2 Preprocess Data

2.1 Load Libraries and Data

Show the code (libraries)

if (!require("pacman")) install.packages("pacman")
library(pacman)

p_load(tidyverse, mctq, data.table, pwr)

set.seed(8675309)

Show the code (load data)

synDiary <- read_csv("data/data-synthetic-raw/synDiary.csv.gz")

# for unknown reasons, read_csv imports some columns incorrectly, so we use data.table::fread
synPanel <- fread("data/data-synthetic-raw/synPanel.csv.gz") 
synIntake <- read_csv("data/data-synthetic-raw/synIntake.csv.gz")

synNintendo <- read_csv("data/data-synthetic-raw/synNintendo.csv.gz") |>
  mutate(sessionEnd = sessionStart + minutes(round(duration)), .after = sessionStart)
synXbox <- read_csv("data/data-synthetic-raw/synXbox.csv.gz") |>
  mutate(sessionEnd = sessionStart + minutes(round(duration)), .after = sessionStart)
synSteam <- read_csv("data/data-synthetic-raw/synSteam.csv.gz")
syniOS <- read_csv("data/data-synthetic-raw/syniOS.csv.gz")
synAndroid <- read_csv("data/data-synthetic-raw/synAndroid.csv.gz")

2.2 Clean intake

Here we:

define a socio-economic status (SES) index
calculate total height in inches
calculate BMI using the formula (weight / (total_inches^2)) * 703

These will be used as covariates in Study 3.

Show the code (clean intake)

synIntakeClean <- synIntake |>
  # define socio-economic status (SES) index
  mutate(
    # Assigning scores to employment categories
    empScore = case_when(
      employment == "Full-Time" ~ 5,
      employment == "Part-Time" ~ 4,
      employment == "Due to start a new job within the next month" ~ 3,
      employment == "Not in paid work (e.g. homemaker', 'retired or disabled)" ~ 2,
      employment == "Unemployed (and job seeking)" ~ 1,
      employment == "Other" ~ NA_real_ ,  # Adjust based on context if necessary
      TRUE ~ NA_real_  # Handle any other cases
    ),
    
    # Assigning scores to education levels
    eduScore = case_when(
      eduLevel == "Graduate or professional degree (MA, MS, MBA, PhD, etc)" ~ 7,
      eduLevel == "University Bachelors Degree" ~ 6,
      eduLevel == "Some University but no degree" ~ 5,
      eduLevel == "Vocational or Similar" ~ 4,
      eduLevel == "Completed Secondary School" ~ 3,
      eduLevel == "Some Secondary" ~ 2,
      eduLevel == "Completed Primary School" ~ 1,
      eduLevel == "Some Primary" ~ 1,
      eduLevel == "Prefer not to say" ~ NA_real_,  # Treat as missing
      TRUE ~ NA_real_  # Handle any other cases
    ),
    
    # Combining the scores into a SES index
    SES_index = empScore + eduScore
  ) |> 
  mutate(
    # Calculate total height in inches
    total_inches = `height#1_1_1` * 12 + `height#1_1_2`,
    
    # Calculate BMI using the formula
    bmi = (weight / (total_inches^2)) * 703,

    age_scaled = as.numeric(scale(age, center = TRUE, scale = TRUE)),
    bmi_scaled = as.numeric(scale(bmi, center = TRUE, scale = TRUE)),
    SES_index_scaled = as.numeric(scale(SES_index, center = TRUE, scale = TRUE))
  )

2.3 Clean diary

Here we:

recode the diary data to numeric values
calculate mean scores of relevant variables
calculate within- and between-person centered variables
recode the displaced activity data into categories (randomly assigned, for now)
calculate some variables based on the telemetry

Show the code (clean diary)

synDiaryClean <- synDiary |>
  mutate(
    across(starts_with(c("bpnsfs", "bangs")), ~ case_when(
      . %in% c("1 \nStrongly Disagree", "1 - Not at all true", "1 - very strongly disagree") ~ 1,
      . %in% c("2", "2 - strongly disagree") ~ 2,
      . %in% c("3", "3 - disagree") ~ 3,
      . %in% c("4Neither Agree nor Disagree", "4", "4 - neither disagree nor agree") ~ 4,
      . %in% c("5", "5 - Completely true", "5 - agree") ~ 5,
      . %in% c("6") ~ 6,
      . %in% c("7 Strongly agree") ~ 7,
      TRUE ~ NA_integer_
    ))
  ) %>%
  # calculate mean scores of relevant variables (there is no missing data within waves)
  mutate(
    globalNS = rowMeans(select(., bpnsfs_1:bpnsfs_3), na.rm = TRUE),
    globalNF = rowMeans(select(., bpnsfs_4:bpnsfs_6), na.rm = TRUE),
    gameNS = rowMeans(select(., bangs_1:bangs_3), na.rm = TRUE),
    gameNF = rowMeans(select(., bangs_4:bangs_6), na.rm = TRUE)
  ) |>
  # Calculate within- and between-person centered variables
  group_by(pid) %>%
  mutate(across(
    c(globalNS, globalNF, gameNS, gameNF),
    list(
      cw = ~ . - mean(., na.rm = TRUE),
      cb = ~ mean(., na.rm = TRUE)
    )
  )) %>%
  ungroup() %>%
  mutate(across(
    ends_with("cb"),
    ~ . - mean(., na.rm = TRUE)
  )) |>
  # to understand displaced activities, we will manually code the true
  # participant activity data into categories. We pre-define 5 problematic
  # displacement categories (work/school, social engagements, sleep, eating,
  # fitness, caretaking) and one catch-all category (other), which may later be
  # broken down into subcategories.
  mutate(
    displacedActivityCategory = ifelse(!is.na(displacedActivity),
      sample(c("work/school", "social engagements", "sleep", "eating", "fitness", "caretaking", "other"),
        n(),
        prob = c(.05, .05, .05, .05, .05, .05, .75),
        replace = TRUE
      ),
      NA_character_
    ),
    displacedCoreDomain = ifelse(displacedActivityCategory %in% c(
      "work/school", "social engagements",
      "sleep", "eating", "fitness", "caretaking", "other"
    ),
    TRUE,
    FALSE
    ),
    .after = displacedActivity
  )

2.4 Clean panel

Here we want to: - recode the panel data to numeric values - calculate mean scores of relevant variables - calculate within- and between-person centered variables

2.4.1 Chronotype

Chronotype or sleep-corrected local time of mid-sleep on work-free days msf_sc() allows you to compute the chronotype, or corrected local time of mid-sleep on work-free days. It takes five arguments: msf (local time of mid-sleep on work-free days), sd_w (sleep duration on workdays), sd_f (sleep duration on work-free days), sd_week(average weekly sleep duration), and alarm_f (a logical object indicating if the respondent uses an alarm clock to wake up on work-free days).

If sd_f is less or equal than sd_w, the output must be msf. Else, it must return msf minus the difference between sd_f and sd_week divided by 2. msf_sc can only be computed if alarm_f is equal to FALSE (the function will return NA when alarm_f == TRUE).

msf_sc applies a correction to msf, removing an estimation of the effect from accumulated sleep debt on workdays that usually is compensated on work-free days. See ?msf_sc to learn more.

Show the code (clean panel)

synPanelClean <- synPanel |>
  mutate(
    across(starts_with(c("bangs", "wemwbs", "promis", "trojan", "BFI", "eps")), ~ case_when(
      . %in% c("Greatly interfered") ~ -3,
      . %in% c("Moderately interfered") ~ -2,
      . %in% c("Slightly interfered") ~ -1,
      . %in% c("No impact","No chance of dozing") ~ 0,
      . %in% c(
        "1 \nStrongly Disagree", "1 - Not at all true", "1 - None of the time", "Never", "1 - Strongly disagree",
        "Disagree strongly", "Slightly supported", "Slight chance of dozing"
      ) ~ 1,
      . %in% c("2", "2 - Rarely", "Rarely", "Disagree a little", "Moderately supported", "Moderate chance of dozing") ~ 2,
      . %in% c("3", "3 - Some of the time", "Sometimes", "Neutral; no opinion", "Greatly supported", "High chance of dozing") ~ 3,
      . %in% c("4Neither Agree nor Disagree", "4", "4 - Often", "Often", "Agree a little") ~ 4,
      . %in% c("5", "5 - Completely true", "5 - All of the time", "Always", "5 - Strongly agree", "Agree strongly") ~ 5,
      . %in% c("6") ~ 6,
      . %in% c("7 Strongly agree") ~ 7,
      TRUE ~ NA_integer_
    ))
  ) %>%
  # calculate mean/sum scores of relevant variables (there is no missing data within waves)
  mutate(
    wemwbs = rowMeans(select(., wemwbs_1:wemwbs_7), na.rm = TRUE),
    promis = rowMeans(select(., promis_1:promis_8), na.rm = TRUE),
    gameNS = rowMeans(select(., bangs_1:bangs_3, bangs_7:bangs_9, bangs_13:bangs_15), na.rm = TRUE),
    gameNF = rowMeans(select(., bangs_4:bangs_6, bangs_10:bangs_12, bangs_16:bangs_18), na.rm = TRUE),
    epsTotal = rowSums(select(., eps_1_1:eps_1_8), na.rm = TRUE),
  ) |>
  # Calculate within- and between-person centered variables
  group_by(pid) %>%
  mutate(across(
    c(wemwbs, promis, gameNS, gameNF),
    list(
      cw = ~ . - mean(., na.rm = TRUE),
      cb = ~ mean(., na.rm = TRUE)
    )
  )) %>%
  ungroup() %>%
  mutate(across(
    ends_with("cb"),
    ~ . - mean(., na.rm = TRUE)
  )) |> 

  # calculate amount of sleep
  mutate(
    psqi_4_1_1_1_hours = as.numeric(`psqi_4#1_1_1`), # Convert hours to numeric
    psqi_4_1_1_2_hours = as.numeric(`psqi_4#1_1_2`) / 60, # Convert minutes to numeric hours
    total_hours_sleep = psqi_4_1_1_1_hours + psqi_4_1_1_2_hours, # Compute total hours of sleep
    .keep = "unused"
  ) |> 

  # ~~~~~~~~~~~~~~~~~~~~~~~~~
  # calculate chronotype ####
  # ~~~~~~~~~~~~~~~~~~~~~~~~
  
  # rename mctq_1 to work  and turn it into a logical variable where 1 is 1 and 2 is 0, other is NA
   mutate(work = case_when(
    mctq_1 == 1 ~ 1,
    mctq_1 == 2 ~ 0,
    TRUE ~ NA_real_
   )) |> 
   rename(
    wd = mctq_2_1, 
    bt_w = mctq_3_1, 
    sprep_w = mctq_3_3, 
    slat_w = mctq_3_4, 
    se_w = mctq_3_5, 
    si_w = mctq_3_6,
    bt_f = mctq_6_1, 
    sprep_f = mctq_6_3, 
    slat_f = mctq_6_4, 
    se_f = mctq_6_5, 
    si_f = mctq_6_6,
    reasons_why_f = mctq_8_1
  ) |> 
  mutate(
    alarm_w = case_when(
      mctq_4_1 == "Yes" ~ 1,
      mctq_4_1 == "No" ~ 0,
      TRUE ~ NA_real_),
    wake_before_w = case_when(
      mctq_5_1 == "Yes" ~ 1,
      mctq_5_1 == "No" ~ 0,
      TRUE ~ NA_real_
    ),
    alarm_f = case_when(
      mctq_7_1 == "Yes" ~ 1,
      mctq_7_1 == "No" ~ 0,
      TRUE ~ NA_real_
    ),
    reasons_f = case_when(
      mctq_7_2 == "Yes" ~ 1,
      mctq_7_2 == "No" ~ 0,
      TRUE ~ NA_real_
    ),
  ) |>
  mutate(
    across("wd", as.integer),
    across(matches("^work$|^alarm_|^wake_|^reasons_f$"), as.logical),
    across(matches("^bt_|^sprep_|^se_"), hms::parse_hm),
    across(matches("^slat_|^si_"), ~ dminutes(as.numeric(.x)))
  ) |> 
  # Calculate sleep onset
  mutate(
    so_w = mctq::so(sprep_w, slat_w),
    so_f = mctq::so(sprep_f, slat_f)
  ) |>
  # Calculate sleep duration
  mutate(
    sd_w = mctq::sdu(so_w, se_w),
    sd_f = mctq::sdu(so_f, se_f)
  ) |>
  # Calculate midsleep time
  mutate(
    msw = mctq::msl(so_w, sd_w),
    msf = mctq::msl(so_f, sd_f)
  ) |>
  # Calculate weekly sleep duration
  mutate(
    sd_week = mctq::sd_week(sd_w, sd_f, wd)
  ) |> 
  # Calculate chronotype
  mutate(
    msf_sc = mctq::msf_sc(msf, sd_w, sd_f, sd_week, alarm_f),
    msf_sc_numeric = as.numeric(msf_sc) / 3600
  )

2.5 Process telemetry

Here we: - join the survey data to each telemetry table - filter for only the sessions/rows that happened immediately after the survey was completed - join these back together with a binary indicator of whether at least one session occurred

Show the code (process telemetry)

# TODO: make sure that sessions happening as late as 6 hours later are still included, even if this isn't on the same day

# in the below, we join the survey data to each telemetry table,
# and filter for only the sessions/rows that happened immediately after the survey was completed.
# after, we join these back together with a binary indicator of whether at least one session occurred

nintendoOverlaps <- synDiaryClean %>%
  left_join(synNintendo, by = c("pid", "day", "date")) |>
  filter(
    sessionEnd >= surveyCompletionTime | # Session ended after the survey time
      sessionStart <= surveyCompletionTime + days(1) # Session started before the end of the time window
  ) |>
  group_by(pid, day, date) |>
  summarize(playedLaterNintendo = TRUE, .groups = "drop")

xboxOverlaps <- synDiaryClean %>%
  left_join(synXbox, by = c("pid", "day", "date")) |>
  filter(
    sessionEnd >= surveyCompletionTime | # Session ended after the survey time
      sessionStart <= surveyCompletionTime + days(1) # Session started before the end of the time window
  ) |>
  group_by(pid, day, date) |>
  summarize(playedLaterXbox = TRUE, .groups = "drop")

steamOverlaps <- synDiaryClean %>%
  left_join(synSteam, by = c("pid", "day", "date")) |>
  filter(
    sessionStart <= surveyCompletionTime + days(1) # Session started before the end of the time window
  ) |>
  group_by(pid, day, date) |>
  summarize(playedLaterSteam = TRUE, .groups = "drop")


# Step 3: Determine if any Nintendo sessions occurred in the time window for each row in df
synDiaryClean <- synDiaryClean |>
  left_join(
    nintendoOverlaps,
    by = c("pid", "day", "date")
  ) |>
  left_join(
    xboxOverlaps,
    by = c("pid", "day", "date")
  ) |>
  left_join(
    steamOverlaps,
    by = c("pid", "day", "date")
  ) |>
  mutate(
    playedLaterNintendo = if_else(is.na(playedLaterNintendo), FALSE, playedLaterNintendo),
    playedLaterXbox = if_else(is.na(playedLaterXbox), FALSE, playedLaterXbox),
    playedLaterSteam = if_else(is.na(playedLaterSteam), FALSE, playedLaterSteam),
    playedLaterAny = ifelse(playedLaterNintendo | playedLaterXbox | playedLaterSteam, TRUE, FALSE)
  )

2.6 Exclusion criteria

We will exclude any telemetry rows wherein players have logged more than 24 hours of playtime on one platform in any single day, or where sessions have taken place in the future, indicating a technical problem or manipulation of the system clock for in-game benefits.

We will further include an attention check in the panel surveys whereby participants are given a random duplicated item from the need satisfaction and frustration measure. Responses where the two duplicate items differ by more than 1 scale point will be flagged for manual inspection of potential careless responding.

We have 4 exclusion criteria total:

No telemetry rows wherein players have logged more than 16 hours of playtime across linked platforms in any single day
No sessions longer than 8 hours
No sessions taking place in the future, relative to when data is collected
No responses failing the attention check (> 1 scale point difference between two duplicated items)

Show the code (exclusion criteria)

# TODO: update exclusion criteria

excluded <- bind_rows(synNintendo, synXbox, synSteam) |> 
  group_by(pid, day) |> 
  summarise(dailyPlay = sum(duration)) |> 
  filter(dailyPlay >= 1440)

synNintendoClean <- synNintendo |> 
  filter(sessionStart <= Sys.time()) |> # no future sessions
  anti_join(excluded, by = c("pid", "day"))

synSteamClean <- synSteam |> 
  filter(sessionStart <= Sys.time()) |> # no future sessions
  anti_join(excluded, by = c("pid", "day"))

synXboxClean <- synXbox |>
  filter(sessionStart <= Sys.time()) |> # no future sessions
  anti_join(excluded, by = c("pid", "day"))

synAndroidClean <- synAndroid |>
  filter(category == "Games") |> 
  mutate(sessionStart = as_datetime(paste(date, sessionStart)),
         sessionEnd = as_datetime(paste(date, sessionEnd))) |>
  rename(titleID = app)

2.7 Positive Control Power

In the manuscript, we specify a variety of positive controls—tests that must be passed in order to justify adhering to our proposed analysis strategy.

It is important that our design has sufficient statistical power to detect such effects if they exist. Here we calculate the power for each of our positive controls.

Applicable Study	Test	Statistical Power
All studies	A significant positive correlation between self-reported video game play and digital trace playtime during the previous 2 weeks	Assuming n = 9,300 panel surveys (after 10% wave-on-wave attrition), a true population value of r = .2, an alpha of .05, and a one-sided test, power > 99%
All studies	There will be no overlapping sessions for a given individual on Nintendo or Xbox (we allow for possible overlap across different platforms, in case the user has two devices active simultaneously) AND there will be no cases where a player logs more than 60 minutes of playtime on Steam between adjacent hourly measurements	(N/A; fails if a single case occurs)
Study 1	Significant positive correlation between need satisfaction in general and daily life satisfaction	Assuming n = 21,000 diary surveys (after 30% total attrition), a true population value of r = .2, and an alpha of .05, power > 99%
Study 2	Significant positive correlation between social jetlag as calculated by the Munich Chronotype Questionnaire and daytime sleepiness.	Assuming n = 4,440 panel surveys with sleep measures (Waves 2, 4, 6 only + 10% wave-on-wave attrition), a true population value of Spearman’s rho = .1, an alpha of .05, and a one-sided test, power > 99%
Study 2	Significant negative correlation between (Pittsburgh Sleep Quality Index sleep quality component) and Warwick-Edinburgh Mental Well-being Scale (WEMWBS).	Assuming n = 4,440 panel surveys with sleep measures (Waves 2, 4, 6 only + 10% wave-on-wave attrition), a true population value of Spearman’s rho = -.1, an alpha of .05, and a one-sided test, power > 99%
Study 3	Significantly higher playtime in shooter games for men as compared to women (Lange et al., 2021)	Assuming telemetry data for n = 1,000 (as attrition during surveys does not prevent us from collecting gameplay data), a true population difference of d = .3, and an alpha of .05, power > 99%