Probability Sampling vs. Non-Probability Sampling

The acquisition of empirical evidence lies at the epistemic core of modern science. Across sociology, epidemiology, political science, econometrics, and operational market research, investigators routinely confront an immutable reality: observing every single constituent element within a target population is almost universally impossible. Resource constraints, geographic dispersion, temporal imperatives, and the physical impossibility of exhaustive measurement dictate that empirical research must rely on a selected microcosm of reality. The discipline through which this microcosm is extracted—sampling methodology—determines whether empirical observations can legitimately substantiate universal claims or merely describe an isolated, non-generalizable cohort. At the very center of this methodological architecture lies the foundational dichotomy between probability sampling and non-probability sampling.

The choice between probability and non-probability designs represents far more than an operational decision concerning logistics and participant recruitment. It embodies a divergence in inferential paradigms, epistemological commitments, and mathematical assumptions regarding the nature of empirical uncertainty. Probability sampling traces its intellectual heritage through frequentist statistics, classical mechanics of randomization, and formal design-based inference, postulating that true population parameters can be estimated with mathematically quantifiable precision when every unit holds a known, non-zero probability of inclusion. In contrast, non-probability sampling arises from pragmatic field realities, interpretive qualitative traditions, and purposive discovery frameworks where statistical generalizability is consciously subordinated to theoretical saturation, operational feasibility, or targeted conceptual depth.

This comprehensive treatise explores the theoretical foundations, operational mechanics, mathematical frameworks, and real-world implications of both sampling traditions. By examining their historical development—from the pioneering insights of Anders Nicolai Kiær and Arthur Lyon Bowley to Jerzy Neyman’s watershed formulations and modern computational post-stratification methods—this analysis provides a rigorous, objective guide for investigators seeking to align their methodological choices with their analytical objectives, financial capacities, and philosophical assumptions.

1. Introduction to Sampling Methodologies

1.1 Defining the Core Concepts of Sampling in Research

Sampling is formally defined as the methodological process of selecting a predetermined subset of individual observations, units, or elements from within a larger, well-defined target population for the purpose of making empirical observations and deriving inferences about the aggregate body. The central aspiration of any sampling enterprise is to bridge the empirical chasm between the observed sample statistic—such as a sample mean, proportion, or regression coefficient—and the unobserved true population parameter. The validity of this bridge depends entirely upon how the subset is selected, structured, and observed.

In rigorous research design, a clear conceptual distinction must be maintained between three distinct layers of population definition: the target population, the sampling frame, and the actual sample units:

• The Target Population: The theoretically defined aggregate of elements to which the research questions, hypotheses, and conclusions are intended to apply. In an epidemiological study, this might be all adult individuals diagnosed with Type 2 diabetes within a sovereign nation; in organizational behavior, it might encompass all mid-level managers operating in multinational technology enterprises.
• The Sampling Frame: The concrete, operationalized list, directory, database, or map from which potential sample elements are directly drawn. The sampling frame serves as the practical proxy for the target population. Discrepancies between the theoretical target population and the operational sampling frame introduce frame error, encompassing both undercoverage (omission of eligible units) and overcoverage (erroneous inclusion of ineligible units or duplicates).
• The Sample Units: The specific entities selected at each stage of the sampling protocol. In single-stage designs, the sampling unit is identical to the observational element (e.g., an individual citizen). In multi-stage or clustered designs, primary sampling units (such as municipalities or school districts) are selected first, followed by secondary sampling units (such as residential blocks or specific schools), ultimately culminating in final observation units (such as individual residents or students).

The mathematical foundation of statistical inference rests upon probability theory. When sample selection is mediated through an objective, randomized protocol, the behavior of sample statistics can be mathematically modeled across hypothetical, repeated draws from the same parent distribution. Under this classical framework, sample statistics function as random variables characterized by an identifiable sampling distribution. This enables researchers to quantify sampling variance, construct confidence intervals, execute hypothesis tests, and invoke the Central Limit Theorem to describe parameter distributions even when parent populations deviate markedly from normality.

Conversely, subjective or purposive sample selection severs this direct connection between probability calculus and inferential certainty. When inclusion is dictated by human judgment, voluntary participation, geographic proximity, or demographic availability, the selection mechanisms remain either completely unknown or mathematically intractable. In these scenarios, sample selection is not governed by an objective stochastic mechanism, rendering the classical frequentist calculation of standard error, margin of error, and p-values conceptually problematic. Methodological rigor therefore demands an acute awareness of this operational trade-off: balancing analytical validity and mathematical defensibility against empirical feasibility, temporal limitations, and material resource constraints.

1.2 The Contextual Debate: Why These Two Paradigms Are Compared

The comparative discourse surrounding probability versus non-probability sampling is frequently obscured by methodological dogmatism. Novice researchers and quantitative purists often operate under the assumption that probability sampling is inherently superior across all investigative contexts. This perspective assumes that any study failing to execute a randomized probability draw suffers from fundamental methodological flaws. Conversely, researchers operating within qualitative, exploratory, or humanistic paradigms frequently dismiss probability designs as rigid, reductionist, prohibitively expensive, and ill-equipped to uncover the nuanced, contextualized realities of human experience. Both extreme stances represent profound misunderstandings of the relationship between research questions, epistemological paradigms, and sampling design.

A persistent conflation occurs regarding when formal statistical generalizability is genuinely required versus when conceptual, theoretical, or contextual transferability is sufficient. If the investigative objective is to determine the prevailing prevalence of a disease, forecast the outcome of a national democratic election, or assess macroeconomic household income distributions, statistical generalizability is non-negotiable. In such cases, failure to utilize probability sampling introduces uncontrolled selection bias that cannot be rectified by expanding sample size alone. However, if the primary objective is to formulate a novel grounded theory, uncover the psychological mechanisms of a rare psychiatric manifestation, or evaluate the usability of an experimental software interface, probability sampling is often neither technically viable nor methodologically necessary. In these exploratory domains, depth of insight, experiential richness, and theoretical saturation supersede the mathematical requirement for representative distributions.

Practical constraints continuously drive methodological selection. Large-scale probability surveys demand comprehensive sampling frames, extensive field infrastructure, sophisticated tracking mechanisms to mitigate non-response, and substantial financial backing. In real-world environments, researchers frequently face compressed timelines, budgetary limitations, institutional restrictions, and target populations that actively resist enumeration or documentation. A rigid insistence on probability sampling under conditions where no reliable sampling frame exists—such as studying unhoused urban populations, undocumented migrant laborers, or individuals engaged in illicit markets—would result in complete methodological paralysis. Consequently, comparing these two sampling paradigms is not an exercise in declaring a universal methodological victor; it is an analytical necessity to understand the exact trade-offs between inferential precision, operational viability, and theoretical utility.

2. Overview of Probability Sampling

2.1 Theoretical Framework and Core Mathematical Principles

Probability sampling is anchored in a foundational mathematical axiom: every single element within the defined sampling frame must possess a known, mathematically calculable, non-zero probability of inclusion into the final sample. Formally designated as the inclusion probability ($\pi_i$), this value must satisfy the condition $\pi_i > 0$ for all elements $i$ within the population of size $N$. When this condition is met, sample selection shifts from subjective human discretion to an objective, randomized physical or algorithmic mechanism, eliminating systematic selection bias and human distortion.

The theoretical framework governing probability sampling is known as design-based inference (or randomization-based inference). Under this paradigm, the population values of the variable of interest, denoted as $y_1, y_2, dots, y_N$, are treated as fixed, deterministic constants rather than random variables. The stochastic nature of the research enterprise arises exclusively from the randomization mechanism governing the sample selection indicator variables:

$$I_i = \begin{\cases} 1 &a\mp; \text{if element } i \text{ is selected into the sample} \ 0 &a\mp; \text{otherwise} \end{\cases}$$

Because the inclusion probability $\pi_i = P(I_i = 1)$ is explicitly known, researchers can derive completely unbiased estimators of population parameters without relying on parametric assumptions about the underlying distribution of the target population. A cornerstone of this framework is the foundational Horvitz–Thompson estimator, developed by D. G. Horvitz and D. J. Thompson in 1952. For a population total $Y = \sum_{i=1}^N y_i$, the Horvitz–Thompson estimator is mathematically formulated as:

$$\hat{Y}_{HT} = \sum_{i in S} \frac{y_i}{\pi_i} = \sum_{i=1}^N I_i \frac{y_i}{\pi_i}$$

Because the expected value of the selection indicator is $E(I_i) = \pi_i$, the mathematical expectation of the estimator yields an unbiased reflection of the population total:

$$E(\hat{Y}_{HT}) = \sum_{i=1}^N E(I_i) \frac{y_i}{\pi_i} = \sum_{i=1}^N \pi_i \frac{y_i}{\pi_i} = \sum_{i=1}^N y_i = Y$$

This proof demonstrates that as long as the inclusion probabilities are known and strictly greater than zero, any unequal chances of selection can be corrected by weighting each observation by the inverse of its selection probability ($w_i = 1 / \pi_i$). This theoretical baseline provides the foundation for calculating standard errors, testing statistical hypotheses, and generating valid confidence intervals.

The historical trajectory of this framework highlights the transition from intuitive balancing to strict mathematical formulation. In 1895, Anders Nicolai Kiær, Director of Statistics Norway, initiated the modern sampling debate at the International Statistical Institute (ISI) session in Bern by advocating for his “representative method” (Den repræsentative Undersøgelsesmethode). Kiær argued that an intentionally selected miniature model of a country, balanced across demographic and geographic variables, could yield reliable estimates without a full census. While initially met with skepticism, Arthur Lyon Bowley formalized this concept in 1906 and 1926 by integrating probability theory, systematic randomization, and the mathematical calculation of sampling errors into survey design.

The definitive paradigm shift occurred in 1934 with Jerzy Neyman’s landmark paper published in the Journal of the Royal Statistical Society: “On the Two Different Aspects of the Representative Method: The Method of Stratified Sampling and the Method of Purposive Selection.” Neyman examined the high-profile failure of an Italian agricultural census sample conducted by Corrado Gini and Luigi Galvani in 1926, which had utilized purposive selection to choose 29 districts matching aggregate census averages. Neyman demonstrated mathematically that purposive selection relied on dangerous, untestable assumptions regarding the correlation of variables across units. He proved that only randomized, probability-based selection could guarantee unbiased estimators, introduced confidence intervals to practical survey sampling, and developed the mathematical formulation for optimal allocation across strata (Neyman allocation). Through Neyman’s work, probability sampling was established as the international benchmark for scientific survey research.

2.2 Primary Modalities of Probability Sampling

Probability sampling is implemented across five core modalities, each tailored to distinct structural, logistical, and geographic environments:

Simple Random Sampling (SRS)

Simple Random Sampling represents the purest conceptual form of probability sampling. In an SRS protocol, every single unit in the sampling frame has an identical probability of inclusion ($\pi_i = n / N$), and every possible combination of $n$ distinct units has the exact same probability of being chosen. Operationally, SRS is executed either with replacement (SRSWR) or without replacement (SRSWOR), with the latter being standard in empirical research because re-sampling the same unit yields no new information. Historically implemented via physical lottery draws, modern execution relies on pseudorandom number generators (PRNGs) applied to computerized lists.

The primary advantage of Simple Random Sampling is its mathematical simplicity; sample statistics directly estimate population parameters without requiring complex weighting adjustments. However, SRS possesses major practical limitations: it requires an exhaustive, computerized enumeration of every unit in the population, is operationally inefficient when surveying geographically dispersed populations, and carries the risk of randomly underrepresenting small but critically important population subgroups due to simple stochastic variation.

Stratified Random Sampling

Stratified Random Sampling addresses subgroup representation and improves precision by dividing the heterogeneous target population into mutually exclusive and collectively exhaustive subgroups, known as strata. These strata are constructed such that variability within each stratum is minimized (homogeneity), while variability between distinct strata is maximized (heterogeneity). Common stratification variables include geographic territory, age cohorts, income brackets, educational attainment, or institutional size.

Once strata are established, independent simple random or systematic samples are drawn from within each individual stratum. Stratified sampling is broadly categorized into two structural approaches:

• Proportional Stratified Sampling: The sample size allocated to each stratum $h$ is strictly proportional to that stratum’s share of the total population ($n_h = n \cdot \frac{N_h}{N}$). This maintains a self-weighting sample structure where all units across the population retain the same overall inclusion probability.
• Disproportional (Optimal) Stratified Sampling: The sample allocation deviates from population proportions to maximize analytical efficiency or meet specific subgroup precision targets. Under Neyman Allocation, strata with greater size and higher internal variance receive larger sample allocations:
  $$n_h = n \frac{N_h S_h}{\sum_{k=1}^L N_k S_k}$$
  Where $S_h$ represents the standard deviation within stratum $h$. When data collection costs vary across strata, the generalized Tschuprow-Neyman Optimum Allocation explicitly incorporates cost per unit ($c_h$), allocating higher sample sizes to larger, more variable, and less expensive strata:
  $$n_h = n \frac{N_h S_h / \sqrt{c_h}}{\sum_{k=1}^L (N_k S_k / \sqrt{c_k})}$$

Cluster Sampling

Cluster Sampling is designed for operational and geographic efficiency when a complete list of individual elements is unavailable, but lists of naturally occurring groupings (clusters) exist or can be readily compiled. In this design, the target population is partitioned into multi-element aggregates—such as city blocks, schools, hospital wards, or electoral precincts. Unlike stratified sampling, which requires internal homogeneity within strata, cluster sampling ideally requires internal heterogeneity within clusters; each cluster should serve as an accurate microcosm of the aggregate population, while clusters themselves remain as homogeneous to one another as possible.

In single-stage cluster sampling, a random sample of clusters is selected, and every single element within the chosen clusters is observed. In two-stage cluster sampling, clusters are randomly selected first (primary sampling units, or PSUs), and then individual elements are randomly selected from within the selected PSUs (secondary sampling units, or SSUs). While cluster sampling significantly reduces field travel costs and simplifies frame construction, it usually increases standard errors relative to SRS. This design penalty is quantified by the Design Effect ($Deff$), driven by the Intraclass Correlation Coefficient ($rho$ or ICC):

$$Deff = 1 + (\bar{m} – 1)\rho$$

Where $\bar{m}$ is the average cluster size. Because elements within the same physical or social environment tend to share similar characteristics ($rho > 0$), cluster samples capture less independent information per observation than an equivalent number of simple random draws.

Systematic Sampling

Systematic Sampling offers an operationally streamlined alternative to simple random selection, particularly in industrial, clinical, and exit-polling settings. Given a sampling frame of size $N$ and a desired sample size $n$, the researcher calculates the sampling interval $k = lfloor N / n rfloor$. A random starting integer $r$ is generated such that $1 le r le k$. The sample is then formed by selecting elements corresponding to indices:

$$r, ; r + k, ; r + 2k, ; r + 3k, ; dots, ; r + (n-1)k$$

When the underlying sampling frame is randomly ordered, systematic sampling operates as a functional equivalent to simple random sampling. If the frame is sorted by a relevant auxiliary variable (such as income or age), systematic sampling functions similarly to proportionate stratified sampling, ensuring balanced spread across that attribute. However, systematic sampling carries one major vulnerability: periodicity. If the sampling frame contains a recurring cyclical pattern that coincides with the sampling interval $k$, the resulting sample will systematically overrepresent or entirely omit critical segments of the population, generating substantial bias.

Multi-Stage Sampling

Multi-Stage Sampling represents the standard design for modern national and international socio-economic, health, and demographic surveys (e.g., the Demographic and Health Surveys [DHS] or the U.S. Current Population Survey). Multi-stage sampling combines stratification, clustering, and simple or systematic randomization across a descending spatial or organizational hierarchy.

A typical national three-stage design illustrates this complexity: At Stage 1, primary sampling units (e.g., census enumeration areas or postal sectors) are stratified by geography and urbanization level, then selected using Probability Proportional to Size (PPS) sampling. In PPS, larger clusters have an inclusion probability proportional to their estimated population, ensuring balanced sample sizes in later stages. At Stage 2, secondary sampling units (e.g., residential blocks) are randomly selected within each chosen PSU. At Stage 3, a complete field listing of households is constructed, and a fixed number of households are selected via systematic sampling. Finally, within each household, an individual respondent may be selected using a randomized protocol (such as the Kish grid or next-birthday rule). This multi-tiered architecture balances fieldwork efficiency with mathematical rigor, though it requires complex weighting adjustments during analytical inference.

2.3 Advantages and Inherent Constraints of Probability Designs

The preeminent advantage of probability sampling is its mathematical foundation for external validity. Because selection probabilities are explicitly known, researchers can derive mathematically verifiable point estimates and quantify sampling variance without assuming that the sample resembles the population by sheer luck. Probability designs allow researchers to calculate clear margins of error, compute standard errors, and generate confidence intervals that accurately describe parameter boundaries at chosen significance levels ($\alpha$). This mathematical foundation underpins parametric hypothesis testing, structural equation modeling, and causal inference in quantitative observational research.

Furthermore, probability sampling provides a systematic defense against unconscious or intentional selection bias. Because the human researcher exercises no direct discretion over which specific units enter the sample, personal prejudices, socioeconomic blind spots, and subconscious convenience preferences are eliminated from unit selection. Through the application of post-stratification, non-response weighting, and calibration adjustments, known imbalances between the final sample and documented demographic benchmarks can be corrected with clear mathematical transparency.

Despite these strengths, probability designs introduce substantial operational constraints:

• Sampling Frame Dependency: The primary logistical barrier is the absolute requirement for an accurate, comprehensive, and up-to-date sampling frame. In many developing regions, conflict zones, or niche social groups, such directories simply do not exist. Constructing an exhaustive frame from scratch (e.g., through physical household listing operations) can consume a major share of an empirical research budget.
• Operational and Financial Cost: The logistics of tracking down randomly selected individuals across wide geographic expanses demand significant field personnel, travel, translation, and administrative oversight. When selected units reside in remote or hazardous locations, researchers cannot substitute them for nearby convenience units without compromising the randomization protocol.
• The Threat of Non-Response: A probability sample is theoretically unbiased only when the initial randomized selection matches the final observed sample. In contemporary research, declining response rates pose a severe threat. If non-respondents systematically differ from respondents on measured variables, non-response bias can distort the advantages of randomization, converting a nominal probability sample into an unrepresentative self-selected cohort.
• Temporal Inefficiency: Implementing a multi-stage probability survey—spanning frame acquisition, stratification, randomized draw, field verification, multi-wave callback protocols, and data auditing—can take months or years. For emergency policy responses, rapid product prototyping, or breaking social developments, the timeline required for rigorous probability sampling is often prohibitively slow.

3. Overview of Non-Probability Sampling

3.1 Conceptual Principles and Selection Logic

Non-probability sampling encompasses all methodological procedures in which individual observations are selected through non-random mechanisms. In a non-probability design, the individual inclusion probabilities ($\pi_i$) are unknown, mathematically undefinable, and frequently zero for broad segments of the target population. Selection is governed by researcher judgment, physical accessibility, subject availability, digital self-selection, or targeted conceptual criteria.

From an inferential perspective, non-probability sampling breaks from design-based inference. Because units lack known selection probabilities, the Horvitz–Thompson framework and classical frequentist design variance calculations cannot be strictly applied. Instead, to extrapolate sample findings to an external population, researchers must rely on model-based inference. Model-based inference requires strong, often untestable assumptions about the relationship between observed sample characteristics, unobserved factors, and the outcome variables of interest. Specifically, it generally assumes the selection mechanism is “ignorable” conditional on observed covariates—an assumption closely aligned with the Missing at Random (MAR) formulation ($Y perp R mid X$, where $Y$ is the outcome variable, $R$ is the sample inclusion indicator, and $X$ is a vector of observed conditioning covariates). If unobserved variables systematically influence sample selection, model-based corrections break down, introducing substantial bias into population estimates.

However, non-probability sampling is not merely a compromised alternative to random selection; it is often the preferred strategy in qualitative, phenomenological, constructivist, and exploratory paradigms. In these contexts, research questions do not seek to estimate population means or percentages ($\bar{Y}$ or $P$). Instead, they aim to uncover latent psychological constructs, map structural processes, generate new conceptual frameworks, or achieve theoretical saturation—the empirical point at which observing additional units yields no new conceptual insights or properties for an emerging theory, as articulated in the Grounded Theory tradition established by Barney Glaser and Anselm Strauss.

In qualitative and exploratory investigations, maximizing the absolute diversity or informational value of respondents is far more analytically valuable than achieving a representative demographic average. Non-probability sampling enables researchers to purposefully target extreme situations, critical incidents, and uniquely qualified key informants. The selection logic in these settings prioritizes information power—a framework emphasizing that the informational richness of a sample depends on the study’s aim, sample specificity, theoretical backing, dialogue quality, and analytical strategy, rather than raw sample size.

3.2 Primary Modalities of Non-Probability Sampling

Non-probability sampling includes several distinct operational designs, each suited to specific qualitative, pragmatic, or computational research contexts:

Convenience Sampling

Convenience sampling involves selecting units that are most accessible, available, and willing to participate. Common implementations include intercept surveys at commercial centers, student participant pools in university laboratory studies, and open-access internet surveys distributed via social media platforms. Because inclusion is driven entirely by accessibility and self-selection, convenience sampling features high operational speed and minimal administrative cost. It is widely used for instrument pre-testing, instructional demonstrations, exploratory pilot trials, and early proof-of-concept testing. However, convenience sampling carries severe methodological limitations: the sample is typically unrepresentative, and its inclusion mechanisms are deeply confounded by geographic, temporal, and self-selection biases.

Purposive (Judgmental) Sampling

Purposive sampling relies on the researcher’s deliberate expertise, diagnostic criteria, and analytical objectives to select units that provide information-rich data regarding the phenomenon of interest. Rather than drawing a random slice of a population, the investigator establishes clear inclusion criteria to target specific profiles. Purposive sampling is executed through several specialized variations:

• Maximum Variation Sampling: Purposefully selecting cases that display wide differences along specific dimensions (e.g., examining enterprise technology adoption across ultra-small startups, mid-tier firms, and global conglomerates). This reveals core commonalities that persist across diverse environments.
• Extreme or Deviant Case Sampling: Selecting unusual, highly atypical, or outlier cases—such as radically successful turnaround schools in high-poverty districts or individuals who completely resist infectious diseases—to identify unique causal mechanisms and operational limits.
• Critical Case Sampling: Selecting cases that permit logical deductions of the type: “If this phenomenon holds true here, it will hold true across almost any scenario.”
• Homogeneous Sampling: Selecting a focused, uniform cohort to study a specific subculture, demographic group, or occupational role in deep detail, reducing confounding variation to isolate core lived experiences.
• Expert Sampling: Systematically assembling panels of recognized domain authorities to establish consensus, calibrate institutional taxonomies, or execute iterative Delphi evaluations.

Snowball (Chain-Referral) Sampling

Snowball sampling is designed to study “hidden,” stigmatized, or hard-to-reach populations lacking formal sampling frames, directories, or institutional registries. Examples include active substance users, undocumented migrants, high-net-worth tax shelter participants, or individuals with rare medical conditions. The process begins with an initial cadre of accessible primary informants (seeds) who meet the study’s inclusion criteria. Following data collection, these seeds refer personal contacts within their network who share those criteria. These secondary contacts are then interviewed and asked for further referrals, generating an expanding, multi-tier chain-referral structure.

While traditional snowball sampling is vulnerable to masking (protecting personal network contacts) and homophily (seeds disproportionately referring individuals similar to themselves in age, social status, and beliefs), methodologists have formalized it into Respondent-Driven Sampling (RDS). Developed by Douglas Heckathorn, RDS applies mathematical network theory to chain-referral data. By tracking recruitment coupons, network sizes, and recruitment recruitment dynamics, RDS uses Markov chain modeling and network reciprocity weights to approximate unbiased population estimates, bridging non-probability field tracking with formal statistical adjustment.

Quota Sampling

Quota sampling is the non-probability counterpart to stratified random sampling, widely utilized in commercial market research, non-probability political polling, and rapid social policy assessments. The target population is categorized into discrete demographic or behavioral subgroups (e.g., age cohorts, gender identities, geographic regions, socio-economic classes). Numerical quotas are established for each subgroup, frequently matching their known census distributions.

Field interviewers or automated online panels then recruit respondents to fill these designated quotas. The fundamental distinction between quota sampling and stratified random sampling lies in the final selection mechanism: once the demographic quota is set, individual selection within that category is left entirely to non-random mechanisms (interviewer choice, immediate foot traffic, or digital volunteerism). While quota sampling ensures that the aggregate sample mirrors the broader population across visible demographic variables, it does not prevent substantial selection bias regarding unmeasured traits, such as psychological engagement, digital literacy, or availability.

Opt-In Digital and Crowdsourced Online Panels

The contemporary sampling landscape has been transformed by commercial opt-in online panels (e.g., Dynata, YouGov, Qualtrics Panels) and crowdsourcing platforms (e.g., Amazon Mechanical Turk, Prolific Academic, CloudResearch). These environments recruit millions of registered digital respondents through financial micro-incentives, gamification systems, and loyalty reward programs. While these panels offer fast turnaround and access to massive datasets, they represent pure non-probability samples. Participants are structurally self-selected, heavily skewed toward high digital literacy, spend disproportionate time online, and often function as professionalized survey takers. As a result, relying on unweighted crowdsourced samples introduces unique analytical challenges, including bot traffic, response automation, and behavioral conditioning.

3.3 Advantages and Inherent Constraints of Non-Probability Designs

The primary advantage of non-probability sampling is its high operational efficiency. By removing the requirement to construct exhaustive sampling frames and execute multi-wave callback protocols, these designs drastically reduce field costs and operational timelines. Large datasets can be collected in days or hours rather than months, making them well-suited for rapidly developing social crises, early-stage commercial research, agile product development, and resource-constrained academic studies.

Second, non-probability designs offer unique access to marginalized, hidden, or unlisted populations. When studying populations that actively avoid official detection or lack institutional documentation, probability sampling frames simply do not exist. Snowball techniques, community-engaged purposive sampling, and key-informant referral networks represent the only viable methodological paths for bringing these groups into empirical and policy research.

Third, non-probability designs are exceptionally effective for generating exploratory depth, identifying complex contextual interactions, and refining theoretical models. In qualitative methodologies—including grounded theory, case study design, and interpretive phenomenology—the objective is not to measure how frequently a phenomenon occurs across a broader population, but to map the internal logic, systemic interactions, and experiential structure of the phenomenon itself.

However, the limitations of non-probability sampling are substantial when applied to descriptive or inferential research questions:

• Absence of Verifiable Inferential Claims: Because selection probabilities are unknown and non-zero access cannot be verified, researchers cannot compute genuine frequentist standard errors, confidence intervals, or margins of error. Applying traditional parametric significance tests to raw non-probability samples violates the mathematical assumptions of those statistics, often generating false-positive conclusions and artificially narrow confidence bands.
• Structural Selection Bias: Inclusion is heavily confounded by participant availability, personal motivation, sociability, digital access, and interviewer discretion. Unlike probability sampling, where non-response error is actively tracked and evaluated, non-probability samples systematically obscure who is missing from the data. Expanding sample size does not resolve this issue; an unrepresentative non-probability sample of 500,000 observations will simply estimate an unrepresentative value with greater statistical precision.
• Model Dependency: Correcting bias in non-probability samples requires sophisticated statistical modeling—such as propensity score matching, raking ratio adjustments, or Multilevel Regression with Poststratification (MRP). These analytical corrections depend entirely on the assumption that researchers have measured all confounding variables that govern sample inclusion, an assumption that is rarely verifiable in field conditions.

4. Key Differences Between Probability and Non-Probability Sampling

4.1 Mechanisms of Selection and Frame Dependency

The operational mechanics that differentiate probability from non-probability sampling center on two core factors: the construction of the sampling frame and the operational mechanism that assigns units to the final sample. In probability sampling, an exhaustive frame listing all potential target units—or well-defined clusters of units—is an absolute prerequisite. This frame serves as the direct link between the physical population and the mathematical probability model. The selection itself is executed through an external, automated randomization process, such as a secure pseudorandom algorithm or a documented random draw. Once an element is selected through this mechanism, it cannot be substituted with an adjacent unit simply because it is easier or cheaper to reach. Every unit’s probability of selection ($\pi_i$) is mathematically defined, strictly positive, and recorded prior to field execution.

In non-probability sampling, research can proceed without a structured, exhaustive census frame. Selection relies on researcher discretion, organizational partnerships, demographic criteria, or participant availability. The probability that any given individual enters the final sample is completely unknown and mathematically untraceable; for large segments of the broader target population, that probability is effectively zero. Because selection is mediated by convenience, self-selection, or subjective judgment, the operational process cannot generate the independent, identically distributed properties required for design-based statistical inference.

Methodological Dimension	Probability Sampling	Non-Probability Sampling
Theoretical Definition	Every unit within the target population has a known, non-zero probability of inclusion ($\pi_i > 0$).	Unit selection probability is unknown, mathematically undefinable, and zero for large segments of the population.
Inclusion Mechanism	Objective randomization protocols (e.g., pseudorandom algorithms, lottery systems).	Subjective judgment, convenience, voluntary participation, targeted characteristics, or quota filling.
Sampling Frame Requirement	Strictly dependent on an exhaustive, enumerated, and verifiable sampling frame.	Can proceed without a pre-existing sampling frame; recruitment occurs organically or purposively.
Inferential Paradigm	Design-based (randomization-based) inference supported by classical probability calculus.	Model-based inference requiring strong, often untestable missing-data assumptions (e.g., MAR/selection ignorability).
Quantification of Error	Sampling error, standard errors, and confidence intervals can be directly calculated.	Sampling error cannot be mathematically isolated from structural selection bias; classical SEs are technically invalid.
Selection Bias Vulnerability	Systematic selection bias is minimized; primary operational risk stems from non-response attrition.	Highly vulnerable to conscious/subconscious researcher bias, accessibility artifacts, and participant self-selection.
Statistical Generalizability	Mathematically justified across the defined target population within known confidence intervals.	Cannot be mathematically generalized to a broader population without structural model assumptions; theoretical transferability.
Operational Cost and Speed	High financial capital, heavy operational oversight, complex field tracking, and prolonged execution timelines.	Rapid turnaround, lower capital expenditure, streamlined administrative overhead, and agile field deployment.
Typical Research Tradition	Quantitative survey research, epidemiology, official state statistics, econometrics, macro-demographics.	Qualitative inquiries, grounded theory, phenomenological exploration, rapid prototyping, early pilot validation.
Core Methodological Variants	Simple Random Sampling, Stratified Sampling, Cluster Sampling, Systematic Sampling, Multi-Stage Sampling.	Convenience Sampling, Purposive (Judgmental) Sampling, Snowball/RDS, Quota Sampling, Opt-in Digital Panels.
Handling Unavailable Units	Requires persistent callback protocols and non-response tracking; arbitrary unit replacement introduces severe bias.	Readily permits substitution of unavailable units with accessible alternatives matching the required criteria.
Sample Size Determination	Governed by formal power analysis, target margins of error, alpha levels, and expected population variance.	Governed by conceptual saturation, informational richness, research objectives, and resource availability.

4.2 Generalizability, Inferential Power, and Analytical Error

The divide between probability and non-probability sampling becomes most critical when evaluating generalizability and inferential power. In probability designs, statistical generalizability is mathematically justified. Because estimators derive their properties directly from the randomized selection mechanism, sample statistics act as unbiased estimators of population parameters across repeated implementations. Sampling variance naturally decreases as sample size ($n$) expands, systematically reducing the standard error roughly in proportion to $1/\sqrt{n}$. Analytical error within probability sampling can be broken down using the classical Mean Squared Error (MSE) framework:

$$MSE(\hat{\theta}) = \text{Var}(\hat{\theta}) + [\text{Bias}(\hat{\theta})]^2$$

Under an ideal, fully realized probability design with complete response, $\text{Bias}(\hat{\theta}) = 0$, meaning the total analytical error consists solely of random sampling variance ($\text{Var}(\hat{\theta})$). This variance can be calculated, monitored, and capped using sample size determination formulas. The operational challenge in probability designs stems primarily from non-sampling errors, such as measurement error, frame coverage deficiencies, and non-response bias. If individuals who refuse to participate differ systematically on the research variables from those who do participate, bias enters the equation, weakening the mathematical advantages of the random design.

In non-probability designs, the analytical error profile is dominated by selection bias. Because inclusion is driven by availability, social networks, or self-selection, $[\text{Bias}(\hat{\theta})]^2$ is typically non-zero, unknown, and largely independent of sample size. This dynamic was demonstrated during the 1936 U.S. Presidential Election poll conducted by the Literary Digest. The magazine gathered over 2.3 million returned mock ballots—an extraordinarily large sample size. Despite this massive volume of data, the poll incorrectly predicted that Republican candidate Alf Landon would defeat the incumbent Franklin D. Roosevelt in a landslide (predicting 57% for Landon versus 43% for Roosevelt). The actual election outcome was a historic landslide victory for Roosevelt, who secured 60.8% of the popular vote against Landon’s 36.5%—a forecasting error exceeding 24 percentage points.

The root cause of this historic failure was severe selection bias. The Literary Digest constructed its sampling frame from telephone directories, automobile registration lists, and club membership registries. In 1936, during the depths of the Great Depression, these assets were disproportionately owned by wealthier citizens who held distinct political attitudes and opposed the New Deal. This sampling bias was further compounded by self-selection bias: individuals dissatisfied with the incumbent administration were far more motivated to return their mail-in ballots. The Literary Digest disaster demonstrated an enduring statistical principle: when sample selection is non-random, expanding the sample size does not reduce bias. It simply estimates an unrepresentative parameter with greater precision.

When researchers attempt to draw population inferences from non-probability data, they must shift from design-based inference to model-based inference. This requires asserting that the model correctly controls for all factors that distinguish the selected sample from the target population. If unobserved variables influence both sample participation and the research outcome, the model will produce confounded estimates. In short, probability sampling provides a mathematical guarantee of design unbiasedness, whereas non-probability sampling shifts the burden of proof entirely onto the theoretical validity of statistical assumptions.

4.3 Resource Allocation, Feasibility, and Operational Complexity

Executing a probability survey requires substantial organizational infrastructure, extended timelines, and significant financial capital. The initial phase—developing an exhaustive, verified sampling frame—often demands purchasing specialized institutional datasets, accessing proprietary government registries, or conducting physical field enumerations. In multi-stage national surveys, researchers must navigate regional logistics, deploy trained field interviewers to remote locations, translate instruments into local dialects, and establish rigorous field-auditing systems to ensure interviewers do not deviate from randomized household selection protocols.

Operational complexity is further compounded by callback and tracking protocols. To prevent non-response bias, modern probability designs require interviewers to make multiple tracking attempts (often 3 to 10 attempts at different times of day and across weekends) before classifying a selected unit as unreachable. Researchers must deploy response-maximization strategies, including introductory letters, monetary incentives, and specialized conversion protocols for initial refusals. Every administrative step increases data collection costs. Consequently, the cost per completed interview in a rigorous probability design typically ranges from dozens to hundreds of dollars for in-person or high-quality telephone surveys.

In contrast, non-probability sampling offers an agile, cost-effective operational profile. If an intercept survey respondent declines to participate, the field interviewer immediately approaches the next available individual. In online crowdsourced environments, recruitment relies on automated algorithms matching demographic quotas across registered user pools, allowing thousands of surveys to be completed in a few days at a fraction of the cost per respondent. For rapid exploratory development, early product usability assessments, or urgent public health evaluations where directional feedback is needed immediately, non-probability designs provide accessible empirical insights that would be unattainable under the resource requirements of probability sampling.

5. Methodological Similarities and Complementary Overlap

5.1 Shared Objectives in Representing Latent Phenomena

Despite their divergent theoretical mechanisms and mathematical foundations, probability and non-probability sampling share a fundamental empirical objective: observing an accessible subset of units to extract meaningful insights about a broader domain of reality. Neither methodology values data collection as a purely isolated exercise; both seek to understand broader patterns, processes, behaviors, or conceptual relationships that extend beyond the specific individuals sitting across from an interviewer or completing an online form.

Both methodological frameworks demand rigorous operational discipline regarding target population definition and eligibility criteria. Whether executing a multi-stage stratified random household survey or an exploratory purposive qualitative interview study, researchers must clearly establish:

• The exact structural boundaries of the target population (inclusion and exclusion criteria).
• The core operational constructs and primary dependent/independent variables under investigation.
• The temporal and spatial boundaries within which the empirical observations are situated.
• Systematic data collection protocols to minimize internal measurement error, questionnaire ambiguity, and interviewer bias.

A poorly constructed survey instrument, confusingly worded question, or biased interviewer will undermine the validity of a study regardless of whether the sample was selected through simple random sampling or convenience intercept. The total quality of an empirical finding depends on the entire data generation process—including construct operationalization, instrument calibration, cognitive testing, and fieldwork ethics—not solely on the mathematical mechanism of sample selection.

5.2 Integration in Mixed-Methods and Multi-Phase Research

Rather than functioning as competing, mutually exclusive methodologies, probability and non-probability sampling are increasingly integrated within multi-phase, mixed-methods research frameworks. Combining both paradigms enables researchers to capitalize on the contextual depth of non-probability selection alongside the statistical generalizability of probability designs.

One common implementation is the Exploratory Sequential Design (Qualitative $\rightarrow$ Quantitative). In the initial exploratory phase, the researcher employs purposive, maximum variation, or theoretical non-probability sampling to conduct in-depth interviews or ethnographic field observations with a focused group of key informants. This non-probability phase uncovers latent dynamics, maps contextual variables, and identifies culturally specific terminology. The qualitative findings are then used to operationalize theoretical constructs, generate formal hypotheses, and construct structured survey questionnaires. In the second phase, this calibrated instrument is administered to a broad probability sample, allowing researchers to measure the prevalence of the discovered phenomena, test hypotheses, and generalize findings across the broader population.

Conversely, in an Explanatory Sequential Design (Quantitative $\rightarrow$ Qualitative), the researcher begins by deploying a wide-scale probability survey to capture aggregate population parameters, identify demographic trends, and locate statistical anomalies or interaction effects. While this probability phase establishes broad generalizability, it often lacks the contextual depth needed to explain why specific trends or outliers occur. To address this, the researcher uses purposive non-probability sampling to recruit participants directly from the initial survey cohort who exhibit specific response profiles (e.g., individuals who succeeded in an intervention despite facing high-risk factors). Detailed qualitative interviews with this non-probability subset then provide explanatory context for the broader quantitative patterns.

Finally, non-probability sampling plays an essential role in instrument calibration and pilot pre-testing before committing to the heavy resource investments required for full-scale probability surveys. Prior to deploying an extensive national survey, researchers routinely field-test questionnaire draft modules using convenience or quota samples across diverse demographic groups. Cognitive interviewing, item-response distribution evaluations, factor structure verifications, and user-experience assessments are conducted on non-probability cohorts to identify confusing phrasing, assess question sensitivity, and verify psychometric properties. Once the survey instrument has been refined and validated through this cost-effective pilot testing, it can be deployed within a rigorous probability sampling architecture.

6. Selection Framework: When to Use Each Sampling Approach

6.1 Criteria Mandating Probability Sampling

Selecting an appropriate sampling strategy requires evaluating the core research questions, epistemological assumptions, practical constraints, and intended use of the empirical findings. The following flowchart describes the primary decision path when designing an empirical investigation:

Specifically, empirical studies must utilize probability sampling designs when the research context meets the following criteria:

• Formulation of Public Policy and National Resource Allocation: Macroeconomic indicators, poverty indices, national labor statistics, and official demographic censuses must be grounded in probability sampling. Policymakers allocating public funds require statistically unbiased estimates that can withstand external auditing and public review.
• Epidemiological Surveillance and Clinical Population Baselines: Tracking infectious disease prevalence, national vaccination rates, and population-wide health risk factors requires probability designs. Relying on convenience samples in health research risks skewing disease indicators, potentially leading to misallocated medical resources.
• Regulatory Submissions and Confirmatory Clinical Trials: Phase III clinical trials and post-market safety evaluations mandated by regulatory agencies (such as the FDA or EMA) require rigorous randomization protocols to eliminate selection confounding and evaluate treatment efficacy against clear statistical criteria.
• Rigorous Parametric Inference and Formal Hypothesis Testing: When research seeks to establish generalizable empirical patterns, verify theoretical predictions through structural equation modeling, or evaluate causal claims in observational settings, probability sampling provides the necessary mathematical justification for frequentist inference.
• Preservation of Longitudinal Comparability: Long-term trend studies that measure socio-demographic changes across decades require stable, documented probability methodologies to ensure that observed shifts reflect genuine social change rather than shifting recruitment dynamics.

6.2 Criteria Mandating Non-Probability Sampling

Conversely, non-probability sampling is methodologically indicated, and often analytically superior, when an empirical investigation aligns with the following conditions:

• Early-Stage Exploratory Studies and Concept Discovery: When investigating a newly emerging phenomenon—such as early social responses to a novel technology or an unstudied cultural trend—the primary objective is to map unknown variables rather than measure their broader distribution. Non-probability purposive selection allows researchers to quickly capture diverse, information-rich insights.
• Qualitative Traditions Seeking Theoretical Saturation: Studies operating within grounded theory, phenomenology, or clinical narrative analysis prioritize conceptual depth over statistical generalization. In these settings, researchers intentionally seek out atypical or contrasting cases to refine theoretical models, a process best served by purposive sampling.
• Studying Hidden, Marginalized, or Hard-to-Reach Cohorts: When target populations lack institutional records, participate in stigmatized behaviors, or actively avoid state authorities (e.g., human trafficking victims, undocumented workers, illicit substance users), probability frames are unavailable. In these environments, snowball sampling, targeted purposive intercept, and respondent-driven sampling provide the only viable entry points for empirical research.
• Operational Pilot Testing and Rapid Instrument Calibration: Evaluating the usability, clarity, and psychometric stability of draft surveys or experimental tasks is efficiently conducted on convenience or quota samples. Investing substantial resources in probability sampling before an instrument has been tested and refined is methodologically inefficient.
• Severe Budgetary, Temporal, or Environmental Crises: During natural disasters, complex humanitarian emergencies, or fast-moving political crises, immediate field assessments are required to guide real-time responses. When operational conditions make probability sampling impossible, well-designed non-probability designs provide actionable data within compressed timelines.

7. Frequently Asked Questions

7.1 Can Non-Probability Sampling Ever Yield Statistically Generalizable Results?

From a strict, classical design-based perspective, non-probability samples cannot yield statistically generalizable results. Because units lack known, non-zero inclusion probabilities, the mathematical link between sample statistics and population parameters cannot be established through probability theory alone. Classical margins of error, standard errors, and p-values cannot be computed without making unverified assumptions about the selection mechanism.

However, through the lens of modern model-based inference, non-probability samples can approximate broader population parameters if specific structural conditions and statistical adjustments are met. The most prominent framework for this adjustment is Multilevel Regression with Poststratification (MRP), alongside propensity score weighting, raking, and covariate calibration. In MRP, researchers model the individual survey response as a function of multiple demographic and geographic covariates using a hierarchical Bayesian framework:

$$P(Y_i = 1) = \text{logit}^{-1}(\beta_0 + \alpha_{\text{age}[i]} + \alpha_{\text{education}[i]} + \alpha_{\text{region}[i]} + dots)$$

Once the model estimates response probabilities across demographic slices, these predictions are weighted by the known census frequencies of those slices (poststratification). This technique has enabled non-probability online panels to produce accurate predictions in political polling and health tracking. However, the validity of this model-based approach relies entirely on the assumption that the selection mechanism is ignorable (Missing at Random, or MAR) conditional on the observed covariates. If the decision to participate is influenced by unobserved factors that also correlate with the research outcome, the model will produce biased estimates. Model-based adjustments can reduce demographic imbalances, but they cannot provide the mathematical guarantee of unbiasedness inherent to a well-executed probability design.

7.2 Why Is Quota Sampling Often Confused with Stratified Random Sampling?

Quota sampling and stratified random sampling are frequently conflated because they share an identical preliminary operational step: both partition the target population into distinct, mutually exclusive demographic subgroups (strata) based on known characteristics, such as age, gender, geographic region, or educational level. In both designs, the researcher specifies target sample counts for each subgroup to ensure that key demographics are represented in the final dataset.

The critical difference lies in how individual units are selected within those defined subgroups:

• Stratified Random Sampling: Requires an exhaustive sampling frame for every stratum. Individual units within each stratum are selected strictly through an objective, randomized protocol (such as an automated pseudorandom draw). The field interviewer has no discretion over who enters the sample. Inclusion probabilities remain mathematically known and non-zero, preserving design-based statistical inference.
• Quota Sampling: Does not require an exhaustive sampling frame within strata. Once the subgroup quotas are set, individual participant selection is left to the discretion of field interviewers, recruitment staff, or automated online panel algorithms. The interviewer or platform recruits whoever is readily available, willing to participate, or easily reachable until the quota is filled.

Because the internal selection mechanism in quota sampling remains non-random, it is vulnerable to systematic availability bias. Interviewers naturally tend to approach individuals who appear more approachable, are readily available during normal working hours, or live in safer, easily accessible neighborhoods. Consequently, while a quota sample may match national census benchmarks on surface-level demographics, it frequently harbors unmeasured biases regarding personality, economic activity, and social attitudes.

7.3 How Does Sample Size Determination Differ Between the Two Paradigms?

The sample size determination process reflects the distinct mathematical and operational foundations of the two paradigms. In probability sampling, determining sample size ($n$) is a formal mathematical procedure derived from statistical power analysis, the chosen Type I error rate ($\alpha$), the desired statistical power ($1 – beta$), the expected population variance ($\sigma^2$), and the acceptable Margin of Error ($MOE$ or $E$). For a simple continuous parameter in an unstratified design, the required sample size is calculated using the classical formula:

$$n = \frac{Z_{alpha/2}^2 \cdot \sigma^2}{E^2}$$

Where $Z_{alpha/2}$ represents the critical value of the standard normal distribution. In complex designs involving clustering, this sample size must be adjusted upward using the Design Effect ($n_{\text{complex}} = n_{\text{srs}} \cdot Deff$) to account for intra-cluster correlation. The objective of sample size determination in probability sampling is to provide sufficient statistical precision to detect specified effect sizes within defined confidence bands.

In non-probability sampling—particularly within qualitative traditions—sample size determination is governed by conceptual criteria rather than frequentist variance calculations. The primary criterion is theoretical saturation: data collection continues until observing additional units yields no new theoretical dimensions, structural properties, or conceptual relationships. In quantitative non-probability designs (such as quota panels or online surveys), sample size is often dictated by pragmatic resource constraints, commercial pricing, or rules of thumb regarding cell counts needed to run complex post-stratification models. Crucially, in non-probability designs, expanding sample size does not resolve underlying selection bias; it simply reduces random variance around what may be a fundamentally biased estimate.

8. Conclusion and Strategic Synthesis

8.1 Synthesizing the Methodological Trade-Offs

The debate between probability and non-probability sampling does not resolve into a hierarchy where one methodology is universally superior to the other. Instead, it presents an analytical trade-off between competing research imperatives. The table below summarizes the core strategic trade-offs that govern this methodological choice:

Ultimately, methodological rigor requires an honest alignment between the chosen sampling technique, the underlying research questions, and the practical operational environment. A research design is not validated simply by incorporating random selection, nor is it invalidated because it relies on purposive recruitment. Rather, rigor depends on whether the sampling methodology can legitimately support the empirical claims being made.

8.2 The Future of Sampling in the Era of Big Data

The rapid expansion of the digital information ecosystem—marked by big data architectures, social media scraping, continuous telemetry, search engine tracking, and transactional data logs—has reignited fundamental debates around sampling methodology. Some data scientists and technology analysts have suggested that massive datasets render traditional sampling theory obsolete, proposing that analyzing millions of digital observations eliminates the need for survey sampling altogether.

However, modern statistical theory has demonstrated that large datasets cannot overcome systematic selection bias. In a landmark 2018 paper published in the Annals of Applied Statistics, statistician Xiao-Li Meng formulated the “Big Data Paradox,” demonstrating mathematically that as non-random datasets grow larger, confidence intervals narrow around increasingly biased estimates. Meng established that when evaluating total estimation error, a non-probability dataset containing millions of observations can harbor less effective information than a carefully drawn simple random sample of a few hundred individuals. This occurs because digital datasets systematically exclude segments of the population that lack internet connectivity, have low digital literacy, or consciously avoid digital platforms, creating severe coverage error.

As a result, the future of empirical research lies in hybrid sampling designs that bridge the scale of big data with the inferential rigor of probability methodology. Emerging research programs use small, carefully administered probability surveys as calibration anchors to audit, benchmark, and correct for participation bias in large non-probability datasets. Rather than replacing survey methodology, the era of big data has reinforced the foundational principles established by Jerzy Neyman nearly a century ago: true empirical insight requires a rigorous understanding of the mechanisms that determine how data are generated, captured, and selected.

References

• Bowley, A. L. (1906). Address to the Economic Science and Statistics Section of the British Association for the Advancement of Science. Journal of the Royal Statistical Society, 69(3), 540–558.
• Bowley, A. L. (1926). Measurement of the precision attained in sampling. Bulletin of the International Statistical Institute, 22(1), 1–62.
• Cochran, W. G. (1977). Sampling techniques (3rd ed.). John Wiley & Sons.
• Glaser, B. G., & Strauss, A. L. (1967). The discovery of grounded theory: Strategies for qualitative research. Aldine Publishing Company.
• Groves, R. M., Fowler, F. J., Couper, M. P., Lepkowski, J. M., Singer, E., & Tourangeau, R. (2009). Survey methodology (2nd ed.). John Wiley & Sons.
• Heckathorn, D. D. (1997). Respondent-driven sampling: A new approach to the study of hidden populations. Social Problems, 44(2), 174–199. https://doi.org/10.2307/3096941
• Horvitz, D. G., & Thompson, D. J. (1952). A generalization of sampling without replacement from a finite universe. Journal of the American Statistical Association, 47(260), 663–685. https://doi.org/10.1080/01621459.1952.10483446
• Kiær, A. N. (1895). Observations et expériences concernant les dénombrements représentatifs. Bulletin de l’Institut International de Statistique, 9(2), 176–183.
• Kish, L. (1965). Survey sampling. John Wiley & Sons.
• Little, R. J., & Rubin, D. B. (2019). Statistical analysis with missing data (3rd ed.). John Wiley & Sons.
• Malterud, K., Siersma, V. D., & Guassora, A. D. (2016). Sample size in qualitative interview studies: Guided by information power. Qualitative Health Research, 26(13), 1753–1760. https://doi.org/10.1177/1049732315617444
• Meng, X.-L. (2018). Statistical paradises and paradoxes in big data (I): Law of large populations, big data paradox, and the 2016 US presidential election. The Annals of Applied Statistics, 12(2), 685–726. https://doi.org/10.1214/18-AOAS1161SF
• Neyman, J. (1934). On the two different aspects of the representative method: The method of stratified sampling and the method of purposive selection. Journal of the Royal Statistical Society, 97(4), 558–625. https://doi.org/10.2307/2342192
• Rubin, D. B. (1976). Inference and missing data. Biometrika, 63(3), 581–592. https://doi.org/10.1093/biomet/63.3.581
• Särndal, C.-E., Swensson, B., & Wretman, J. (2003). Model assisted survey sampling. Springer Science & Business Media.