KUDER-RICHARDSON FORMULAS

Introduction and Historical Context

The field of psychometrics, which focuses on the theory and technique of psychological measurement, places paramount importance on the accurate assessment of test quality. Central to this assessment is the concept of reliability, which refers to the consistency of a measurement tool. Among the various methods developed to estimate reliability, measures of internal consistency are critical, assessing how well different items on a test measure the same underlying construct. The Kuder-Richardson (KR) formulas represent a seminal contribution to this area, providing a robust statistical framework for calculating the internal consistency of tests scored dichotomously.

Developed and published in 1937 by psychologists George Kuder and Mary Richardson, these formulas emerged during a period of intense focus on standardizing psychological assessment instruments. Their work was aimed at addressing the limitations of earlier split-half reliability methods, which often yielded inconsistent results depending on how the test was arbitrarily divided. The introduction of the KR formulas offered a mathematically superior solution that essentially considered the correlation of every possible split-half configuration simultaneously, thus providing a single, more stable estimate of internal consistency. This innovation cemented the KR formulas as foundational tools in psychometric theory, particularly for aptitude and achievement tests where responses are typically scored as correct (1) or incorrect (0).

While often generalized in modern research by the later development of Cronbach’s Alpha (1951), the KR formulas, specifically KR-20 and KR-21, remain essential references. They are widely taught and applied, especially when dealing exclusively with binary data. Understanding their derivation and underlying assumptions is crucial for any researcher utilizing standardized testing, as they provide the mathematical backbone for establishing the trustworthiness of instruments designed to measure constructs across various domains, including education, clinical assessment, and organizational psychology.

Defining Internal Consistency Reliability

Internal consistency reliability is a measure based on the correlational relationships between different items on the same test, all administered at a single point in time. It estimates the degree to which all items measure a single, common characteristic or domain. When a test exhibits high internal consistency, it suggests that respondents who perform well on one item tend to perform well on other items, implying a high degree of content homogeneity. This type of reliability is essential because if items within a scale are measuring disparate constructs, the overall test score loses its meaning as a unitary measure of the intended trait.

Unlike test-retest reliability, which examines consistency over time, or inter-rater reliability, which examines consistency across different scorers, internal consistency is derived exclusively from a single administration of the test. The core psychometric concept behind the KR formulas involves the partitioning of variance. The total variance observed in the test scores is conceptually divided into variance attributable to the true underlying construct (true score variance) and variance attributable to random measurement error. A higher ratio of true score variance to total variance results in a higher reliability coefficient, which theoretically approaches the maximum value of 1.00.

The reliability coefficient obtained from the KR formulas can be interpreted as the estimated correlation between the actual test and all other possible tests of equal length that measure the same construct. Consequently, a high coefficient (e.g., above 0.80 for standard research) indicates that the test items are highly homogeneous and reliably measure the intended psychological trait. Conversely, a low coefficient suggests that the items are heterogeneous, or that the score is significantly contaminated by random measurement error, potentially rendering interpretations derived from the test invalid or misleading for individual assessment.

The Kuder-Richardson Formula 20 (KR-20): Derivation and Application

The Kuder-Richardson Formula 20 (KR-20) is the most widely recognized and mathematically rigorous of the original formulas. It is specifically designed for instruments where items are scored dichotomously, meaning responses must be classified as one of two outcomes, typically represented by 0 (incorrect/negative) or 1 (correct/positive). The KR-20 formula provides a lower-bound estimate of the true reliability of the test, meaning the true reliability is likely equal to or greater than the calculated KR-20 value. This formula is particularly valuable because its calculation requires only a single administration of the test and relies solely on item statistics derived from that specific testing event.

The mathematical representation of KR-20 requires the calculation of the proportion of examinees who answered each item correctly and the total variance of the observed scores. The formula is structured to compare the variability within individual items against the overall variability of the total test scores. The standard presentation of the formula is:

• $$KR-20 = (k / (k – 1)) times [1 – ((text{Sum of item variances}) / (text{Total test score variance}))]$$

More formally, where ‘k’ represents the total number of items on the test, ‘$$p_i$$’ is the proportion of examinees who pass item ‘i’, ‘$$q_i$$’ is the proportion who fail item ‘i’ ($$q_i = 1 – p_i$$), and ‘$$sigma_x^2$$’ represents the total variance of the observed test scores. The term $$p_i q_i$$ is mathematically equivalent to the variance of a single dichotomous item. Thus, the crucial component, $$sum p_i q_i$$, is the sum of the variances of all individual items. This sum represents the error variance if the items were independent, and it is compared against the total variance of the whole test score, $$sigma_x^2$$.

The KR-20 formula is highly robust because it intrinsically accounts for differences in item difficulty levels across the test. If items vary widely in difficulty (meaning the $$p_i$$ values are diverse), the KR-20 calculation remains accurate, provided the underlying psychometric assumption of unidimensionality (measuring a single construct) is strictly met. This inherent precision makes KR-20 the preferred calculation when detailed item statistics are available, offering a comprehensive and unbiased assessment of item coherence.

Assumptions of the KR-20 Formula

The accurate application and interpretation of the KR-20 coefficient depend on satisfying several critical assumptions inherent to its mathematical derivation. The most critical assumption is that the test must be unidimensional, meaning that all items are designed to measure a single, common underlying latent trait or construct. If a test inadvertently measures multiple distinct psychological factors—for instance, if a math test measures both computational skills and reading comprehension—the resulting KR-20 coefficient may be misleading. It might underestimate the reliability of the intended construct or, conversely, provide an inflated overall estimate that masks poor item performance on one of the unintended factors.

A second foundational assumption is the requirement that all items must be dichotomous, meaning they must be scored strictly 0 or 1. This is non-negotiable for the KR formulas. If an instrument utilizes polytomous scoring—such as a five-point Likert scale, or allows for partial credit—the KR-20 formula is mathematically inappropriate because the calculation of item variance using the $$p_i q_i$$ term is only valid for binary outcomes. In such cases, the researcher must utilize its generalization, Cronbach’s Alpha, which can accommodate the variance calculations required for continuous or ordinal data.

Furthermore, KR-20 relies on the assumption of essential tau-equivalence. This implies that while the items may differ in difficulty (allowing for varying $$p_i$$ values), they are measuring the same true score construct using the same scale of measurement, and importantly, the errors associated with different items must be uncorrelated. If item errors are correlated—perhaps because a cluster of items is systematically easier or harder due to their physical proximity on the test or external factors like shared stimulus material—the calculated KR-20 coefficient may be artificially inflated. Researchers must therefore conduct thorough item analysis and factor analysis to confirm that these fundamental psychometric conditions are met before reporting the KR-20 coefficient as a reliable estimate of internal consistency.

The Kuder-Richardson Formula 21 (KR-21): Simplified Calculation and Limitations

The Kuder-Richardson Formula 21 (KR-21) was introduced primarily as a computational shortcut when calculating individual item variances was tedious or impossible prior to the widespread availability of computer processing. While the KR-20 requires detailed statistics for every item, the KR-21 makes a powerful, simplifying assumption: that all items are of equal difficulty. This allows the calculation to rely solely on the mean score of the total test, rather than requiring the individual item proportion statistics ($$p_i$$ and $$q_i$$ values).

The mathematical structure of KR-21 incorporates the mean score (M) of the test directly. The formula is written as:

• $$KR-21 = (k / (k – 1)) times [1 – ((M times (k – M)) / (k times sigma_x^2))]$$

Where ‘k’ is the number of items, ‘M’ is the mean test score (the sum of all correct responses divided by the number of examinees), and ‘$$sigma_x^2$$’ is the total test variance. The term $$M(k – M)/k$$ serves as an approximation of the sum of item variances ($sum p_i q_i$) used in KR-20, assuming that if all items were equally difficult, the sum of their variances could be reliably estimated from the mean score alone.

Due to its simplifying assumption, KR-21 is almost universally a lower-bound estimate of the reliability compared to KR-20, unless the test items happen to be perfectly equal in difficulty, a rare occurrence in real-world educational or psychological testing. If item difficulties vary significantly, KR-21 will yield a coefficient that substantially underestimates the true internal consistency. Consequently, while KR-21 historically offered ease of calculation, psychometricians today strongly prefer KR-20 because it does not require the unrealistic assumption of equal item difficulty. KR-21 is now largely relegated to pedagogical settings for illustrating the relationship between mean score and variance, rather than being used for rigorous reporting in research publications.

Distinguishing KR Formulas from Cronbach’s Alpha

The relationship between the KR formulas and Cronbach’s Alpha ($alpha$) is a critical concept in psychometrics. Both coefficients measure the same attribute—internal consistency reliability derived from a single test administration—but they differ strictly in their domain of application based on the scoring format of the test items. This distinction is not mathematical but operational.

The essential difference lies in the nature of the data. KR formulas, specifically KR-20 and KR-21, are mathematically restricted to tests consisting solely of dichotomous items (scored 0 or 1). Cronbach’s Alpha, developed by Lee Cronbach in 1951, is a direct statistical generalization of the KR-20 formula. Alpha is utilized when test items are scored polytomously, meaning they involve graded responses, such as Likert-type scales (e.g., measuring agreement on a 1-to-5 scale), or instruments using partial credit scoring.

Crucially, if a researcher calculates Cronbach’s Alpha for a set of items that are all strictly dichotomously scored, the resulting coefficient will be mathematically identical to the KR-20 coefficient. Alpha’s flexibility stems from its variance calculation, which accommodates continuous or ordinal data where the variance of an item is calculated using the standard variance formula, rather than the simplified $$p_i q_i$$ term used exclusively for binary data. Therefore, the choice between KR-20 and Alpha is dictated solely by the measurement format of the instrument. The underlying psychometric assumptions, including unidimensionality and essential tau-equivalence, remain constant for both statistics, confirming that they are two applications of the same core reliability model.

Practical Applications in Psychometrics

The KR formulas hold significant practical importance across various domains of psychological measurement, particularly in the development and validation of standardized assessments used in educational and occupational settings. Before a test can be deemed suitable for high-stakes decision-making, such as placement, selection, or certification, its internal consistency must be rigorously documented using KR-20, provided the test is structured with binary scoring.

In educational psychology, KR-20 is routinely calculated for standardized achievement batteries, classroom exams, and licensure assessments. A high KR-20 coefficient provides assurance to educators and policymakers that the test is consistently measuring the intended subject matter knowledge and is not unduly influenced by random error or ambiguous item wording. If the KR-20 value is low, it serves as a powerful diagnostic signal indicating the urgent need for item revision or deletion. This prompts researchers to conduct detailed item analysis to identify poorly functioning items that may be measuring extraneous constructs, thereby contaminating the overall test score.

Beyond the validation phase, the structure of the KR formulas is also integral to estimating the necessary length of a test. Psychometric theory dictates that, generally, longer tests achieve higher reliability. The components of the KR formulas, especially the item variance structure, can be used in conjunction with the Spearman-Brown prophecy formula to predict how the reliability would change if the test were lengthened or shortened. This allows test developers to optimize the number of items required to achieve a desired minimum level of reliability while maintaining efficiency in the assessment process.

Interpretation and Reporting of KR Coefficients

The calculated KR coefficient is a unitless value ranging from 0.00 to 1.00. A coefficient of 1.00 signifies perfect consistency (zero measurement error), while 0.00 indicates that the test scores are composed entirely of random error. Interpreting the magnitude of the coefficient is context-dependent, relying heavily on the nature and stakes associated with the test being administered.

General guidelines often suggest different thresholds for acceptable reliability. For basic research purposes, where conclusions are drawn about group differences, a coefficient of 0.70 is often considered the minimally acceptable floor. However, for applied clinical or high-stakes decision-making involving individual scores (e.g., psychological diagnosis, professional certification, or selection), a much higher reliability is required, typically ranging from 0.90 to 0.95. When reporting results, researchers must explicitly state the specific formula used (KR-20 or KR-21), the characteristics of the sample tested, and the calculated coefficient value, ensuring transparency in the psychometric quality assessment.

It is fundamentally important to recognize that a high KR coefficient indicates internal consistency, but it does not guarantee validity. A test can be highly reliable (consistent) yet systematically measure the wrong construct (invalid). Conversely, a low KR coefficient immediately signals a severe problem with the test structure itself. If the reliability is low, the scores cannot be trusted as consistent measures of any trait, regardless of the test’s theoretical validity. Therefore, KR-20 serves as a necessary, though insufficient, prerequisite for establishing the overall psychometric soundness of a dichotomously scored instrument.

Limitations and Alternatives

Despite their enduring historical significance and ongoing utility for data scored dichotomously, the Kuder-Richardson formulas possess inherent limitations that restrict their application in modern psychological practice. The most obvious restriction is their strict requirement for binary scoring, which excludes them from being used with the vast majority of contemporary psychological scales that employ graded response formats, such as personality inventories, clinical symptom checklists, or attitude measures.

Furthermore, the reliance on the strong assumption of unidimensionality poses a recurring challenge. If a test is intentionally designed to be multidimensional (i.e., measuring several related but distinct factors), calculating a single, overall KR-20 coefficient for the entire scale will yield a misleading result that is often lower than the true reliability of the subcomponents. In these cases, the correct procedure is to calculate separate KR-20 coefficients for each identified subscale or factor, ensuring that the internal consistency is assessed strictly within the bounds of each distinct dimension. Failure to first confirm the factor structure through appropriate statistical methods can lead to erroneous conclusions about item quality and test construction.

The principal alternative and generalization to the KR-20 formula is Cronbach’s Alpha. Alpha is now the standard coefficient reported in almost all psychological research for scales that use polytomous scoring. In highly specialized or advanced psychometric contexts, particularly when the assumption of essential tau-equivalence is violated, researchers may opt for even more sophisticated alternatives like McDonald’s Omega ($omega$). Omega, typically derived through confirmatory factor analysis (CFA), relaxes the restrictive assumption of tau-equivalence and is argued by some to provide a potentially more accurate estimate of composite reliability, especially in tests exhibiting complex factor structures. Nevertheless, for basic, standardized tests based on binary scoring, KR-20 remains the fundamental and most appropriate calculation.

Conclusion

The Kuder-Richardson formulas, particularly KR-20, represent a foundational achievement in the history of psychometrics. Developed specifically to overcome the inherent inconsistencies of earlier reliability estimation methods, they provide a statistically sound and empirically tested method for quantifying the internal consistency of instruments scored dichotomously. The accurate application of KR-20 is fundamental to the validation of achievement, aptitude, and knowledge tests, ensuring that researchers and practitioners rely on measures that are consistent, coherent, and trustworthy.

Although the simplified KR-21 formula has largely been superseded due to its unrealistic assumption of equal item difficulty, and while Cronbach’s Alpha has become the indispensable standard for polytomous data, the legacy of Kuder and Richardson remains central to measurement theory. Understanding the derivation, specific assumptions, and limitations of KR-20 is essential for proper test construction and rigorous evaluation. By systematically comparing the sum of individual item variances against the total test variance, KR-20 offers a powerful diagnostic tool for assessing item quality and establishing the necessary reliability for valid psychological inference.

Ultimately, the rigorous use of these coefficients serves to uphold the ethical and scientific standards of psychological research and assessment. The calculation and reporting of high KR coefficients confirm that the measurement instrument is internally coherent and free from excessive random error, a critical prerequisite for making informed and responsible decisions based on observed test scores.