partner),
WPM_ratio = WPM / (WPM_partner + WPM),
TW_ratio_partner = TW_partner / (TW + TW_partner),
ST_ratio_partner = ST_partner / (ST + ST_partner),
WPM_ratio_partner = WPM_partner / (WPM_partner + WPM))
data_diff_ratio |>
select(iSubject, TW_ratio, ST_ratio, WPM_ratio) |>
head()## iSubject TW_ratio ST_ratio WPM_ratio
## 1 101 0.5248391 0.4869316 0.5429260
## 2 102 0.4751609 0.5130684 0.4570740
## 3 103 0.4017355 0.4447448 0.4585935
## 4 104 0.5982645 0.5552552 0.5414065
## 5 105 0.5042816 0.5157180 0.4856391
## 6 106 0.4957184 0.4842820 0.5143609Total Words Synchrony
Total Words Similarity
Absolute Distance Between Total Words within Dyad
TW_diff_mean <- mean(data_diff$TW_diff)
TW_diff_med <- median(data_diff$TW_diff)
TW_diff_SD <- sd(data_diff$TW_diff)
cat(c("Mean:", TW_diff_mean, "\nMedian:", TW_diff_med, "\nSD:", TW_diff_SD))## Mean: 441.55
## Median: 288
## SD: 385.356428921232ggplot(data_diff, aes(x = TW_diff)) +
geom_density(fill = "#A1D3E8") +
geom_vline(xintercept = TW_diff_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Total Words Ratio
TW_ratio: Words(SubjectA) / (Words(SubjectA) +
Words(SubjectB))
TW_ratio_SD <- sd(data_diff_ratio$TW_ratio)
cat(c("SD:", TW_ratio_SD))## SD: 0.0863473224183988ggplot(data_diff_ratio, aes(x = TW_ratio)) +
geom_density(fill = "#89CFF0") +
theme_classic()
Speaking Time Synchrony
Speaking Time Similarity
Absolute Distance Between Speaking Time within Dyad
ST_diff_mean <- mean(data_diff$ST_diff)
ST_diff_med <- median(data_diff$ST_diff)
ST_diff_SD <- sd(data_diff$ST_diff)
cat(c("Mean:", ST_diff_mean, "\nMedian:", ST_diff_med, "\nSD:", ST_diff_SD))## Mean: 155.996716666667
## Median: 104.34
## SD: 139.306805596914ggplot(data_diff, aes(x = ST_diff)) +
geom_density(fill = "#BEE3AD") +
geom_vline(xintercept = ST_diff_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Speaking Time Ratio
ST_ratio_SD <- sd(data_diff_ratio$ST_ratio)
cat(c("SD:", ST_ratio_SD))## SD: 0.0740915221037298ggplot(data_diff_ratio, aes(x = ST_ratio)) +
geom_density(fill = "#A8D5BA") +
theme_classic()
Speech Rate Synchorny
Speech Rate Similarity
Absolute Distance Between Speech Rate within Dyad
WPM_diff_mean <- mean(data_diff$WPM_diff)
WPM_diff_med <- median(data_diff$WPM_diff)
WPM_diff_SD <- sd(data_diff$WPM_diff)
cat(c("Mean:", WPM_diff_mean, "\nMedian:", WPM_diff_med, "\nSD:", WPM_diff_SD))## Mean: 18.5636318826667
## Median: 16.08311715
## SD: 13.0876795158243ggplot(data_diff, aes(x = WPM_diff)) +
geom_density(fill = "#CBA9DA") +
geom_vline(xintercept = WPM_diff_mean, linetype = "dashed", color = "gray40") +
theme_classic()
Speech Rate Ratio
WPM_ratio_SD <- sd(data_diff_ratio$WPM_ratio)
cat(c("SD:", WPM_ratio_SD))## SD: 0.042719434410943ggplot(data_diff_ratio, aes(x = WPM_ratio)) +
geom_density(fill = "#BBA4C3") +
theme_classic()
# Apply log transformation for diff and ratio
data <- data_diff_ratio |>
mutate(TW_diff_log = log2(TW_diff),
TW_ratio_log = log2(TW_ratio),
TW_ratio_dyad_log = log2(TW_ratio) + log2(TW_ratio_partner),
ST_diff_log = log2(ST_diff),
ST_ratio_log = log2(ST_ratio),
ST_ratio_dyad_log = log2(ST_ratio) + log2(ST_ratio_partner),
WPM_diff_log = log2(WPM_diff),
WPM_ratio_log = log2(WPM_ratio),
WPM_ratio_dyad_log = log2(WPM_ratio) + log2(WPM_ratio_partner))
data |>
select(iSubject, iDyad, TW, TW_partner, TW_diff, TW_diff_log, TW_ratio, TW_ratio_log, TW_ratio_dyad_log) |>
head()## iSubject iDyad TW TW_partner TW_diff TW_diff_log TW_ratio TW_ratio_log
## 1 101 101_102 2039 1846 193 7.592457 0.5248391 -0.9300528
## 2 102 101_102 1846 2039 193 7.592457 0.4751609 -1.0735120
## 3 103 103_104 1713 2551 838 9.710806 0.4017355 -1.3156823
## 4 104 103_104 2551 1713 838 9.710806 0.5982645 -0.7411445
## 5 105 105_106 1060 1042 18 4.169925 0.5042816 -0.9876984
## 6 106 105_106 1042 1060 18 4.169925 0.4957184 -1.0124074
## TW_ratio_dyad_log
## 1 -2.003565
## 2 -2.003565
## 3 -2.056827
## 4 -2.056827
## 5 -2.000106
## 6 -2.000106GOI VS Speech Synchrony: Correlation Matrix
# Calculate the correlation matrix and p-values
#corr_results <- corr(as.matrix(data |>
# select(GOI, TW_diff, TW_ratio, ST_diff, ST_ratio, WPM_diff, WPM_ratio)))
#corr_matrix <- corr_results$r
#p_matrix <- corr_results$P
# Plot the heatmap with correlation coefficients and p-values using a built-in palette
#corrplot(corr_matrix, method = "color", col = colorRampPalette(brewer.pal(n = 8, name = "RdYlBu"))(200),
# type = "upper", p.mat = p_matrix, sig.level = c(0.001, 0.01, 0.05),
# insig = "label_sig", addCoef.col = "black", number.cex = 0.9,
# tl.col = "black", tl.srt = 45, pch.cex = 0.7)Model Selection
In this section, we investigate how speech characteristics contribute to the variability in interaction quality, as measured by GOI scores, across dyads. We begin by constructing baseline models with random intercepts to account for the inherent variation in GOI among different dyads. This approach allows us to first quantify the extent to which differences between dyads influence interaction quality before systematically introducing predictors related to total speech output (TW, ST, WPM) and the differences in speech attributes between dyad members (TW_diff, ST_diff, WPM_diff).
TW Model
model0 <- lmer(GOI ~ (1|iDyad), data = data)
model1 <- lmer(GOI ~ TW + (1|iDyad), data = data)
model2 <- lmer(GOI ~ TW_diff + (1|iDyad), data = data)
model3 <- lmer(GOI ~ TW * TW_diff + (1 | iDyad), data = data)## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescalingsjPlot::tab_model(model0, model1, model2, model3, show.std = TRUE, show.ci = FALSE)| GOI | GOI | GOI | GOI | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Predictors | Estimates | std. Beta | p | Estimates | std. Beta | p | Estimates | std. Beta | p | Estimates | std. Beta | p | std. p |
| (Intercept) | 68.43 | 0.00 | <0.001 | 59.97 | 0.00 | <0.001 | 61.60 | 0.00 | <0.001 | 46.07 | 0.13 | <0.001 | 0.234 |
| TW | 0.01 | 0.21 | 0.029 | 0.01 | 0.20 | 0.014 | 0.094 | ||||||
| TW diff | 0.02 | 0.31 | 0.003 | 0.04 | 0.29 | 0.001 | 0.013 | ||||||
| TW × TW diff | -0.00 | -0.22 | 0.012 | 0.012 | |||||||||
| Random Effects | |||||||||||||
| σ2 | 224.66 | 239.03 | 224.66 | 230.01 | |||||||||
| τ00 | 147.43 iDyad | 111.00 iDyad | 115.51 iDyad | 90.01 iDyad | |||||||||
| ICC | 0.40 | 0.32 | 0.34 | 0.28 | |||||||||
| N | 60 iDyad | 60 iDyad | 60 iDyad | 60 iDyad | |||||||||
| Observations | 120 | 120 | 120 | 120 | |||||||||
| Marginal R2 / Conditional R2 | 0.000 / 0.396 | 0.044 / 0.347 | 0.094 / 0.402 | 0.151 / 0.390 | |||||||||
ranova(model3)## ANOVA-like table for random-effects: Single term deletions
##
## Model:
## GOI ~ TW + TW_diff + (1 | iDyad) + TW:TW_diff
## npar logLik AIC LRT Df Pr(>Chisq)
## <none> 6 -530.73 1073.5
## (1 | iDyad) 5 -533.12 1076.2 4.7731 1 0.02891 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1TW Report
The mixed-effects analysis shows that the interaction between total words spoken (TW) and the difference in word count between dyad members (TW_diff) significantly impacts interaction quality (GOI). Specifically, the interaction term is negative and significant (Estimate = -0.00, p = 0.012), indicating that while TW_diff generally enhances GOI (Estimate = 0.04, p = 0.001), this effect diminishes as TW increases. The model’s explanatory power is reflected in a marginal R² of 0.151 and a conditional R² of 0.390, showing that the fixed effects explain 15.1% of the variance in GOI, with dyad-specific factors accounting for the rest. The significance of the random intercept for dyads was confirmed by a likelihood ratio test (p = 0.02891).
TW Interaction Effects Visualization
# Function to create interaction plots based on quartile grouping of a moderator variable
# Arguments: data, variable, moderator
create_interaction_plot <- function(data, variable_x, variable_m) {
# Calculate IQR thresholds for grouping
iqr_values <- quantile(data[[variable_m]], probs = c(0.25, 0.5, 0.75), na.rm = TRUE)
iqr_breaks <- c(-Inf, iqr_values[1], iqr_values[2], iqr_values[3], Inf)
# Create groups based on the IQR
data <- data %>%
mutate(Quartile_Group = cut(data[[variable_m]],
breaks = iqr_breaks,
labels = c("Q1", "Q2", "Q3", "Q4")))
# Generate the interaction plot
p <- ggplot(data, aes(x = .data[[variable_x]], y = GOI, color = Quartile_Group)) +
geom_point(alpha = 0.6, size = 3) +
geom_smooth(method = "lm", aes(group = Quartile_Group), se = FALSE, size = 1) +
scale_color_viridis_d(option = "magma") +
labs(title = paste("Moderating Effect of Grouped", variable_m, "on", variable_x, "and GOI"),
subtitle = "Regression lines for each level of grouped Quartile",
x = paste(variable_x, "[s]"),
y = "Interaction Quality (GOI)",
color = paste("Grouped", variable_m, "Levels")) +
theme_minimal()
# Return the plot
return(p)
}TW_model <- create_interaction_plot(data, "TW_diff", "TW")## Warning: Using `size` aesthetic for lines was deprecated in ggplot2 3.4.0.
## ℹ Please use `linewidth` instead.
## This warning is displayed once every 8 hours.
## Call `lifecycle::last_lifecycle_warnings()` to see where this warning was
## generated.print(TW_model)## `geom_smooth()` using formula = 'y ~ x'
TW Plot Report
The interaction plot visualizes the moderating effect of total words spoken (TW) on the relationship between word count difference (TW_diff) within dyads and the quality of interaction (GOI). As indicated by the regression lines, a nuanced moderation effect emerges across the quartiles of TW. Notably, in the highest quartile (Q4), the interaction shows a negative trend (β = -0.22, p = 0.012), suggesting that dyads with smaller differences in spoken words tend to evaluate their interactions more favorably. This pattern contrasts with the generally positive, though decreasing, trend observed in the lower quartiles (Q1, Q2, and Q3), where higher TW_diff is associated with enhanced interaction quality. These findings highlight the complex role of speech balance within dyads, particularly under varying conditions of total speech output.
ST Model
model0 <- lmer(GOI ~ (1|iDyad), data = data)
model1 <- lmer(GOI ~ ST + (1|iDyad), data = data)
model2 <- lmer(GOI ~ ST_diff + (1|iDyad), data = data)
model3 <- lmer(GOI ~ ST * ST_diff + (1 | iDyad), data = data)## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescalingsjPlot::tab_model(model0, model1, model2, model3, df.method = "satterthwaite", show.std = TRUE, show.ci = FALSE)## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescaling| GOI | GOI | GOI | GOI | ||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Predictors | Estimates | std. Beta | p | Estimates | std. Beta | p | Estimates | std. Beta | p | Estimates | std. Beta | p | std. p |
| (Intercept) | 68.43 | 0.00 | <0.001 | 58.95 | 0.00 | <0.001 | 61.67 | 0.00 | <0.001 | 49.30 | 0.07 | <0.001 | 0.538 |
| ST | 0.01 | 0.21 | 0.031 | 0.02 | 0.15 | 0.062 | 0.179 | ||||||
| ST diff | 0.04 | 0.31 | 0.003 | 0.09 | 0.30 | 0.010 | 0.009 | ||||||
| ST × ST diff | -0.00 | -0.14 | 0.091 | 0.091 | |||||||||
| Random Effects | |||||||||||||
| σ2 | 224.66 | 242.13 | 224.66 | 224.71 | |||||||||
| τ00 | 147.43 iDyad | 107.28 iDyad | 114.57 iDyad | 107.85 iDyad | |||||||||
| ICC | 0.40 | 0.31 | 0.34 | 0.32 | |||||||||
| N | 60 iDyad | 60 iDyad | 60 iDyad | 60 iDyad | |||||||||
| Observations | 120 | 120 | 120 | 120 | |||||||||
| Marginal R2 / Conditional R2 | 0.000 / 0.396 | 0.044 / 0.338 | 0.097 / 0.402 | 0.128 / 0.411 | |||||||||
ranova(model3)## ANOVA-like table for random-effects: Single term deletions
##
## Model:
## GOI ~ ST + ST_diff + (1 | iDyad) + ST:ST_diff
## npar logLik AIC LRT Df Pr(>Chisq)
## <none> 6 -528.28 1068.6
## (1 | iDyad) 5 -531.46 1072.9 6.3528 1 0.01172 *
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1ST Report
The analysis reveals that while longer speaking times (ST) have a marginal positive association with GOI (Estimate = 0.01, p = 0.062), the difference in speaking time between dyad members (ST_diff) has a more pronounced positive effect (Estimate = 0.09, p = 0.010). The interaction between ST and ST_diff is not statistically significant (Estimate = -0.00, p = 0.091), suggesting only a potential moderation effect. The model’s marginal R² is 0.128, and the conditional R² is 0.411, indicating that the fixed effects account for 12.8% of the variance in GOI. The random effects highlight significant variability between dyads, with the significance of the random intercept supported by a likelihood ratio test (p = 0.01172).
ST Interaction Effect Visualization
ST_model <- create_interaction_plot(data, "ST_diff", "ST")
print(ST_model)## `geom_smooth()` using formula = 'y ~ x'
ST Plot Report
The interaction plot illustrates the moderating effect of total speaking time (ST) on the relationship between speaking time difference (ST_diff) within dyads and the quality of interaction (GOI). The regression lines reveal a nuanced moderation effect across the quartiles of ST. In the highest quartile (Q4), the interaction trend is relatively flat, with the interaction term between ST and ST_diff showing a modest negative effect (β = -0.14, p = 0.091). This suggests that larger differences in speaking time do not significantly alter the perceived interaction quality. Conversely, in the lower quartiles (Q1, Q2, and Q3), there is a generally positive trend where increased ST_diff is associated with higher GOI. These results underscore the complexity of speaking time balance within dyads and its variable impact on interaction quality depending on the overall speaking time.
WPM Model
model0 <- lmer(GOI ~ (1|iDyad), data = data)
model1 <- lmer(GOI ~ WPM + (1|iDyad), data = data)
model2 <- lmer(GOI ~ WPM_diff + (1|iDyad), data = data)
model3 <- lmer(GOI ~ WPM * WPM_diff + (1 | iDyad), data = data)## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescalingsjPlot::tab_model(model0, model1, model2, model3, df.method = "satterthwaite")## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescaling| GOI | GOI | GOI | GOI | |||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Predictors | Estimates | CI | p | Estimates | CI | p | Estimates | CI | p | Estimates | CI | p |
| (Intercept) | 68.43 | 64.27 – 72.59 | <0.001 | 56.96 | 33.40 – 80.52 | <0.001 | 69.68 | 62.38 – 76.98 | <0.001 | 42.09 | -6.65 – 90.82 | 0.090 |
| WPM | 0.08 | -0.09 – 0.26 | 0.329 | 0.21 | -0.15 – 0.57 | 0.259 | ||||||
| WPM diff | -0.07 | -0.39 – 0.25 | 0.677 | 0.57 | -1.13 – 2.27 | 0.510 | ||||||
| WPM × WPM diff | -0.00 | -0.02 – 0.01 | 0.450 | |||||||||
| Random Effects | ||||||||||||
| σ2 | 224.66 | 225.56 | 224.66 | 222.84 | ||||||||
| τ00 | 147.43 iDyad | 146.10 iDyad | 151.11 iDyad | 156.41 iDyad | ||||||||
| ICC | 0.40 | 0.39 | 0.40 | 0.41 | ||||||||
| N | 60 iDyad | 60 iDyad | 60 iDyad | 60 iDyad | ||||||||
| Observations | 120 | 120 | 120 | 120 | ||||||||
| Marginal R2 / Conditional R2 | 0.000 / 0.396 | 0.008 / 0.398 | 0.002 / 0.403 | 0.018 / 0.423 | ||||||||
ranova(model3)## ANOVA-like table for random-effects: Single term deletions
##
## Model:
## GOI ~ WPM + WPM_diff + (1 | iDyad) + WPM:WPM_diff
## npar logLik AIC LRT Df Pr(>Chisq)
## <none> 6 -523.87 1059.7
## (1 | iDyad) 5 -529.21 1068.4 10.689 1 0.001077 **
## ---
## Signif. codes: 0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1WPM Report
The analysis finds that neither the overall speech rate (WPM) nor the difference in speech rate between dyad members (WPM_diff) significantly impacts interaction quality (GOI). The coefficients for WPM (Estimate = 0.21, p = 0.259) and WPM_diff (Estimate = 0.57, p = 0.510) are not statistically significant, and the interaction term is also non-significant (Estimate = -0.00, p = 0.450). Despite the lack of significant findings, the model’s marginal R² is 0.018, and the conditional R² is 0.423, reflecting that only a small portion of the variance in GOI is explained by the fixed effects, with dyad-specific factors accounting for the majority. The significance of the random intercept was confirmed by a likelihood ratio test (p = 0.001077).
WPM_model <- create_interaction_plot(data, "WPM_diff", "WPM")
print(WPM_model)## `geom_smooth()` using formula = 'y ~ x'
WPM Plot Report
The interaction plot examines the moderating effect of overall speech rate (WPM) on the relationship between the difference in speech rate (WPM_diff) within dyads and the quality of interaction (GOI). The analysis finds no strong evidence for either a main effect of WPM (β = 0.21, p = 0.259) or WPM_diff (β = 0.57, p = 0.510), nor for an interaction effect between the two (β = -0.00, p = 0.450). These findings suggest that variations in speech rate, both in terms of individual totals and differences between dyad members, do not significantly influence interaction quality.
Discussion
This study explored how various speech characteristics within dyadic interactions—specifically, total words spoken (TW), speaking time (ST), and speech rate (WPM)—influence the perceived quality of the interaction, as measured by the Goodness of Interaction (GOI) questionnaire. The analyses focused on both the individual effects of these speech characteristics and their differences between dyad members (TW_diff, ST_diff, WPM_diff), as well as the potential moderating effects of overall speech output.
Key Findings Total Words (TW) and Word Count Difference (TW_diff): The analysis revealed that the relationship between word count difference (TW_diff) and interaction quality (GOI) is significantly moderated by the total words spoken by dyad members (TW). The interaction effect is particularly evident in the highest quartile (Q4), where dyads with smaller differences in word counts tend to evaluate their interactions more favorably. This suggests that when participants in a dyad both speak more, a smaller difference in the amount of words spoken is associated with higher interaction quality. Conversely, in the lower quartiles (Q1, Q2, and Q3), the relationship between TW_diff and GOI is generally positive, indicating that in less verbose conversations, a larger difference in word count might be perceived more favorably.
Speaking Time (ST) and Speaking Time Difference (ST_diff): Similar to the findings on total words, the difference in speaking time (ST_diff) between dyad members was found to have a positive relationship with interaction quality (GOI). However, the interaction effect between total speaking time (ST) and ST_diff was not statistically significant, suggesting that while speaking time differences do influence interaction quality, this effect does not vary significantly with the overall amount of time spoken.
Speech Rate (WPM) and Speech Rate Difference (WPM_diff): The analysis of speech rate (WPM) and its difference between dyad members (WPM_diff) found no significant impact on interaction quality (GOI). Both the main effects and the interaction terms were not statistically significant, indicating that variations in speech rate, whether absolute or relative between dyad members, do not appear to influence how participants perceive the quality of their interaction.
Implications These findings highlight the nuanced role that speech characteristics play in shaping the perceived quality of social interactions. Specifically, the results suggest that in more verbose conversations (higher TW), maintaining a balance in the amount of speech between participants may be key to fostering positive evaluations of the interaction. This may be because balanced contributions are more likely to be perceived as equitable, thereby enhancing the overall interaction quality.
On the other hand, the lack of significant findings related to speech rate (WPM) implies that while the pace of speech is often considered an important aspect of communication, it may not play a critical role in how interaction quality is assessed in this context. This could be due to the fact that speech rate differences are less noticeable or less impactful in the semi-structured dialogue format used in this study.
Limitations and Future Research One limitation of this study is its reliance on semi-structured dialogues conducted via Zoom, which may not fully capture the dynamics of face-to-face interactions. Additionally, the relatively small sample size (120 dyads) and the exploratory nature of the analysis suggest that these findings should be interpreted with caution and verified in larger, more diverse samples.
Future research could explore these dynamics in different communication contexts, such as in-person interactions or more informal conversational settings, to determine whether these patterns hold across various modes of communication. Moreover, examining other speech characteristics, such as prosody or the use of pauses, could provide a more comprehensive understanding of the factors that contribute to interaction quality.
Conclusion In summary, this study contributes to our understanding of how speech characteristics influence the perceived quality of social interactions. The findings suggest that while the balance of speech within dyads (in terms of word count and speaking time) is important, other factors like speech rate may play a less significant role. These insights could have implications for improving communication strategies in settings where interaction quality is critical, such as in therapy, education, or team collaboration.
Time Domain
Load long format of data, indicating the speaking characteristic within tho four part of the experiment.
data_long <- read.csv("Data/Data_Long.csv")
data_long |>
select(iSubject, Part, Words, SpeakingTime.s, WPM) |>
head()## iSubject Part Words SpeakingTime.s WPM
## 1 101 1 480 160.228 179.7439
## 2 101 2 548 235.757 139.4656
## 3 101 3 540 241.493 134.1654
## 4 101 4 471 208.610 135.4681
## 5 102 1 360 206.901 104.3978
## 6 102 2 467 243.549 115.0487Calculate Similarity
data_long <- data_long |>
mutate(
Words_diff = as.numeric(abs(Words - Words_other)),
ST_diff = abs(SpeakingTime.s - SpeakingTime.s_other),
WPM_diff = abs(WPM - WPM_other))
data_long |>
select(iSubject, Part, Words_diff, ST_diff, WPM_diff) |>
head()## iSubject Part Words_diff ST_diff WPM_diff
## 1 101 1 120 46.673 75.34611
## 2 101 2 81 7.792 24.41692
## 3 101 3 51 15.265 22.57914
## 4 101 4 43 6.215 15.92894
## 5 102 1 120 46.673 75.34611
## 6 102 2 81 7.792 24.41692Time Domain Models
model0_words <- lmer(Words_diff ~ (1|Part), data = data_long)
model0_words_diff <- lmer(Words ~ (1|Part), data = data_long)
model1 <- lmer(GOI ~ Words * Words_diff + (1|Part) , data = data_long)## Warning: Some predictor variables are on very different scales: consider
## rescaling## boundary (singular) fit: see help('isSingular')## Warning: Some predictor variables are on very different scales: consider
## rescalingsjPlot::tab_model(model0_words, model0_words_diff, model1, df.method = "satterthwaite", show.std = TRUE)## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescaling## boundary (singular) fit: see help('isSingular')| Words_diff | Words | GOI | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Predictors | Estimates | std. Beta | CI | standardized CI | p | Estimates | std. Beta | CI | standardized CI | p | Estimates | std. Beta | CI | standardized CI | p | std. p |
| (Intercept) | 119.16 | 0.00 | 102.35 – 135.98 | -0.15 – 0.15 | <0.001 | 368.29 | 0.00 | 327.64 – 408.95 | -0.21 – 0.21 | <0.001 | 50.50 | 0.10 | 45.57 – 55.42 | 0.01 – 0.19 | <0.001 | 0.037 |
| Words | 0.04 | 0.21 | 0.03 – 0.05 | 0.11 – 0.31 | <0.001 | <0.001 | ||||||||||
| Words diff | 0.10 | 0.26 | 0.07 – 0.14 | 0.15 – 0.36 | <0.001 | <0.001 | ||||||||||
| Words × Words diff | -0.00 | -0.19 | -0.00 – -0.00 | -0.27 – -0.12 | <0.001 | <0.001 | ||||||||||
| Random Effects | ||||||||||||||||
| σ2 | 13229.28 | 38125.18 | 326.05 | |||||||||||||
| τ00 | 0.76 Part | 335.33 Part | 0.00 Part | |||||||||||||
| ICC | 0.00 | 0.01 | ||||||||||||||
| N | 4 Part | 4 Part | 4 Part | |||||||||||||
| Observations | 476 | 476 | 476 | |||||||||||||
| Marginal R2 / Conditional R2 | 0.000 / 0.000 | 0.000 / 0.009 | 0.124 / NA | |||||||||||||
model0_ST <- lmer(ST_diff ~ (1|Part), data = data_long)## boundary (singular) fit: see help('isSingular')model0_ST_diff <- lmer(SpeakingTime.s ~ (1|Part), data = data_long)
model1 <- lmer(GOI ~ SpeakingTime.s * ST_diff + (1|Part) , data = data_long)## Warning: Some predictor variables are on very different scales: consider
## rescaling## boundary (singular) fit: see help('isSingular')## Warning: Some predictor variables are on very different scales: consider
## rescalingsjPlot::tab_model(model0_ST, model0_ST_diff, model1, df.method = "satterthwaite", show.std = TRUE)## boundary (singular) fit: see help('isSingular')## Warning: Some predictor variables are on very different scales: consider
## rescaling
## Warning: Some predictor variables are on very different scales: consider
## rescaling## boundary (singular) fit: see help('isSingular')| ST_diff | SpeakingTime.s | GOI | ||||||||||||||
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| Predictors | Estimates | std. Beta | CI | standardized CI | p | Estimates | std. Beta | CI | standardized CI | p | Estimates | std. Beta | CI | standardized CI | p | std. p |
| (Intercept) | 43.37 | -0.00 | 39.62 – 47.11 | -0.09 – 0.09 | <0.001 | 161.76 | 0.00 | 144.95 – 178.58 | -0.22 – 0.22 | <0.001 | 52.77 | 0.05 | 47.53 – 58.01 | -0.04 – 0.14 | <0.001 | 0.272 |
| SpeakingTime s | 0.07 | 0.17 | 0.04 – 0.10 | 0.07 – 0.26 | <0.001 | 0.001 | ||||||||||
| ST diff | 0.23 | 0.25 | 0.14 – 0.33 | 0.15 – 0.35 | <0.001 | <0.001 | ||||||||||
| SpeakingTime s × ST diff | -0.00 | -0.12 | -0.00 – -0.00 | -0.19 – -0.05 | 0.001 | 0.001 | ||||||||||
| Random Effects | ||||||||||||||||
| σ2 | 1729.18 | 5960.46 | 333.67 | |||||||||||||
| τ00 | 0.00 Part | 61.71 Part | 0.00 Part | |||||||||||||
| ICC | 0.01 | |||||||||||||||
| N | 4 Part | 4 Part | 4 Part | |||||||||||||
| Observations | 476 | 476 | 476 | |||||||||||||
| Marginal R2 / Conditional R2 | 0.000 / NA | 0.000 / 0.010 | 0.103 / NA | |||||||||||||