L09 — Sampling Distributions & the CLT by Simulation

1Objectives¶

By the end of this lesson you will be able to:

1. Explain what a sampling distribution is and why it’s different from the distribution of the raw data.

2. Simulate a sampling distribution of the mean with the do(1000) * mean(~x, data = resample(D)) pattern.

3. Plot a simulated sampling distribution with gf_histogram() and read its shape, center, and spread.

4. Compare a simulated standard error to the theoretical value $\sigma / \sqrt{n}$.

5. Explain, using your own simulation, what the Central Limit Theorem (CLT) says and why it matters for inference.

2One sample, or many samples?¶

Every statistic you’ve computed so far — favstats()'s mean, tally()'s proportion — came from one sample of 120 students. But that one sample was never guaranteed to look exactly like a different sample of 120 students would have. A sampling distribution is the distribution you’d get if you could repeat “take a sample, compute a statistic” over and over — not the distribution of individual students, but the distribution of the statistic itself, across many hypothetical samples. You can’t actually resurvey CSUB a thousand times, but you can simulate it, and that simulation is this lesson’s whole point. Reload the toolkit and data if you’re starting fresh:

suppressMessages({library(mosaic); library(BSDA)})
survey <- read.csv("data/survey_sim.csv")
set.seed(2200)

3A skewed population, on purpose¶

To see the sampling-distribution idea do real work, we need a population that isn’t already bell-shaped — otherwise “the sampling distribution looks normal” wouldn’t be surprising. commute_min (minutes to campus) is exactly that: it’s generated from a right-skewed Gamma model in data/make_survey_sim.R (a few very long commutes stretch the distribution out to the right).

favstats(~ commute_min, data = survey)

 min Q1 median Q3 max     mean      sd   n missing
   7 15     22 30  79 26.21667 16.0043 120       0

The mean (26.2) sits noticeably above the median (22) — a classic right-skew signature, confirmed by the plot:

gf_histogram(~ commute_min, data = survey, binwidth = 5,
             fill = "#D55E00", color = "white") %>%
  gf_labs(title = "Population: commute minutes, all 120 surveyed students",
          x = "Commute time (minutes)", y = "Number of students")

This whole survey data frame — all 120 rows — is our stand-in population $D$ for this lesson. Its true mean is $\mu = 26.22$ minutes and its true standard deviation is $\sigma = 16.00$ minutes; both are known exactly because we have the entire population, not just a sample of it.

4resample(): one simulated sample from D¶

resample(D, size = n) draws n rows from data frame D with replacement — a mosaic function built for exactly this kind of simulation. One call gives you one simulated sample:

set.seed(2200)
one_resample <- resample(survey, size = 30)
dim(one_resample)
head(one_resample[, c("student_id", "commute_min")], 4)

[1] 30 10
   student_id commute_min
44       S044          28
65       S065           9
47       S047          23
24       S024          17

30 rows, same 10 columns as survey — and because it’s with replacement, the same student (row 44, 65, ...) can appear more than once across many calls, and some students won’t appear at all in a given call. That “with replacement” detail matters: it’s what lets resample() stand in for “draw a fresh sample from the population” even though it’s technically reusing the same 120 rows every time.

5The core pattern: do(1000) * mean(~x, data = resample(D))¶

One resample gives one simulated sample mean. Wrapping it in do(1000) * repeats the whole “resample, then compute the mean” process 1000 times, building an actual simulated sampling distribution:

set.seed(2200)
sim30 <- do(1000) * mean(~ commute_min, data = resample(survey, size = 30))
head(sim30, 4)

      mean
1 26.70000
2 32.23333
3 26.43333
4 23.06667

Read this the way you’d read any do() output (L08 used the same pattern for rflip()): each of the 1000 rows is one simulated “resample 30 students, compute their mean commute time” experiment. Row 1’s simulated sample happened to average 26.70 minutes; row 2’s, 32.23 minutes; and so on — 1000 different simulated sample means, one whole distribution’s worth. Summarize and plot that distribution exactly like any other numeric column, because that’s exactly what it is:

favstats(~ mean, data = sim30)

      min       Q1   median       Q3      max     mean       sd    n missing
 18.73333 24.19167 25.96667 28.03333 35.46667 26.14273 2.864828 1000       0

gf_histogram(~ mean, data = sim30, binwidth = 1,
             fill = "#0072B2", color = "white") %>%
  gf_labs(title = "Sampling distribution of the mean, n = 30 (1000 reps)",
          x = "Sample mean commute time (minutes)", y = "Number of simulated samples")

Compare this figure to the population histogram above: individual commute times are right-skewed and spread from 7 to 79 minutes, but sample means of 30 commute times are roughly symmetric and bell-shaped, clustered tightly between about 19 and 35 minutes. That gap between “what one observation looks like” and “what an average of many observations looks like” is the entire reason sampling distributions matter for inference — it’s coming back in L10 the moment we build a confidence interval.

6Simulated SE vs. the theoretical formula $\sigma / \sqrt{n}$¶

The standard error (SE) of the mean is the standard deviation of the sampling distribution itself — how much the sample mean typically bounces around from sample to sample. You just simulated that distribution, so you can read its SE straight off the simulation, with no formula at all:

sd(~ mean, data = sim30)

[1] 2.864828

Statistical theory gives a formula for the same quantity, using only the population SD $\sigma$ and the sample size $n$:

sd(~ commute_min, data = survey) / sqrt(30)

[1] 2.921972

2.86 (simulated) vs. 2.92 (formula) — close, not identical, because a simulation of 1000 reps is itself a sample of the true sampling distribution and carries its own small amount of noise. Run more reps (do(10000) * ...) and the two numbers converge even closer. This agreement is the real payoff of simulating first: the formula isn’t a mysterious rule to memorize, it’s a shortcut for something you can also just go and simulate directly.

7The Central Limit Theorem, illustrated by changing $n$¶

The Central Limit Theorem (CLT) says: as sample size $n$ grows, the sampling distribution of the mean gets closer to a normal (bell) shape and narrower around $\mu$ — no matter how skewed the original population is. commute_min’s skewed population is the perfect test case. Repeat the exact same simulation at three sample sizes:

set.seed(2200)
sim5   <- do(1000) * mean(~ commute_min, data = resample(survey, size = 5))
set.seed(2200)
sim100 <- do(1000) * mean(~ commute_min, data = resample(survey, size = 100))

sd(~ mean, data = sim5)
sd(~ commute_min, data = survey) / sqrt(5)

sd(~ mean, data = sim100)
sd(~ commute_min, data = survey) / sqrt(100)

[1] 7.362373
[1] 7.157339

[1] 1.578016
[1] 1.60043

Every row agrees closely, and the SE shrinks as $n$ grows — exactly what $\sigma/\sqrt{n}$ predicts, since a bigger $n$ sits in a bigger denominator. Faceting the three sampling distributions side by side shows the shape half of the CLT, not just the spread:

sim5$n   <- "n = 5"
sim30$n  <- "n = 30"
sim100$n <- "n = 100"
all_sims <- rbind(sim5[, c("mean", "n")], sim30[, c("mean", "n")], sim100[, c("mean", "n")])
all_sims$n <- factor(all_sims$n, levels = c("n = 5", "n = 30", "n = 100"))

gf_histogram(~ mean, data = all_sims, binwidth = 2,
             fill = "#009E73", color = "white") %>%
  gf_facet_wrap(~ n, scales = "free_y") %>%
  gf_labs(title = "Sampling distribution of the mean narrows and normalizes as n grows",
          x = "Sample mean commute time (minutes)", y = "Number of simulated samples")

The n = 5 panel still shows a hint of the population’s right skew (a longer tail on the right) and is noticeably wide. By n = 100, the skew has essentially vanished and the panel is tight and symmetric around the population mean. Nothing about commute_min itself changed between panels — only the sample size used to build each mean. That is the Central Limit Theorem, not as a memorized statement, but as something you just watched happen in three histograms, generated from data you know is skewed to begin with.

mosaic’s %>% gf_facet_wrap(~n) is the identical faceting tool from L07 — one more place a skill you already have keeps paying off.

8Summary¶

• A sampling distribution is the distribution of a statistic (like $\bar{x}$) across many hypothetical samples — not the distribution of the raw data.

• resample(D, size = n) draws one simulated sample of n rows from data frame D, with replacement; do(1000) * mean(~x, data = resample(D, size = n)) repeats that 1000 times, building a full simulated sampling distribution you can favstats() and gf_histogram() like any other numeric column.

• The simulated standard error (sd() of the sampling distribution) closely matches the theoretical formula $SE(\bar{x}) = \sigma/\sqrt{n}$ — the formula is a shortcut for what you can also simulate directly.

• The Central Limit Theorem: as $n$ grows, the sampling distribution of the mean narrows (SE shrinks, since $n$ is in the denominator) and its shape moves toward normal — even when, as with commute_min, the underlying population is skewed.

• The full script that generated every figure and number in this lesson is committed at data/make_L09_figures.R — run it yourself to reproduce every output exactly.