VALIDITY AND RELIABILITY

For the statistical consultant working with social science researchers the estimation of reliability and validity is a task frequently encountered.  Measurement issues differ in the social sciences in that they are related to the quantification of abstract, intangible and non observable constructs.  In many instances, then, the meaning of quantities is only inferred.

Let us begin by a general description of the paradigm that we are dealing with.  Most concepts in the behavioral sciences have meaning within the context of the theory that they are a part of.  Each concept, thus, has an operational definition which is governed by the overarching theory.  If a concept is involved in the testing of hypothesis to support the theory it has to be measured.  So the first decision that the research is faced with is “how shall the concept be measured?” That is the type of measure.  At a very broad level the type of measure can be observational, self-report, interview, etc.  These types ultimately take shape of a more specific form like observation of ongoing activity, observing video-taped events, self-report measures like questionnaires that can be open-ended or close-ended, Likert-type scales, interviews that are structured, semi-structured or unstructured and open-ended or close-ended.  Needless to say, each type of measure has specific types of issues that need to be addressed to make the measurement meaningful, accurate, and efficient.

Another important feature is the population for which the measure is intended.  This decision is not entirely dependent on the theoretical paradigm but more to the immediate research question at hand.

A third point that needs mentioning is the purpose of the scale or measure.  What is it that the researcher wants to do with the measure?  Is it developed for a specific study or is it developed with the anticipation of extensive use with similar populations?

Once some of these decisions are made and a measure is developed, which is a careful and tedious process, the relevant questions to raise are “how do we know that we are indeed measuring what we want to measure?” since the construct that we are measuring is abstract, and “can we be sure that if we repeated the measurement we will get the same result?”.  The first question is related to validity and second to reliability.  Validity and reliability are two important characteristics of behavioral measure and are referred to as psychometric properties.

It is important to bear in mind that validity and reliability are not an all or none issue but a matter of degree.

Validity:

Very simply, validity is the extent to which a test measures what it is supposed to measure.  The question of validity is raised in the context of the three points made above, the form of the test, the purpose of the test and the population for whom it is intended.  Therefore, we cannot ask the general question “Is this a valid test?”.  The question to ask is “how valid is this test for the decision that I need to make?” or “how valid is the interpretation I propose for the test?”  We can divide the types of validity into logical and empirical.

Content Validity:

When we want to find out if the entire content of the behavior/construct/area is represented in the test we compare the test task with the content of the behavior.  This is a logical method, not an empirical one.  Example, if we want to test knowledge on American Geography it is not fair to have most questions limited to the geography of New England.

Face Validity:

Basically face validity refers to the degree to which a test appears to measure what it purports to measure.

Criterion-Oriented or Predictive Validity:

When you are expecting a future performance based on the scores obtained currently by the measure, correlate the scores obtained with the performance.  The later performance is called the criterion and the current score is the prediction.  This is an empirical check on the value of the test – a criterion-oriented or predictive validation.

Concurrent Validity:

Concurrent validity is the degree to which the scores on a test are related to the scores on another, already established, test administered at the same time, or to some other valid criterion available at the same time.  Example, a new simple test is to be used in place of an old cumbersome one, which is considered useful, measurements are obtained on both at the same time.  Logically, predictive and concurrent validation are the same, the term concurrent validation is used to indicate that no time elapsed between measures.

Construct Validity:

Construct validity is the degree to which a test measures an intended hypothetical construct.  Many times psychologists assess/measure abstract attributes or constructs.  The process of validating the interpretations about that construct as indicated by the test score is construct validation.  This can be done experimentally, e.g., if we want to validate a measure of anxiety.  We have a hypothesis that anxiety increases when subjects are under the threat of an electric shock, then the threat of an electric shock should increase anxiety scores (note:  not all construct validation is this dramatic!)

A correlation coefficient is a statistical summary of the relation between two variables.  It is the most common way of reporting the answer to such questions as the following:  Does this test predict performance on the job?  Do these two tests measure the same thing?  Do the ranks of these people today agree with their ranks a year ago?

(rank correlation and product-moment correlation)

According to Cronbach, to the question “what is a good validity coefficient?” the only sensible answer is “the best you can get”, and it is unusual for a validity coefficient to rise above 0.60, though that is far from perfect prediction.

All in all we need to always keep in mind the contextual questions:  what is the test going to be used for? how expensive is it in terms of time, energy and money? what implications are we intending to draw from test scores?

Reliability:

Research requires dependable measurement.  (Nunnally) Measurements are reliable to the extent that they are repeatable and that any random influence which tends to make measurements different from occasion to occasion or circumstance to circumstance is a source of measurement error.  (Gay) Reliability is the degree to which a test consistently measures whatever it measures.  Errors of measurement that affect reliability are random errors and errors of measurement that affect validity are systematic or constant errors.

Test-retest, equivalent forms and split-half reliability are all determined through correlation.

Test-retest Reliability:

Test-retest reliability is the degree to which scores are consistent over time.  It indicates score variation that occurs from testing session to testing session as a result of errors of measurement.  Problems:  Memory, Maturation, Learning.

Equivalent-Forms or Alternate-Forms Reliability:

Two tests that are identical in every way except for the actual items included.  Used when it is likely that test takers will recall responses made during the first session and when alternate forms are available.  Correlate the two scores.  The obtained coefficient is called the coefficient of stability or coefficient of equivalence.  Problem:  Difficulty of constructing two forms that are essentially equivalent.

Both of the above require two administrations.

Split-Half Reliability:

Requires only one administration.  Especially appropriate when the test is very long.  The most commonly used method to split the test into two is using the odd-even strategy.  Since longer tests tend to be more reliable, and since split-half reliability represents the reliability of a test only half as long as the actual test, a correction formula must be applied to the coefficient.  Spearman-Brown prophecy formula.

Split-half reliability is a form of internal consistency reliability.

Rationale Equivalence Reliability:

Rationale equivalence reliability is not established through correlation but rather estimates internal consistency by determining how all items on a test relate to all other items and to the total test.

Internal Consistency Reliability:

Determining how all items on the test relate to all other items.  Kudser-Richardson-> is an estimate of reliability that is essentially equivalent to the average of the split-half reliabilities computed for all possible halves.

Standard Error of Measurement:

Reliability can also be expressed in terms of the standard error of measurement.  It is an estimate of how often you can expect errors of a given size.

REFERENCES

Berk, R., 1979.  Generalizability of Behavioral Observations:  A Clarification of Interobserver Agreement and Inter observer Reliability.  American Journal of Mental Deficiency, Vol. 83, No. 5, p. 460-472.

Cronbach, L., 1990.  Essentials of psychological testing.  Harper & Row, New York.

Carmines, E., and Zeller, R., 1979.  Reliability and Validity Assessment. Sage Publications, Beverly Hills, California.