By the end

You should be able to:

1. state what one row represents;
2. inspect a table before changing it;
3. choose an operation that matches the biological question;
4. inspect a derived variable before comparing it;
5. distinguish a result, its interpretation and the evidence supporting it.

How this session works

LOOK → PREDICT → TRY IN RSTUDIO → DISCUSS → REVEAL

• The slides contain the code and precomputed output.
• Short pauses are marked Now in RStudio.
• Work with your neighbour; the TAs will circulate around the room.

The mouse experiment

sex · cage · batch provide experimental context for each mouse.

Did treated mice gain more weight than control mice?

Treatment varies within cages; treatment and cage are not the same grouping.

Before seeing the table

Predict:

• What should one row represent?
• Which columns are needed?
• Which column should identify a mouse?
• Where should the unit of weight be recorded?

Before seeing the table

Predict:

• What should one row represent?
• Which columns are needed?
• Which column should identify a mouse?
• Where should the unit of weight be recorded?

What does one row represent?

Choose one:

1. one weight measurement;
2. one mouse;
3. one cage;
4. one treatment group.

What does one row represent?

Choose one:

1. one weight measurement;
2. one mouse;
3. one cage;
4. one treatment group.

Rows, columns and cells

One value per cell; units belong in metadata or the column name.

Variables have different roles

The role follows from the experiment, not merely the R storage type.

Checkpoint: read the representation

For the cell in row 2 under final_weight_g:

• What does its row represent?
• What does its column represent?
• What does the value represent?
• Where does its unit come from?

Checkpoint: read the representation

For the cell in row 2 under final_weight_g:

• What does its row represent?
• What does its column represent?
• What does the value represent?
• Where does its unit come from?

Now in RStudio

Run once in the Console:

library(tidyverse)
data(mouse_trial, package = "BIOSCI504")

Keep mouse_trial unchanged. We will build temporary pipelines from it.

A preview is not an inspection

head(mouse_trial)

can show examples of values.

It cannot establish:

• the total number of records;
• whether every ID is unique;
• missingness elsewhere;
• complete group counts;
• the full range of a measurement.

Inspection starts with expectations

Inspection is a sequence of questions, not one command.

Now in RStudio: inspect structure

Predict what each command will tell you, then run it:

dim(mouse_trial)
names(mouse_trial)
glimpse(mouse_trial)

Check the observational unit

mouse_trial |>
  summarise(
    rows = n(),
    distinct_mice = n_distinct(mouse_id)
  )

What relationship do you expect between the two results?

The identifiers support the row claim

mouse_trial |>
  summarise(
    rows = n(),
    distinct_mice = n_distinct(mouse_id)
  )

  rows distinct_mice
1   48            48

Rows equal distinct mouse IDs in this table.

Now in RStudio: groups and missingness

Before running, predict the number of groups and what 0 would mean:

mouse_trial |>
  count(treatment)

mouse_trial |>
  summarise(
    across(
      c(baseline_weight_g, final_weight_g),
      ~ sum(is.na(.x))
    )
  )

Now in RStudio: groups and missingness

Before running, predict the number of groups and what 0 would mean:

mouse_trial |>
  count(treatment)

  treatment  n
1   control 26
2 treatment 22

mouse_trial |>
  summarise(
    across(
      c(baseline_weight_g, final_weight_g),
      ~ sum(is.na(.x))
    )
  )

  baseline_weight_g final_weight_g
1                 1              1

Now in RStudio: check ranges against units

Predict a plausible weight range in grams, then run:

mouse_trial |>
  summarise(
    across(
      ends_with("_weight_g"),
      ~ list(range(.x, na.rm = TRUE))
    )
  )

The suffix _g makes the expected scale testable.

Now in RStudio: check ranges against units

Predict a plausible weight range in grams, then run:

mouse_trial |>
  summarise(
    across(
      ends_with("_weight_g"),
      ~ list(range(.x, na.rm = TRUE))
    )
  )

  baseline_weight_g final_weight_g
1        22.3, 28.3     23.1, 34.1

Checkpoint: choose the revealing check

Which check is most direct?

Checkpoint: choose the revealing check

Which check is most direct?

The question determines the operation

These are different biological questions.

Same final weight, different change

Mouse A gained 5 g. Mouse B gained 1 g. Both finished at 25 g.

Valid code can answer the wrong question

Question: Did treated mice gain more weight?

mouse_trial |>
  group_by(treatment) |>
  summarise(
    n = sum(!is.na(final_weight_g)),
    mean_final_g = mean(final_weight_g, na.rm = TRUE)
  )

This code runs. Which question does it actually answer?

Valid code can answer the wrong question

Question: Did treated mice gain more weight?

mouse_trial |>
  group_by(treatment) |>
  summarise(
    n = sum(!is.na(final_weight_g)),
    mean_final_g = mean(final_weight_g, na.rm = TRUE)
  )

This code runs. Which question does it actually answer?

Name the quantity before coding

For each mouse:

\[ \text{weight change (g)} = \text{final weight (g)} - \text{baseline weight (g)} \]

Expected sign: positive for gain, negative for loss.

Now in RStudio: derive the variable

mouse_trial |>
  mutate(
    weight_change_g = ____________________
  )

What should remain unchanged after mutate()?

Inspect the new state

mouse_trial |>
  mutate(
    weight_change_g = final_weight_g - baseline_weight_g
  ) |>
  summarise(
    type = typeof(weight_change_g),
    missing = sum(is.na(weight_change_g)),
    minimum = min(weight_change_g, na.rm = TRUE),
    maximum = max(weight_change_g, na.rm = TRUE)
  )

    type missing minimum maximum
1 double       2     0.4     6.1

Missing inputs propagate

final weight     baseline weight     change
    25.3 g    −       NA          =    NA
      NA      −      24.2 g       =    NA

The mouse remains a row. The derived measurement is missing.

Checkpoint: predict the transformation

For one mouse:

baseline = 24.1 g
final    = 25.3 g

Before calculating:

• What sign should the change have?
• What is its value and unit?
• What does the row represent after adding the column?

Checkpoint: predict the transformation

For one mouse:

baseline = 24.1 g
final    = 25.3 g

Before calculating:

• What sign should the change have?
• What is its value and unit?
• What does the row represent after adding the column?

Plot individual observations first

A summary deliberately removes detail

Same mean; different evidence.

Predict the grouped summary

tibble(
  comparison_group = rep(c("A", "B"), each = 5),
  outcome = c(1, 2, 3, 4, 5, 2.8, 2.9, 3, 3.1, 3.2)
) |>
  group_by(comparison_group) |>
  summarise(
    n = n(),
    mean = mean(outcome),
    .groups = "drop"
  )

Before running: how many rows should the result contain?

Predict the grouped summary

tibble(
  comparison_group = rep(c("A", "B"), each = 5),
  outcome = c(1, 2, 3, 4, 5, 2.8, 2.9, 3, 3.1, 3.2)
) |>
  group_by(comparison_group) |>
  summarise(
    n = n(),
    mean = mean(outcome),
    .groups = "drop"
  )

# A tibble: 2 × 3
  comparison_group     n  mean
  <chr>            <int> <dbl>
1 A                    5     3
2 B                    5     3

One row now represents one group.

Uncertainty belongs with an estimate

None establishes that the design, data handling or code are correct.

Before choosing a test, record the assumptions

• What is the observational unit?
• Are observations independent enough for the proposed comparison?
• How were treatment and measurements assigned or collected?
• Are missing values plausibly ignorable?
• Do spread, shape or unusual values challenge the summary?
• Are all measurements expressed in the documented unit?

Visualization is verification evidence

A plot can reveal:

• spread hidden by a mean;
• an observation on the wrong scale;
• a group with few usable values;
• an unexpected relationship between baseline and final weight;
• a transformation whose sign or range is implausible.

The figure should be checked before it is interpreted.

Result is not interpretation

The design limits the claim

The table records:

• assigned treatment;
• paired weights;
• sex, cage and batch;
• missing measurements.

It does not automatically establish:

• the cause of every individual difference;
• independence within cages;
• why a measurement is missing;
• generality beyond the studied mice and conditions.

Why I trust this

Trust points to evidence. Remaining doubt should also be recorded.

AI can explain one step

Give it the local context:

Here is one inspection command and its output.

1. Explain what the command checks.
2. Name one concerning result and why it matters.
3. Do not repair the data or continue the analysis.

Then test the explanation against the table or a small code change.

Generated code can reduce typing. The question, context, output and assumptions still require scientific evaluation.

The exercise follows the same reasoning sequence

Situation
  → question and prediction
  → how the data arose
  → analysis specification
  → inspect
  → compare final weight
  → derive and compare change
  → verify
  → result
  → interpretation
  → why I trust this

Begin with the question

Open:

exercises/tabular-comparison/analysis.qmd

Start with:

1. question and prediction;
2. how the data arose;
3. analysis specification;
4. inspection.

Do not begin by requesting the complete analysis from AI.

Questions

Then move into the tabular-comparison exercise.