STATISTICAL SAMPLING

METHODS
Sampling means selecting a subset of individuals or items from
a larger population to draw conclusions about the whole.
The methods are broadly classified into two main categories:
Probability Sampling and Non-Probability Sampling.
I. Probability Sampling Methods:
Every member of the population has a known, non-zero
chance of being selected for the sample. This involves
random selection, which minimizes bias and allows
researchers to generalize the findings to the entire
population with a measurable margin of error.

 Simple Random Sampling (SRS):

Most basic form of probability sampling. Every
individual and every possible sample of a given
size have an equal chance of being selected.
Small to medium, fairly homogeneous populations
with a clear list.
EXAMPLES AND USE-CASE:
Typical in introductory experiments, classroom
surveys, quality checks, and pilot data for
modelling.
Estimating the average income of residents in a
city, where each resident has an equal likelihood of
being chosen for the survey.
Election polls, unbiased surveys.








 Advantages:
Unbiased in theory; supports valid inference.
Easy concept; simple formulas for mean, variance.
 Disadvantages:
Needs a full list of the population.
Can be costly/slow for large or scattered
populations.

 Systematic Sampling:
Select every k -th unit after a random starting
point: choose a random start between 1 and k ,
then pick units at intervals of k .
EXAMPLES AND USE-CASE:
Assembly quality check, population entry surveys,
Environmental monitoring,

Advantages:
Very easy to implement (every k -th unit).
Good for production lines, queues, data streams.
Often cheaper and faster than SRS.
Disadvantages:
Risk of bias if there is a hidden pattern with
period k (e.g., machine cycles).
Needs ordered list/flow of units.

 Stratified Sampling:
 The population is divided into non-overlapping
subgroups (strata) based on relevant
characteristics, and then a random sample is
drawn from each stratum. This method ensures
representation of all subgroups, which can improve
the representativeness of the sample, particularly
when the variable of interest varies significantly
between strata.
EXAMPLES AND USE-CASE:
Large-scale surveys or studies where specific
demographic representation is critical for accurate
analysis
To survey university students, you treat each year
(1st, 2nd, 3rd, 4th) as a stratum and randomly
sample students proportionally from each year so
their share in the sample matches the population.

Advantages:
 Ensures representation of key subgroups (strata).
 Often more precise (lower variance) than SRS.
 Allows separate estimates for each stratum.
Disadvantages:
 Need info to define strata beforehand.
 More complex design and analysis.
 Wrong or poor stratification can remove benefits.

Cluster Sampling:
Divide the population into naturally occurring groups (clusters),
such as schools, neighborhoods, factories, or offices.
Select some clusters randomly; in “single-stage” cluster
sampling, survey everyone in selected clusters; in “two-stage”
or multistage, sample individuals within chosen clusters.
EXAMPLES AND USE-CASE:
Large, geographically spread populations where listing all
individuals is hard or expensive: public health surveys,
education research, national household surveys.
 Education: Randomly select 20 schools in a state, then
test all students in those schools (single-stage cluster).
 Online platform: Select cities as clusters, then randomly
sample users within each selected city to study
engagement
Advantages:
 Much cheaper/logistically easier for wide areas.
 No need list of all individuals, only clusters (e.g., schools).
 Good for large-scale field surveys.
Disadvantages:
 Individuals within a cluster are similar → higher variance
than SRS.
 Requires correct definition of clusters.
 Analysis more complex (must account for clustering).
Multistage sampling:
 Apply clustering and/or stratification in multiple levels:
country → state → city → neighborhood → households →
individuals.
 At each stage, take a random sample of units
EXAMPLES AND USE-CASE:
National and international surveys where direct SRS is
impossible due to logistics: census sample surveys,
demographic and health surveys, large market research
panels.
A social media company: randomly sample countries, then
states, then cities, then users within each city to measure
ad performance

Advantages:
 Very flexible for national surveys.
 Saves cost by sampling at several levels.
 Still probability-based → valid inference if modeled
correctly.
Disadvantages:
 Design and analysis can be complex.
 Sampling error accumulates over stages.
 Needs good frames at each stage.

Special variants:
 Reservoir sampling:
 Core concept: Maintain a fixed-size random sample
from a data stream of unknown length using a
streaming algorithm.
 Use-case: Log analysis, monitoring events, big data
pipelines where storing all items is impossible.

Advantages:
  Memory-efficient: Fixed upper bound on storage,
ideal for massive or streaming data where full
dataset can't fit in RAM.
 Single-pass processing: No need to know total size
upfront or rewind; processes data sequentially in
real-time.
 Unbiased randomness: Each element has equal
inclusion probability, suitable for representative
sampling in logs, sensors, or ML training subsets.
Disadvantages:
 Generates random numbers for every item (O(n)
time), slow for very large streams despite single
pass.
 Basic version assumes uniform stream; biased or
evolving streams (e.g., time-weighted) need
extensions that add complexity.
 Not ideal for small datasets or when multiple passes
are feasible; lacks stratification for subgroup
representation.
 Stratified systematic sampling:
 Combine stratification and systematic selection
within each stratum for operational efficiency and
subgroup representation.
Advantages:
 Combines stratification benefits (subgroup
representation, lower variance) with systematic ease
(faster/cheaper than pure random within strata).
 Ensures balance across key variables (e.g., age,
region) while avoiding full randomization overhead.
 More precise estimates than simple systematic; good
for operational data like quality control by shift or
batch.
Disadvantages:
  Requires prior strata knowledge and proportional
sizing; poor stratification wastes effort.
 Inherits systematic risks: Bias if hidden periodicity
aligns with k within a stratum (e.g., daily cycles in
production data).
 Complex analysis: Must weight strata and account for
design in variance calculations.
Probability Proportional to Size (PPS) Sampling:
 Larger clusters have higher chance of selection.
 Example: Towns with more people have higher
selection probability in migration study.
 Used in: Population survey, business employee
surveys

II. Non-probability sampling methods:

These methods are techniques where samples are
selected based on non-random criteria such as
convenience, judgment, or accessibility, rather than
giving every member of the population an equal chance
of selection. They are widely used in exploratory
research, case studies, and situations where probability
sampling is impractica
Convenience sampling:
 Choose units that are easiest to access; no random
mechanism and no full frame
 USE CASE AND EXEAMPLES:
Early exploratory work, classroom demonstrations, pilot
usability tests, quick feedback from nearby or online
audiences.
Asking friends in your lab or the first 100 visitors to a
website to fill out a quick survey; results are fast but not
representative
Advantages:
 Very quick and cheap.
 Good for pilots, UX tests, classroom demos.
 Easy when population frame is unknown.
Disadvantages:
 Strong selection bias; not representative.
 Results can’t be generalized reliably.
 Often over-represents easily reachable people.

Voluntary/volunteer sampling:
 Participants opt in by responding to open calls (e.g.,
online polls, “rate your experience” forms), often
leading to self-selection bias.
USE-CASE AND EXAMPLES:
  Feedback widgets, product reviews, online rating
systems, public opinion polls on news websites where
anyone can vote.
 A news website puts a poll “Do you support X
policy?” and visitors who care enough click and vote;
people with strong opinions are over-represented.
Advantages:
 Easy to collect large n online (polls, reviews).
 Good for satisfaction/feedback systems.
 Cheap, scalable.
Disadvantages:
 Self-selection bias: mainly people with strong opinions
respond.
 Not representative; poor external validity.
 Can distort estimates badly (e.g., political polls).
 Purposive/judgmental sampling:
 Researcher deliberately selects units that are particularly
informative, expert, or typical based on judgment and
study objectives.
 USE-CASE AND EXAMPLES:
 Qualitative research, expert panels, case studies,
rare or specialized populations (e.g., domain experts,
key decision-makers).
 Choosing 15 experienced data engineers and 10
product managers from different tech companies to
interview about MLOps practices.
Advantages:
 Efficient for rare/elite/very specific groups.
 Researcher can target “information-rich” cases.
 Useful in qualitative or case-study research.
Disadvantages:
 Strongly depends on researcher’s judgment → bias.
 No known selection probabilities; no formal inference.
 Hard to defend as representative.

Quota sampling:
  Set quotas for important subgroups (e.g., 50% male,
50% female; or specific age brackets), then select
convenient or purposive respondents until each
quota is filled, without random selection within
groups.
 USE-CASE AND EXAMPLES:
 Market research and opinion polls where time and
budget are limited but some demographic balance is
required.
 For a quick brand survey, a market research agency
wants 200 respondents: 100 men and 100 women;
interviewers recruit people on the street until each
gender quota is met.
Advantages:
 Ensures certain proportions (e.g., age, gender) quickly.
 Cheaper and faster than stratified random sampling.
 Common in market research street/online interviews.
Disadvantages:
 Within each quota, selection is non-random → bias.
 Interviewer convenience strongly affects sample.
 Cannot compute sampling error properly.

Snowball / chain sampling:

 Begin with a small set of initial participants; ask them
to refer others in their network; the sample
“snowballs” as new participants recruit more.
 USE-CASE AND EXAMPLES:
 Hard-to-reach or hidden populations (e.g., certain
patient groups, people engaged in illegal or
stigmatized activities), social network studies.
  To study an underground music subculture, a
researcher interviews initial artists, then asks each to
recommend two more artists, and so on
Advantages
 Can reach hidden or stigmatized populations.
 Leverages trust networks; good for sensitive topics.
 Useful for social network structure studies.
Disadvantages
 Strong network bias (friends of friends).
 Slow and hard to control composition.
 Very weak generalizability to the full population.

Other variants:
 Judgment/purposive subtypes: typical case sampling,
extreme/deviant case sampling, maximum variation
sampling, focused on particular patterns or extremes
(often in qualitative research).
 Dense or double sampling (in some classifications):
 Used to collect more detailed information from a
subset or to complement other sampling stages,
often combining probability and non-probability
phases.