Randomized block design: controlling variance

Progettazione Sperimentale in Python

James Chapman

Curriculum Manager, DataCamp

Understanding blocking

Reduce variance by grouping similar units
Each block receives all treatments
Focus on treatment effects, controlling for block effects

An example of block randomization showing two grids of orange and white squares

Block design data example

athletes.head()

   Athlete_ID Initial_Fitness_Level  Muscle_Gain_kg
0         113              Beginner        3.225102
1          30              Advanced        3.976548
2         183          Intermediate        5.165449
3         200              Beginner        2.188297
4         194              Beginner        4.724162

Implementing randomized block design

Use .groupby() to shuffle within blocks

blocks = athletes.groupby('Initial_Fitness_Level').apply(

    lambda x: x.sample(frac=1)

)

blocks = blocks.reset_index(drop=True)
blocks

     Athlete_ID Initial_Fitness_Level  Muscle_Gain_kg
0           198              Advanced           5.742
1           146              Advanced           6.248
2           157              Advanced           6.049
..          ...                   ...             ...
198         164          Intermediate           6.134
199         178          Intermediate           6.591

Implemented randomized blocks

numpy.random.choice() for random treatment assignment within blocks

blocks['Treatment'] = np.random.choice(
    ['Cardio', 'Strength Training', 'Mixed'],
    size=len(blocks))

blocks.sample(n=5)

     Athlete_ID  Initial_Fitness_Level  Muscle_Gain_kg          Treatment
 87         194               Beginner           4.724             Cardio
 54           3               Advanced           3.731  Strength Training
177          80           Intermediate           6.758              Mixed
146         183           Intermediate           5.165  Strength Training          
 60         190               Advanced           3.763             Cardio

Visualizing treatment effects within blocks

import seaborn as sns
sns.boxplot(x='Initial_Fitness_Level', y='Muscle_Gain_kg', hue='Treatment', data=blocks)
plt.show()

Boxplot within blocks

ANOVA within blocks

Assume a significance level $\alpha$ of 0.05

from scipy.stats import f_oneway
blocks.groupby('Initial_Fitness_Level').apply(
  lambda x: f_oneway(x[x['Treatment'] == 'Cardio']['Muscle_Gain_kg'], 
                     x[x['Treatment'] == 'Mixed']['Muscle_Gain_kg'],
                     x[x['Treatment'] == 'Strength Training']['Muscle_Gain_kg'])
)

Block
Initial_Fitness_Level
Advanced        (0.7951054385317405, 0.4555687666120679)
Beginner        (0.1085790370950905, 0.8972754969684291)
Intermediate    (0.5678877824942661, 0.5698403547950377)
dtype: object

Visualizing effects across blocks

import seaborn as sns
sns.boxplot(x='Initial_Fitness_Level', y='Muscle_Gain_kg', data=blocks)
plt.show()

Boxplot across blocks

ANOVA between blocks

f_oneway(
  blocks[blocks['Initial_Fitness_Level'] == "Advanced"]['Muscle_Gain_kg'], 
  blocks[blocks['Initial_Fitness_Level'] == "Beginner"]['Muscle_Gain_kg'], 
  blocks[blocks['Initial_Fitness_Level'] == "Intermediate"]['Muscle_Gain_kg']
)

F_onewayResult(statistic=2.325058605244051, pvalue=0.10045536062209368)

Let's practice!

Progettazione Sperimentale in Python