4-Hour Switches

This notebook shows the visualizations of different calendars that can be generated with switchback splitters

In [1]:

from cluster_experiments import PowerAnalysis
import pandas as pd
import numpy as np
import plotnine as pn
# Remove warnings
import warnings
warnings.filterwarnings('ignore')

np.random.seed(42)

pn.theme_set(pn.theme_minimal())

from cluster_experiments import PowerAnalysis import pandas as pd import numpy as np import plotnine as pn # Remove warnings import warnings warnings.filterwarnings('ignore') np.random.seed(42) pn.theme_set(pn.theme_minimal())

In [2]:

def plot_calendar(splitter, df, calendar_name):
    treatments = splitter.assign_treatment_df(df)
    
    treatment_df = (
        treatments
        .loc[:, ["city", "treatment", "time"]]
        .drop_duplicates()
    )
    # Convert time column to datetime format and extract week and day
    treatment_df['week_number'] = treatment_df['time'].dt.isocalendar().week.astype(int).astype(str)
    treatment_df['week_number_num'] = treatment_df['time'].dt.isocalendar().week
    # add time with the week set to the minimum week number
    treatment_df['time_in_week'] = treatment_df['time'] - pd.to_timedelta(treatment_df['week_number_num'], unit='W') + pd.to_timedelta(treatment_df['week_number_num'].min(), unit='W')

    # Custom date formatter function
    def custom_date_formatter(dates):
        return [date.strftime('%a %H:%M') for date in dates]

    # Create the plot
    plot = (
        pn.ggplot(treatment_df, pn.aes(x='time_in_week', y='week_number', fill='treatment'))
        + pn.geom_tile()
        + pn.theme_minimal()
        + pn.labs(
            title=f"Treatment assignment using {calendar_name}",
            x="Time since start of week",
            y="Week number",
            fill="Treatment"
        )
        + pn.facet_wrap('~city')
        + pn.scale_x_datetime(labels=custom_date_formatter)
        # vertical labels
        + pn.theme(axis_text_x=pn.element_text(angle=90, vjust=1.0))
    )

    # Display the plot
    return plot

def plot_calendar(splitter, df, calendar_name): treatments = splitter.assign_treatment_df(df) treatment_df = ( treatments .loc[:, ["city", "treatment", "time"]] .drop_duplicates() ) # Convert time column to datetime format and extract week and day treatment_df['week_number'] = treatment_df['time'].dt.isocalendar().week.astype(int).astype(str) treatment_df['week_number_num'] = treatment_df['time'].dt.isocalendar().week # add time with the week set to the minimum week number treatment_df['time_in_week'] = treatment_df['time'] - pd.to_timedelta(treatment_df['week_number_num'], unit='W') + pd.to_timedelta(treatment_df['week_number_num'].min(), unit='W') # Custom date formatter function def custom_date_formatter(dates): return [date.strftime('%a %H:%M') for date in dates] # Create the plot plot = ( pn.ggplot(treatment_df, pn.aes(x='time_in_week', y='week_number', fill='treatment')) + pn.geom_tile() + pn.theme_minimal() + pn.labs( title=f"Treatment assignment using {calendar_name}", x="Time since start of week", y="Week number", fill="Treatment" ) + pn.facet_wrap('~city') + pn.scale_x_datetime(labels=custom_date_formatter) # vertical labels + pn.theme(axis_text_x=pn.element_text(angle=90, vjust=1.0)) ) # Display the plot return plot

In [3]:

# Define data with random dates
df_raw = pd.DataFrame(
    {   
        "time": pd.date_range("2021-03-01", "2021-03-15", freq="1s")[
            np.random.randint(14 * 24 * 60 * 60, size=14 * 24 * 60 * 60)
        ],
        "y": np.random.randn(14 * 24 * 60 * 60),
    }
).assign(
    day_of_week=lambda df: df.time.dt.dayofweek,
    hour_of_day=lambda df: df.time.dt.hour,
    hour_4h=lambda df: df.time.dt.hour // 4,
    week=lambda df: df.time.dt.isocalendar().week,
)
df = pd.concat([df_raw.assign(city=city) for city in ("TGN", "NYC", "LON", "REU")])

# Define data with random dates df_raw = pd.DataFrame( { "time": pd.date_range("2021-03-01", "2021-03-15", freq="1s")[ np.random.randint(14 * 24 * 60 * 60, size=14 * 24 * 60 * 60) ], "y": np.random.randn(14 * 24 * 60 * 60), } ).assign( day_of_week=lambda df: df.time.dt.dayofweek, hour_of_day=lambda df: df.time.dt.hour, hour_4h=lambda df: df.time.dt.hour // 4, week=lambda df: df.time.dt.isocalendar().week, ) df = pd.concat([df_raw.assign(city=city) for city in ("TGN", "NYC", "LON", "REU")])

In [4]:

# We have data of 2-weeks in 4 cities
df.head()

# We have data of 2-weeks in 4 cities df.head()

Out[4]:

	time	y	day_of_week	hour_of_day	hour_4h	week	city
0	2021-03-02 09:52:38	0.328784	1	9	2	9	TGN
1	2021-03-08 18:25:55	0.489899	0	18	4	10	TGN
2	2021-03-02 12:38:52	-0.559446	1	12	3	9	TGN
3	2021-03-03 23:59:38	1.644260	2	23	5	9	TGN
4	2021-03-02 06:37:48	-0.418535	1	6	1	9	TGN

In [5]:

# Completely random calendar
from datetime import timedelta


config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback",
    "cluster_cols": ["time", "city"],
    "target_col": "y",
    "washover": "constant_washover",
    "washover_time_delta": timedelta(hours=2),
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "random calendar")

# Completely random calendar from datetime import timedelta config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback", "cluster_cols": ["time", "city"], "target_col": "y", "washover": "constant_washover", "washover_time_delta": timedelta(hours=2), } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "random calendar")

Out[5]:

<ggplot: (304288799)>

In [6]:

# Balanced calendar
config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback_balance",
    "cluster_cols": ["time", "city"],
    "target_col": "y",
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "balanced Calendar")

# Balanced calendar config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback_balance", "cluster_cols": ["time", "city"], "target_col": "y", } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "balanced Calendar")

Out[6]:

<ggplot: (304498911)>

In [7]:

# Stratified - city level
config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback_stratified",
    "cluster_cols": ["time", "city"],
    "strata_cols": ["city"],
    "target_col": "y",
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "stratified calendar (city level)")

# Stratified - city level config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback_stratified", "cluster_cols": ["time", "city"], "strata_cols": ["city"], "target_col": "y", } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "stratified calendar (city level)")

Out[7]:

<ggplot: (304646891)>

In [8]:

# Stratified - city + week level
config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback_stratified",
    "cluster_cols": ["time", "city"],
    "strata_cols": ["week", "city"],
    "target_col": "y",
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "stratified calendar (week + city level)")

# Stratified - city + week level config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback_stratified", "cluster_cols": ["time", "city"], "strata_cols": ["week", "city"], "target_col": "y", } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "stratified calendar (week + city level)")

Out[8]:

<ggplot: (304292286)>

In [9]:

# Stratified - city + day of week level
config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback_stratified",
    "cluster_cols": ["time", "city"],
    "strata_cols": ["day_of_week", "city"],
    "target_col": "y",
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "stratified calendar (day of week + city level)")

# Stratified - city + day of week level config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback_stratified", "cluster_cols": ["time", "city"], "strata_cols": ["day_of_week", "city"], "target_col": "y", } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "stratified calendar (day of week + city level)")

Out[9]:

<ggplot: (304915270)>

In [10]:

# Stratified - city + day of week + hour of day level
config = {
    "time_col": "time",
    "switch_frequency": "4h",
    "perturbator": "constant",
    "analysis": "ols_clustered",
    "splitter": "switchback_stratified",
    "cluster_cols": ["time", "city"],
    "strata_cols": ["day_of_week", "hour_4h", "city"],
    "target_col": "y",
}

power = PowerAnalysis.from_dict(config)

plot_calendar(power.splitter, df, "stratified calendar (day of week + hour of day + city level)")

# Stratified - city + day of week + hour of day level config = { "time_col": "time", "switch_frequency": "4h", "perturbator": "constant", "analysis": "ols_clustered", "splitter": "switchback_stratified", "cluster_cols": ["time", "city"], "strata_cols": ["day_of_week", "hour_4h", "city"], "target_col": "y", } power = PowerAnalysis.from_dict(config) plot_calendar(power.splitter, df, "stratified calendar (day of week + hour of day + city level)")

Out[10]:

<ggplot: (305061328)>