---
title: "American Time Use Survey (ATUS)"
subtitle: "Exercises"
format:
  html: 
    code-overflow: wrap
    toc: true
    toc-title: "Table of Contents"
include-in-header:
  - text: |
      <style>
      .cell-output-stdout code {
        word-break: break-wor !important;
        white-space: pre-wrap !important;
      }
      </style>
execute:
  warning: false
  message: false
---

```{r}
# as usual, we will use the tidyverse
library(tidyverse)

# responses concerning personal activity
atus_act = read_csv('http://euclid.nmu.edu/~joshthom/dat309/data/ATUS/atusact_2025.dat')

# responses concerning demographics
atus_rost = read_csv('http://euclid.nmu.edu/~joshthom/dat309/data/ATUS/atusrost_2025.dat')

# responses concerning types of jobs
atus_job = read_csv('http://euclid.nmu.edu/~joshthom/dat309/data/ATUS/atusresp_2025.dat')

# respondent info with labor force status & earnings
atus_resp = read_csv('http://euclid.nmu.edu/~joshthom/dat309/data/ATUS/atusresp_2025.dat')

atus_dict = read_csv('http://euclid.nmu.edu/~joshthom/dat309/data/ATUS/atus_dict.csv')
```


## Exercise: Weighted vs. Unweighted Estimates

Every ATUS respondent carries a survey weight, `TUFINLWGT`, that adjusts for the survey's sampling design so that estimates generalize to the U.S. civilian, noninstitutional population. If we ignore this weight and just count respondents directly, our estimates only describe the *sample* we happened to collect, not the population.

In this exercise we'll compare the *unweighted* proportion of respondents in a given job category to the *weighted* proportion (which is what you would actually report).

```{r weighted-vs-unweighted}

atus_job <- atus_job |> rename("job" = TRDTOCC1)

# atus_job already has `job` (renamed from TRDTOCC1) and TUFINLWGT
jobs_subset <- atus_job |>
  filter(job == 3 | job == 5 | job == 13) |>
  mutate(
    job = case_when(
      job == 3  ~ "Computer/Math",
      job == 5  ~ "Life/Physical/Social Science",
      job == 13 ~ "Food Prep/Serving"
    )
  )

# Count the number of respondents in each job category, ignoring weights
jobs_subset |>
  count(job)

# Count again, but as a proportion of the total sample (unweighted)
# Unweighted: each respondent counts equally
unweighted <- jobs_subset |>
  count(job) |>
  mutate(prop_unweighted = n / sum(n))

head(unweighted)

# Weighted: each respondent counts proportional to TUFINLWGT
weighted <- jobs_subset |>
  group_by(job) |>
  summarize(wgt_total = sum(TUFINLWGT)) |>
  mutate(prop_weighted = wgt_total / sum(wgt_total))

head(weighted)

# Combine for comparison
unweighted |>
  select(job, prop_unweighted) |>
  left_join(weighted |> select(job, prop_weighted), by = "job")
```

**Discussion questions**

1. How different are the weighted and unweighted proportions? Are the differences large enough to matter?
2. Why might some groups be over- or under-represented in the raw sample relative to their weighted share of the population? (Hint: think about who tends to respond to time-use surveys, and how ATUS's sampling design works.)
3. If you were writing a news article citing "X% of workers are in food service," which number — weighted or unweighted — should you report, and why?

**Try it yourself:** Repeat this comparison for a different set of job categories, or for a completely different variable (e.g., `TESEX` in `atus_rost`, or an activity code in `atus_act`). Does the gap between weighted and unweighted estimates get bigger or smaller?

## Exercise: Cross-File Join and Group Comparison

ATUS spreads information across several files that all share the `TUCASEID` identifier: `atus_rost` holds one row per *household member*, including demographics like sex (`TESEX`) and age (`TEAGE`), while `atus_resp` holds one row per *respondent*, including summary variables like `TRTALONE` (total minutes spent alone on the diary day). To ask a question like "does time spent alone differ by sex?" we need to bring these files together with joins.

`atus_rost` contains a row for every person in a respondent's household, not just the respondent, so we restrict to `TULINENO == 1` (the respondent) before joining, to avoid a many-to-many join.

```{r join-and-compare}
# Demographics: one row per respondent
respondent_demo <- atus_rost |>
  filter(TULINENO == 1) |>
  select(TUCASEID, TESEX, TEAGE) |>
  mutate(
    sex = case_when(
      TESEX == 1 ~ "Male",
      TESEX == 2 ~ "Female"
    )
  )

# Join demographics onto respondent-level summary data (TRTALONE lives here, one row per person)
resp_with_demo <- atus_resp |>
  select(TUCASEID, TRTALONE) |>
  left_join(respondent_demo, by = "TUCASEID")

# Compare average time spent alone (TRTALONE) by sex
resp_with_demo |>
  filter(!is.na(sex)) |>
  group_by(sex) |>
  summarize(
    n = n(),
    mean_minutes_alone = mean(TRTALONE, na.rm = TRUE),
    median_minutes_alone = median(TRTALONE, na.rm = TRUE)
  )

# Visualize the distribution
resp_with_demo |>
  filter(!is.na(sex)) |>
  ggplot(aes(x = sex, y = TRTALONE)) +
  geom_boxplot() +
  labs(
    x = "Sex",
    y = "Minutes spent alone (per diary day)",
    title = "Time spent alone by sex"
  )
```

**Discussion questions**

1. Why did we filter `atus_rost` to `TULINENO == 1` before joining? What would go wrong if we skipped that step?
2. `TRTALONE` lives in `atus_resp`, one row per respondent — not in `atus_act`, which has one row per *activity*. Why does it make more sense to store a daily total like `TRTALONE` at the respondent level rather than repeating it on every activity row?
3. Looking at the boxplot, does time spent alone appear to differ meaningfully by sex? What would you want to check before concluding this reflects a real difference (e.g., sample size, weighting, age composition)?

**Try it yourself:** Repeat this join-and-compare workflow using a different grouping variable from `atus_rost` (e.g., age bucketed into ranges) or a different outcome variable from `atus_resp` (e.g., `TRTFAMILY`, time spent with family). Does accounting for `TUFINLWGT` change your conclusions, as in the previous exercise?

## Exercise: Data Dictionary Scavenger Hunt

ATUS variable names are cryptic by design (`TEERNHRY`, `TRDTOCC1`, `TUFINLWGT`, ...), and the naming convention itself carries information: the second letter tells you the variable's *type* — `E` = edited response, `R` = recode calculated from other variables, `RT` = summary/time-total variable, `X` = allocation flag, `T` = topcode flag. Reading documentation carefully is a core data science skill, and this dataset — assembled from six separate files by many different hands over many years — rewards that carefullyj.

`TUFINLWGT` is the final annual weight for an ATUS respondent-day record. Multiplying by this weight and summing across respondents gives population-level estimates (e.g., "average minutes per day the U.S. population age 15+ spent on X").

`TUFINLWGT` is not a simple inverse-probability weight — it bakes in CPS's original weight, ATUS's day-based subsampling, nonresponse correction, and demographic post-stratification. For any population estimate, you should use it (e.g., weighted.mean(), or survey/srvyr package with the ATUS replicate weights if you need correct standard errors — ATUS also provides replicate weight variables for variance estimation, which plain weighted means won't reflect).


In this exercise, you'll pick unfamiliar variables, look them up in the data dictionary, classify them, and use what you find to build something.

```{r scavenger-hunt-setup}
# atus_dict has one row per variable, with its name and description
atus_dict |> slice_sample(n = 8)
```

**Part A: Look them up.** For each variable below, find its row in `atus_dict` (or search [atus_dictionary_simplified.qmd](../week3/atus_dictionary_simplified.qmd) directly) and write down, in your own words, what it measures:

- `TEAGE`
- `TELFS`
- `TRDTOCC1`
- `TUDIARYDAY`
- `TRTALONE`
- `TXAGE`

```{r lookup}
target_vars <- c("TEAGE", "TELFS", "TRDTOCC1", "TUDIARYDAY", "TRTALONE", "TXAGE")

atus_dict |>
  filter(name %in% target_vars)
```

**Part B: Classify by naming convention.** Using the convention described above (second letter = `E`, `R`, `X`, or `T`), classify each variable in `target_vars` as an edited response, a recode, an allocation flag, or something else. Check your answers against the description text — does it match what you'd expect from the letter code?

**Part C: Build something.** Choose *three* variables you were not already using in Exercises 1 or 2 — at least one of them unfamiliar to you before this exercise — and use them together to produce one exploratory visualization or summary table from `atus_act`, `atus_rost`, or `atus_job`. For example, you might explore how `TUDIARYDAY` (day of week) relates to `TRTALONE`, or how `TELFS` (labor force status) relates to earnings.

**Discussion questions**

1. Did any variable's actual meaning surprise you, or differ from what you guessed based on its name alone?
2. Pick one allocation flag (`X`-prefixed) variable related to a variable you used in Exercise 1 or 2. What does a nonzero flag value tell you about that observation, and would you want to exclude or flag those rows in a real analysis?
3. Why might a dataset like ATUS — built from government surveys, revised across years, and merged from six files — need this much documentation to use correctly, compared to, say, a dataset you might create yourself in a spreadsheet?