Random Sampling in Python: random.sample, NumPy and pandas

Researchers searching for this usually have a list of participants, records or samples and need to draw a random subset reproducibly, then report how they did it. Python offers four standard routes: random.sample(), random.choices(), NumPy’s Generator.choice() and pandas’ DataFrame.sample(). The short answer: use DataFrame.sample() when your sampling frame is a table, Generator.choice() when it is an array or a list of IDs, and the standard-library random.sample() for a plain list. Details below follow each library’s official documentation (read 2026-11-08).

What random sampling in Python is used for in research

A simple random sample gives every unit in the sampling frame the same chance of selection. In practice researchers use it to pick participants from a registry, to select records for a manual audit or coding check, to build a validation subset, or to draw a pilot sample from a larger dataset. Two choices matter before any code runs: whether units may be selected more than once (sampling with replacement) and whether selection probabilities are equal or weighted. Python’s functions differ on exactly these points.

Install or access

The random module ships with Python itself (docs.python.org). NumPy and pandas are separate packages from their official projects; install them with pip install numpy pandas and check the installed versions with numpy.__version__ and pandas.__version__, which you will want to record for your methods section.

Choosing the right function

Function	Replacement	Weights	Seed / reproducibility	Input
random.sample(population, k, *, counts=None)	Without: returns k unique elements	None (the counts argument, added in 3.9, repeats elements instead)	random.seed(a) on the module	A sequence (since Python 3.11 a set is no longer converted automatically)
random.choices(population, weights=None, *, cum_weights=None, k=1)	With	weights or cum_weights, same length as the population	random.seed(a)	A sequence
Generator.choice(a, size=None, replace=True, p=None, axis=0, shuffle=True)	Either, set by replace	p, must sum to 1	Seed passed to np.random.default_rng()	Array-like, or an integer n meaning arange(n)
DataFrame.sample(n=None, frac=None, replace=False, weights=None, random_state=None, axis=None, ignore_index=False)	Either, set by replace	weights: a column name or an aligned Series	random_state (int, BitGenerator, RandomState or Generator)	A DataFrame or Series

Sources: Python random docs, NumPy Generator.choice docs and pandas DataFrame.sample docs (all read 2026-11-08). The default of replace differs: Generator.choice defaults to True, while random.sample and DataFrame.sample sample without replacement by default.

Step-by-step: draw a reproducible simple random sample

1. Sampling from a table with pandas

1. Load your sampling frame into a DataFrame, one row per eligible unit, with duplicates and ineligible units already removed. Sampling cannot correct a flawed frame.
2. Call sample with either a count or a fraction, and a fixed seed:

   sample_df = df.sample(n=100, random_state=42)

   or, for a proportion of the rows:

   sample_df = df.sample(frac=0.10, random_state=42)

   n and frac are mutually exclusive; if neither is given, n=1.

3. Set replace=True only if you want sampling with replacement. It is required when frac is greater than 1.
4. Optionally pass ignore_index=True to relabel the result 0 to n-1.

2. Sampling IDs or an array with NumPy

1. Create a generator with a recorded seed: rng = np.random.default_rng(12345).
2. Draw without replacement: chosen = rng.choice(ids, size=100, replace=False). For a 2-D array, axis=0 (the default) selects rows.

According to the NumPy documentation, rng.choice(5, 3, replace=False) is equivalent to rng.permutation(np.arange(5))[:3], and supplying p uses a slower sampler.

3. Sampling a plain list with the standard library

1. import random, then optionally random.seed(2026). An integer seed is used directly; with None, Python uses system time or operating-system randomness, so results will differ on each run.
2. subset = random.sample(ids, k=100) returns a list of 100 unique elements.
3. For sampling with replacement or unequal weights use random.choices(ids, weights=w, k=100).

4. Record what you did

For the methods section, report the sampling frame and its size, the function used, the sample size, whether sampling was with or without replacement, the seed, and the Python, NumPy and pandas versions. A seed alone does not reproduce a draw across libraries, so name the exact function. This supports reproducibility of the selection step.

Output and how to read it

Each function returns the selected elements: random.sample and random.choices return a list, Generator.choice an array, and DataFrame.sample a DataFrame or Series whose index labels are those of the original rows (unless ignore_index=True). In the pandas documentation example, df["num_legs"].sample(n=3, random_state=1) returns three index-labelled rows (fish, spider, falcon): the labels show which original records were drawn, so keep them as your record of selection. Row order in the output is random too; sort by index if you need the original order.

Weighted draws: pandas normalises weights that do not sum to 1, treats NaN weights as zero and does not allow infinite weights. A selected sample is one realisation; its estimates differ from the population value by sampling error, which a larger n reduces.

Troubleshooting

Symptom	Cause (per official docs)	Fix
ValueError from random.sample	k larger than the population	Reduce k, or sample with replacement using random.choices
Error passing a set to random.sample (Python 3.11+)	Population must be a sequence; sets are no longer converted	Convert first, for example sorted(my_set)
TypeError from random.choices	Both weights and cum_weights were given	Supply only one
ValueError from random.choices	All weights are zero (3.9+)	Give at least one non-zero weight
IndexError from random.choices	Empty population	Check the filter that built the list
ValueError from Generator.choice	replace=False with size larger than the population	Reduce size or set replace=True
Error from DataFrame.sample with frac above 1	replace must be True when frac is greater than 1	Set replace=True or use frac of 1 or less
Error combining n and frac	The two are mutually exclusive	Pass one
Weighted pandas draw rejected	With replace=False, n*max(weights)/sum(weights) above 1 is not allowed	Lower n or use replace=True
Different sample each run	No seed set (seed(None), no random_state)	Set and record a seed

Do not use the random module for security purposes; the Python documentation directs security uses to the secrets module. Sampling within groups (stratified draws) is available through pandas’ DataFrameGroupBy.sample, listed under “See also” in the DataFrame.sample documentation; see our stratified sampling guide for the design side.

Alternatives within Python

All four options below are part of Python or the NumPy and pandas packages. Pick by input type:

If your data is	Use	Why (from the docs)
A DataFrame	DataFrame.sample	Keeps index labels, accepts frac, takes weights from a column
A NumPy array or ID vector	Generator.choice	Seeded generator object; replace and p arguments
A short list, no extra packages	random.sample	Standard library; unique elements only
Need replacement or weights on a list	random.choices	Weights or cumulative weights, k draws with replacement

For deeper treatment of NumPy’s function see NumPy random choice; for fixed-interval selection see systematic sampling.

FAQ

How do I take a random sample without replacement in Python?

Use random.sample(population, k), rng.choice(a, size, replace=False) with a NumPy generator, or df.sample(n=...) in pandas, which does not replace by default.

How do I make a random sample reproducible?

Set a seed: random.seed(n) for the standard library, np.random.default_rng(n) for NumPy, or random_state=n in pandas. Record the seed and library versions.

What is the difference between random.sample and random.choices?

random.sample returns unique elements (without replacement); random.choices draws with replacement and accepts weights.

Can I sample a percentage of rows in pandas?

Yes: df.sample(frac=0.2) returns 20% of rows. Do not combine frac with n.

Is Python’s random module suitable for cryptography?

No. The documentation says to use secrets for security purposes.

Sources

• Python Software Foundation, random: Generate pseudo-random numbers, read 2026-11-08.
• pandas documentation, pandas.DataFrame.sample (pandas 3.1.1 docs), read 2026-11-08.
• NumPy documentation, numpy.random.Generator.choice, read 2026-11-08.