Lecture 9 – Data 100, Fall 2024¶

Data 100, Fall 2024

Acknowledgments Page

Updated by Joseph Gonzalez, Dominic Liu, Fernando Pérez.

In [1]:
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns

sns.set_theme(style='darkgrid', font_scale = 1.5,
              rc={'figure.figsize':(7,5)})

rng = np.random.default_rng()

Barbie v. Oppenheimer¶

To study how various sampling strategies work we will use a (fictional) census -- a complete survey of all Berkeley residents (our population). For the purposes of this fictional demo, assume:

  • wears_birkenstocks indicates if a resident identifies as male.
  • There are only two movies they can watch on July 21st: Barbie and Oppenheimer.
  • Every resident watches a movie (either Barbie or Oppenheimer) on July 21st.
In [2]:
census = pd.read_csv("movie_census.csv")
census['Barbie'] = census['movie'] == 'Barbie'
census
Out[2]:
age wears_birkenstocks movie Barbie
0 35 False Barbie True
1 42 True Oppenheimer False
2 55 False Barbie True
3 77 True Oppenheimer False
4 31 False Barbie True
... ... ... ... ...
1299995 62 True Barbie True
1299996 78 True Oppenheimer False
1299997 68 False Oppenheimer False
1299998 82 True Oppenheimer False
1299999 23 False Barbie True

1300000 rows × 4 columns

What fraction of Berkeley residents chose Barbie?

In [3]:
actual_barbie = census["Barbie"].mean()
actual_barbie
Out[3]:
np.float64(0.5302792307692308)

This is the actual outcome of the competition. Based on this result, Barbie would win. How did our sample of retirees do?

Convenience sample: Undergrads in Prof. Gonzalez OH¶

In [4]:
undergrads = census[(18 <= census['age']) & (census['age'] <= 22)].sample(10, replace=False)
undergrads["Barbie"].mean()
Out[4]:
np.float64(0.6)

Based on this result, we would have predicted that Oppenheimer would win! What happened?

  1. Is the sample too small / noisy?
In [5]:
len(undergrads)
Out[5]:
10
In [6]:
print("Percent of Berkeley:", len(undergrads)/len(census) * 100)

Percent of Berkeley: 0.0007692307692307692

Convenience sample: Elderly at a Campus Event¶

In [7]:
elderly = census[census['age'] >= 65].sample(100)
elderly["Barbie"].mean()
Out[7]:
np.float64(0.29)

Based on this result, we would have predicted that Oppenheimer would win! What happened?

  1. Is the sample too small / noisy?
In [8]:
len(elderly)
Out[8]:
100
In [9]:
print("Percent of Berkeley:", len(elderly)/len(census) * 100)

Percent of Berkeley: 0.007692307692307693

Check for bias¶

Let us aggregate all choices by age and visualize the fraction of Barbie views, split by footwear.

In [10]:
votes_by_barbie = (
    census
    .groupby(["age","wears_birkenstocks"])
        .agg("mean", numeric_only=True)
    .reset_index())

votes_by_barbie
Out[10]:
age wears_birkenstocks Barbie
0 18 False 0.819594
1 18 True 0.667001
2 19 False 0.812214
3 19 True 0.661252
4 20 False 0.805281
... ... ... ...
125 80 True 0.259731
126 81 False 0.394946
127 81 True 0.256759
128 82 False 0.398970
129 82 True 0.248060

130 rows × 3 columns

In [11]:
import plotly.express as px
px.scatter(votes_by_barbie, x = "age", y = "Barbie", 
           color = "wears_birkenstocks",
           title= "Preferences by Demographics")
  • We see that retirees (in Berkeley) tend to watch Oppenheimer.
  • We also see that residents who don't routinely wear Birkenstocks tend to prefer Barbie (nothing wrong with Birkenstocks).

Simple Random Sample¶

What if we instead took a simple random sample (SRS) to collect our sample?

Suppose we took an SRS of the same size as our undergrad sample:

In [12]:
## By default, replace = False
n = 2000
random_sample = census.sample(n, replace = False)

random_sample["Barbie"].mean()
Out[12]:
np.float64(0.538)

This is very close to the actual vote!

In [13]:
actual_barbie
Out[13]:
np.float64(0.5302792307692308)

It turns out that we can get similar results with a much smaller sample size, say, 800:

In [14]:
n = 800
random_sample = census.sample(n, replace = False)

# Compute the sample average and the resulting relative error
sample_barbie = random_sample["Barbie"].mean()
err = abs(sample_barbie-actual_barbie)/actual_barbie

# We can print output with Markdown formatting too...
from IPython.display import Markdown
Markdown(f"**Actual** = {actual_barbie:.4f}, **Sample** = {sample_barbie:.4f}, "
         f"**Err** = {100*err:.2f}%.")
Out[14]:

Actual = 0.5303, Sample = 0.5450, Err = 2.78%.

We'll learn how to choose this number when we (re)learn the Central Limit Theorem later in the semester.

Quantifying chance error¶

In our SRS of size 800, what would be our chance error?

Let's simulate 1000 versions of taking the 800-sized SRS from before:

In [15]:
nrep = 1000   # number of simulations
n = 800       # size of our sample
poll_result = []
for i in range(0, nrep):
    random_sample = census.sample(n, replace = False)
    poll_result.append(random_sample["Barbie"].mean())

Visualizing the distribution of outcomes:

In [16]:
fig = px.histogram(poll_result, histnorm='probability density', nbins=50)
fig.add_vline(x=actual_barbie, line_width=3, line_dash="dash", line_color="orange")
fig.update_layout(showlegend=False)

# Add Kernel Density Estimate curve
from scipy import stats
from plotly import graph_objects as go
x = np.linspace(min(poll_result), max(poll_result), 100)
fig.add_trace(go.Scatter(
    x=x, 
    y=stats.gaussian_kde(poll_result)(x), # Library for KDE (auto selects bandwidth)
    mode='lines', line=dict(color='red', width=3)) # Formatting
    )

Using seaborn instead:

In [17]:
sns.histplot(poll_result, stat='density', kde=True);
plt.axvline(actual_barbie, color='orange', linestyle='dashed', linewidth=2)
Out[17]:
<matplotlib.lines.Line2D at 0x16bde76d0>
No description has been provided for this image

What fraction of these simulated samples would have predicted Barbie?

In [18]:
poll_result = pd.Series(poll_result)
np.sum(poll_result > 0.5)/1000
Out[18]:
np.float64(0.944)



Return to Slides



Simulating from a Multinomial Distribution¶

Sometimes instead of having individual reports in the population, we have aggregate statistics. For example, we could have only learned that 53% of election voters voted Democrat. Even so, we can still simulate probability samples if we assume the population is large.

Specifically, we can use multinomial probabilities to simulate random samples with replacement.

Marbles¶

Suppose we have a very large bag of marbles with the following statistics:

  • 60% blue
  • 30% green
  • 10% red

We then draw 100 marbles from this bag at random with replacement.

In [19]:
np.random.multinomial(100, [0.60, 0.30, 0.10])
Out[19]:
array([61, 25, 14])

We can repeat this simulation multiple times, say 20:

In [20]:
np.random.multinomial(100, [0.60, 0.30, 0.10], size=20)
Out[20]:
array([[56, 35,  9],
       [64, 26, 10],
       [54, 40,  6],
       [61, 27, 12],
       [60, 27, 13],
       [59, 32,  9],
       [55, 34, 11],
       [63, 27, 10],
       [63, 26, 11],
       [55, 36,  9],
       [58, 31, 11],
       [55, 35, 10],
       [65, 26,  9],
       [56, 34, 10],
       [54, 32, 14],
       [59, 36,  5],
       [61, 28, 11],
       [55, 41,  4],
       [62, 30,  8],
       [56, 31, 13]])
In [ ]: