3 Descriptives and Visualizations

3.1 Overview

In this section, I describe and visualize the sample and variables. We have variables on the meta-level (about the survey), the person-level, the app-level, and the day-level. App-level data is in the apps_long data file; all other in the dat data file.

Meta-level

• Duration of the entry survey, when participants reported traits (duration_personality)
• Duration of the exit survey, when participants reported their screen time (duration_screen_time)

Person-level

• Participant identifier (id)
• age in years
• ethnicity
• Notifications of social media apps over the past week (weekly_notifications)
• Basic Psychological Need Satisfaction (autonomy_trait, competence_trait, relatedness_trait) plus their individual items (starting with bpns_)
• Big Five (extraversion, agreeableness, conscientiousness, neuroticsim, openness) plus their individual items (starting with big_five_)

App-level

• What app participants report use for (app)
• On what rank was that app on participants’ top ten (rank)
• Notifications for that app for the week (notifications_per_week)
• Pickups for that app on that day (pickups)
• Screen time for that app on that day (social_media_objective)

Day-level

• Duration of the survey on that day (duration_diary)
• day the survey was answered
• Estimated time on social media on that day (social_media_subjective)
• Estimated pickups of social media apps on that day (pickups_subjective)
• Estimated notifications of social media apps on that day (notifications_subjective)
• Objective time on social media on that day (social_media_objective)
• Objective pickups of social media apps on that day (pickups_objective)
• Well-being on that day (well_being) plus its individual items (starting with low_ and high_)
• Basic psychological needs on that day (autonomy_state, competence_state, relatedness_state) plus their individual items (starting with autonomy_, competence_, relatedness_ respectively)
• Experiences of satisfaction, boredom, stress, enjoyment on that day (satisfied, boring, stressful, enoyable)

3.2 Meta-level

I begin with describing and plotting the duration of the entry and exit surveys. Table 3.1 shows descriptive stats; Figure 3.1 shows that twp participants had their entry surveys open for a day before pressing send, which skews the mean massively. However, those people’s data look good, so I wouldn’t exclude them here. Note: Colors are from here.

Table 3.1: Duration of entry and exit surveys

variable	mean	sd	median	min	max	range	cilow	cihigh
duration_personality	1H 5M 56S	4H 36M 55S	15M 40S	1M 22S	1d 9H 33M 36S	1d 9H 32M 14S	9M 50S	2H 2M 3S
duration_screen_time	20M 21S	20M 8S	13M 18S	1M 8S	1H 51M 29S	1H 50M 21S	16M 17S	24M 26S

Figure 3.1: Duration of surveys

3.3 Person-level

Let’s have a look at the final sample. Overall, our sample size is N = 96. The sample has a mean age of M = 20.45, SD = 1.32. The age ranges from 18 to 25 The sample consists mostly of women (66 women, 30 men, and one non-binary participant).

Most participants are Asian, followed by White, Black, and Hispanic, see Table 3.2

Table 3.2: Ethnicity of the sample

ethnicity	count	percent
Asian	40	42
White	26	27
Black or African American	11	11
Hispanic or Latino	10	10
Multiracial	6	6
NA	2	2
Native Hawaiian or Other Pacific Islander	1	1

Alright, next we look at the objective count of notifications over the past week, aggregated across all apps. Table 3.3 shows that participants received quite a lot of notifications from social media apps only. That distribution is heavily skewed (Figure 3.2 by a couple of participants who received several thousand notifications over the week.

Table 3.3: Weekly notifications (objective) across all apps

variable	mean	sd	median	min	max	range	cilow	cihigh
weekly_notifications	997.6354	673.4178	868.5	126	3382	3256	861.1883	1134.083

Figure 3.2: Weekly notifications (objective) across all apps

Now we look at the trait variables: the basic psychological need satisfaction and the big five. Note that I follow recent recommendations and calculate \(\omega\) for reliability. Table 3.4 shows the descriptive information of the three psychological needs and the big five. Figure 3.3 shows their distribution. The sample isn’t too large, so considering the small size, I’d say everything looks pretty good.

Table 3.4: Descriptives for trait variables

variable	mean	sd	median	min	max	range	cilow	cihigh	omega
autonomy_trait	4.51	0.88	4.50	1.25	6.88	5.62	4.33	4.68	0.82
competence_trait	5.16	0.70	5.31	3.50	6.50	3.00	5.01	5.30	0.75
relatedness_trait	4.54	1.03	4.56	2.38	6.62	4.25	4.33	4.75	0.88
extraversion	3.13	0.74	3.00	1.62	5.00	3.38	2.98	3.29	0.86
agreeableness	3.68	0.56	3.59	2.33	5.00	2.67	3.57	3.80	0.75
conscientiousness	3.50	0.58	3.50	2.00	4.89	2.89	3.39	3.62	0.75
neuroticism	3.29	0.50	3.38	2.00	4.38	2.38	3.19	3.39	0.59
openness	3.54	0.48	3.60	2.10	4.80	2.70	3.45	3.64	0.66

Figure 3.3: Distribution of trait variables

In Figure 3.4 we see the correlations between those traits. As expected psychological needs are correlated highly with each other. Credit for the lm lines goes to data prone, whose idea I adapted.

Figure 3.4: Correlation matrix of trait level variables

3.4 App-level

First, Figure 3.5 shows what apps mostly nominated (i.e., used). We see that out of the sample, most participants had Messaging, Snapchat, Whatsapp etc. as part of their top ten.

Figure 3.5: Percentage of nominated apps

Next, I visualize how many minutes each of those apps was used across the sample. For that, I need to reshape the data a bit to get the mean minutes per app across all days and all participants. In Figure 3.6, I show the average objective time per app. Note that the CIs are across the entire data and not nested by app or day. Also, a high mean doesn’t mean that much because it could just be from one participant who used it a lot on two days. The size of the points shows how often an app was reported across the entire sample. For apps that only had one entry, those CI will be nonexistent. In addition, I now exclude entries on social_media_objective that have NA. The NA here can mean participants just didn’t fill in anything, or they had zero duration on that day. Because adding up the raw scores across apps was so close to the daily total, I’ll exclude NAs here.

Figure 3.6: Average daily objective time for all apps across participants and days

I’ll do the same for objective pickups per app, averaged across day and participant. Figure 3.7 shows that the same apps that got a lot of screen time had a lot of pickups.

Figure 3.7: Average daily pickups for all apps across participants and days

Last, I check which apps got the most notifications over the week in Figure 3.8, on average. It’s interesting to see that Facebook had a lot of screen time and pickups, but much fewer notifications. Also, these notifications are per week, and not per day, as the previous two figures.

Figure 3.8: Average notifications per week for all apps across participants

3.5 Day level

Alright, we’re at the most interesting section, the daily surveys. I first look at how long people typically took for a survey. Table 3.5 shows that the mean is highly skewed because of outliers and the median more appropriate to describe the duration. In Figure 3.9 we see that a couple of people took a long time from opening to submitting the survey. I checked those participants who took a long time in the data processing section. The maximum duration here from someone who didn’t open the survey on a Friday. So that duration is just the survey closing automatically after two days, which really drives up the mean.

Table 3.5: Duration of daily surveys

variable	mean	sd	median	min	max	range	cilow	cihigh
duration_diary	50M 22S	2H 44M 17S	16M 0S	25S	2d 0H 27M 48S	2d 0H 27M 23S	35M 5S	1H 5M 40S

Figure 3.9: Duration of daily surveys

Alright, next I inspect overall response rate in the final sample, aka how many valid surveys do we have among the final sample. Each participant received five surveys, one for each day, so 96 participants x 5 = 480. We have 435 surveys in the final sample where participants actually responded, which means a 91% response rate among the final sample.

Let’s inspect response rate per day. As is to be expected, participants lost motivation over the course of the week. However, even the response rate on Friday is really high (at least among our sample of valid responses). We should still consider to take the day grouping into account when modelling the data later in the analysis.

Figure 3.10: Survey responses per day

Next, I describe and plot the distributions of the social media use variables. The distribution and CI is of the entire sample, not aggregated by participant or day first. Table 3.6 shows that participants weren’t too far off in their estimates, which is interesting. As expected (Figure 3.11), the social media variables are a bit skewed, but overall, they look fine.

Table 3.6: Descriptive information on social media variables

variable	mean	sd	median	min	max	range	cilow	cihigh
social_media_subjective	153	112	130	5	692	687	143	164
social_media_objective	139	94	118	2	565	563	130	148
error	46	156	-3	-96	1186	1282	31	61
pickups_subjective	34	43	17	0	259	259	30	38
pickups_objective	49	31	44	0	196	196	46	52
notifications_subjective	61	99	30	0	700	700	52	71

Figure 3.11: Distribution of social media variables

I also want to see how much variability there is between the objective and subjective measures. In Figure 3.12 we see per participant the difference between objective and subjective social media use. The numbers in the grey box show whether the subjective report is an underestimate (negative number) or overestimate (positive number). Inspiration for the plot from here and here.

Figure 3.12: Difference between subjective and objective social media use (difference in grey box)

Now let’s look at the state well-being and psychological needs variables plus the four experiences (e.g., boring). Again, I calculate \(\omega\), but this time for the entire sample in Table 3.7. That will necessarily bias the estimate because there’s multiple measures per person. I’m not aware of a consensus reliability procedure for repeated measures. Figure 3.13 shows that the data look pretty good.

Table 3.7: Descriptives for state variables

variable	mean	sd	median	min	max	range	cilow	cihigh	omega
well_being_state	3.23	0.70	3.17	1.25	5	3.75	3.17	3.30	0.85
autonomy_state	4.50	1.16	4.25	1.00	7	6.00	4.39	4.60	0.69
competence_state	4.60	1.27	4.50	1.00	7	6.00	4.48	4.72	0.79
relatedness_state	5.33	1.07	5.50	2.50	7	4.50	5.23	5.43	0.70
satisfied	4.61	1.33	5.00	1.00	7	6.00	4.49	4.74	NA
boring	3.40	1.55	3.00	1.00	7	6.00	3.25	3.55	NA
stressful	3.96	1.81	4.00	1.00	7	6.00	3.79	4.13	NA
enjoyable	4.27	1.41	4.00	1.00	7	6.00	4.14	4.40	NA

Figure 3.13: Distribution of state variables

In Figure 3.14 we see the correlations between variables on the state level.

Figure 3.14: Correlation matrix of state level variables. social = screen time on social media; _s = subjective; _o = objective; not = notifications

In Figure 3.15 we see the correlations between use variables on the state level and the trait level.

Figure 3.15: Correlation matrix of use variables (state) and personality traits. social = screen time on social media; _s = subjective; _o = objective; not = notifications; extra = extraversion; agree = agreeableness; con = conscientiousness; neuo = neuroticism; open = openness

3.6 Demographics and social media

To compare our findings with previous research which found gender and age differences in social media use, we also check for those differences in our data set. We don’t report those results in the paper because a) they’re less relevant to our research questions, b) space limits.

First, let’s create a correlation matrix between age and the social media use variables. Figure 3.16 shows that age is pretty much unrelated to the social media use variables, possibly because the age range is quite narrow.

Figure 3.16: Correlation between age and social media use

Next, I visualize differences in the three social media use variables between genders. Figure (fig:visualize-gender-differences) shows little visual indication that there might be differences.

Figure 3.17: Comparison of social media variables between gender

I’ll test that formally with a t-test on the aggregated data. For none of the outcomes is the difference between genders significant.

my_t_test <- function(model) {
  out1 <- broom::tidy(model) %>% 
    select(
      contains("estimate"),
      statistic,
      p.value
    ) %>% 
    rename(
      Difference = estimate,
      Men = estimate1,
      Women = estimate2,
      p = p.value
    )
}

tmp %>% 
  group_by(variable) %>% 
  group_modify(~my_t_test(t.test(value ~ gender, data = .x))) %>% 
  rename(Outcome = variable) %>% 
  kable(digits = c(2,2,2,2))

Outcome	Difference	Men	Women	statistic	p
error	15.71	59.50	43.79	0.50	0.62
social_media_objective	-13.12	131.38	144.50	-0.68	0.50
social_media_subjective	-11.47	147.57	159.04	-0.52	0.60

3.7 Plots for paper

Here, I’ll create summary figures for the paper. I’ll begin with plotting the traits.

For the plot, the data need to be in the long format.

tmp <- 
  dat %>% 
  group_by(id) %>% 
  slice(1) %>% 
  ungroup() %>% 
  select(all_of(c("id", trait_descriptives$variable))) %>% 
  pivot_longer(
    -id,
    names_to = "variable",
    values_to = "value"
  )

rename_levels <- c(
  "Autonomy",
  "Competence",
  "Relatedness",
  "Agreeableness",
  "Conscientiousness",
  "Extraversion",
  "Neuroticism",
  "Openness"
)

my_string <- "_trait"

# reorder and rename factor levels
clean_plot_data <- 
  function(
    dat,
    levels_to_rename,
    string_to_remove
  ){
    dat <- 
      dat %>% 
      mutate(
        # in case it's social media variables
        variable = case_when(
          variable == "social_media_objective" ~ "Objective (h)",
          variable == "social_media_subjective" ~ "Subjective (h)",
          variable == "error" ~ "Accuracy (%)",
          TRUE ~ variable
        ),
        # remove _trait at the end and capitalize
        variable = str_to_sentence(str_remove(variable, string_to_remove)),
        variable = as.factor(variable),
        variable = str_replace(variable, "_", "-"),
        
        # reorder factor levels
        variable = fct_relevel(
          variable,
          levels_to_rename
        )
      )
    return(dat)
  }

tmp <- clean_plot_data(tmp, rename_levels, my_string)
trait_descriptives <- clean_plot_data(trait_descriptives, rename_levels, my_string)

Okay, we already have the aggregated info in trait_descriptives, so we can get to plotting.

# function for breaks
my_breaks <- 
  function(x) {
    if (max(x) > 5){
      1:7
    } else {
    1:5
    }
  }

# function for limits
my_limits <- 
  function(x) {
    if (max(x) > 5){
      c(1,7)
    } else {
    c(1,5)
    }
  }

# color palette (not needed anymore after review that colors made the figure harder to see)
cb_palette <- c("#000000", "#E69F00", "#56B4E9", "#009E73", "#F0E442", "#0072B2", "#D55E00", "#CC79A7")

# plot
ggplot(
  tmp,
  aes(
    x = value,
    y = 1
  )
) +
  geom_quasirandom(groupOnX=FALSE, size = 0.7, shape = 20, color = "black") +
  facet_wrap(
    ~ variable, 
    scales = "free_x"
  ) +
  scale_x_continuous(breaks = my_breaks, limits = my_limits) +
  geom_text(
    data = trait_descriptives,
    aes(
      x = 1.6,
      y = 1.4,
      label = paste0("M = ", mean),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black"
  ) +
  geom_text(
    data = trait_descriptives,
    aes(
      x = 1.6,
      y = 1.3,
      label = paste0("SD = ", sd),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black"
  ) +
  geom_text(
    data = trait_descriptives,
    aes(
      x = 1.6,
      y = 1.2,
      label = paste0("\u03a9 = ", omega),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black"
  ) +
  theme_cowplot() +
  # scale_colour_manual(values=cb_palette) +
  # scale_fill_manual(values = cb_palette) +
  theme(
    axis.text.y = element_blank(),
    axis.title.x = element_blank(),
    axis.title.y = element_blank(),
    axis.ticks.y = element_blank(),
    axis.line.y = element_blank(),
    strip.background.x = element_blank(),
    strip.background.y = element_blank(),
    legend.position = "none",
    text = element_text(family = "Corbel")
  ) -> figure1

figure1

ggsave(
  here("figures", "figure1.tiff"),
  plot = figure1,
  width = 21 * 0.8,
  height = 29.7 * 0.4,
  units = "cm",
  dpi = 300
)

Okay, next the state variables.

tmp <- 
  dat %>% 
  select(all_of(c("id", state_descriptives$variable))) %>% 
  pivot_longer(
    -id,
    names_to = "variable",
    values_to = "value"
  )

rename_levels <- c(
  "Autonomy",
  "Competence",
  "Relatedness",
  "Boring",
  "Enjoyable",
  "Satisfied",
  "Stressful",
  "Well-being"
)

my_string <- "_state"

tmp <- clean_plot_data(tmp, rename_levels, my_string)
state_descriptives <- clean_plot_data(state_descriptives, rename_levels, my_string)

# position of the text (so it doesn't overlap with points)
state_descriptives <- 
  state_descriptives %>% 
  mutate(
    x_position = 1.6,
    y_position = 1.45
  )

Then to plotting.

# plot
ggplot(
  tmp,
  aes(
    x = value,
    y = 1
  )
) +
  geom_quasirandom(groupOnX=FALSE, size = 0.7, shape = 20, color = "black") +
  facet_wrap(
    ~ variable, 
    scales = "free_x"
  ) +
  scale_x_continuous(breaks = my_breaks, limits = my_limits) +
  geom_text(
    data = state_descriptives,
    aes(
      x = x_position,
      y = y_position + 0.1,
      label = paste0("M = ", mean),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black"
  ) +
  geom_text(
    data = state_descriptives,
    aes(
      x = x_position,
      y = y_position,
      label = paste0("SD = ", sd),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black"
  ) +
  geom_text(
    data = state_descriptives,
    aes(
      x = x_position,
      y = y_position - 0.1,
      label = paste0("\u03a9 = ", omega),
      family = "Corbel"
    ),
    size = 2.5,
    color = "black",
    alpha = if_else(is.na(state_descriptives$omega), 0, 1) # one-item measure don't have omega, so I make those see through
  ) +
  theme_cowplot() +
  # scale_colour_manual(values=cb_palette) +
  # scale_fill_manual(values = cb_palette) +
  theme(
    axis.text.y = element_blank(),
    axis.title.x = element_blank(),
    axis.title.y = element_blank(),
    axis.ticks.y = element_blank(),
    axis.line.y = element_blank(),
    strip.background.x = element_blank(),
    strip.background.y = element_blank(),
    legend.position = "none",
    text = element_text(family = "Corbel")
  ) -> figure2

figure2

ggsave(
  here("figures", "figure2.tiff"),
  plot = figure2,
  width = 21 * 0.8,
  height = 29.7 * 0.4,
  units = "cm",
  dpi = 300
)

Alright, last the smartphone use variables.

tmp <- 
  dat %>% 
  select(all_of(c("id", "social_media_objective", "social_media_subjective", "error"))) %>%
  # turn to hours
  mutate(
    across(
      contains("social_media"),
      ~ .x /60
    )
  ) %>% 
  pivot_longer(
    -id,
    names_to = "variable",
    values_to = "value"
  )

rename_levels <- c(
  "Objective (h)",
  "Subjective (h)",
  "Accuracy (%)"
)

my_string <- "_state"

tmp <- clean_plot_data(tmp, rename_levels, my_string)
social_media2 <- clean_plot_data(social_media, rename_levels, my_string) %>% 
  filter(variable %in% c("Objective (h)", "Subjective (h)", "Accuracy (%)")) %>% 
  # add x axis position for geom_text
  mutate(
    x_position = case_when(
      variable == "Accuracy (%)" ~ 1200*0.8,
      TRUE ~ 0.8*13
    )
  )

And the plot.

# function for breaks
my_breaks <- 
  function(x) {
    if (max(x) < 100){
      seq(0, 13, 2)
    } else {
    c(-200, 0, 400, 800, 1200)
    }
  }

# function for limits
my_limits <- 
  function(x) {
    if (max(x) < 100){
      c(0, 13)
    } else {
    c(-200, 1200)
    }
  }

# plot
ggplot(
  tmp,
  aes(
    x = value,
    y = 1
  )
) +
  geom_quasirandom(groupOnX=FALSE, size = 0.7, shape = 20, color = "black") +
  facet_wrap(
    ~ variable, 
    scales = "free_x"
  ) +
  scale_x_continuous(breaks = my_breaks, limits = my_limits) +
  geom_text(
    data = social_media2,
    aes(
      x = x_position,
      y = 0.7,
      label = paste0("M = ", mean),
      family = "Corbel"
    ),
    size = 3,
    color = "black"
  ) +
  geom_text(
    data = social_media2,
    aes(
      x = x_position,
      y = 0.67,
      label = paste0("SD = ", sd),
      family = "Corbel"
    ),
    size = 3,
    color = "black"
  ) +
  theme_cowplot() +
  scale_colour_manual(values=cb_palette) +
  scale_fill_manual(values = cb_palette) +
  theme(
    axis.text.y = element_blank(),
    axis.title.x = element_blank(),
    axis.title.y = element_blank(),
    axis.ticks.y = element_blank(),
    axis.line.y = element_blank(),
    strip.background.x = element_blank(),
    strip.background.y = element_blank(),
    legend.position = "none",
    text = element_text(family = "Corbel")
  ) -> figure3

figure3

ggsave(
  here("figures", "figure3.tiff"),
  plot = figure3,
  width = 21 * 0.8,
  height = 29.7 * 0.4,
  units = "cm",
  dpi = 300
)