560 likes | 1.4k Vues
Validity – Outline. Definition Validity: Two Different Views Types of Validity Face Content Criterion Predictive vs. Concurrent Validity Coefficients Construct Convergent Discriminant. Validity measures agreement between a test score and the characteristic it is believed to measure.
E N D
Validity – Outline • Definition • Validity: Two Different Views • Types of Validity • Face • Content • Criterion • Predictive vs. Concurrent • Validity Coefficients • Construct • Convergent • Discriminant
Validity measures agreement between a test score and the characteristic it is believed to measure The basic question is: are you measuring what you think you’re measuring? Validity – Definition
Traditional: Validity is a property of tests Does the test measure what you think it measures? Validity: two very different views
Traditional Recent (e.g, Messick, 1989; Committee on Standards for Educational and Psychological Testing (CSEPT)): Validity is a property of test score interpretations Validity exists when actions based on the interpretation are justified given a theoretical basis and social consequences Validity: two very different views
Does the test measure what you think it measures? Validity exists when actions based on the interpretation are justified given a theoretical basis and social consequences Note the difference:
Who is to say the ‘social consequences’ of test use are good or bad? According to CSEPT validity is a subjective judgment In my view, this makes the concept useless: “if you like the result the test gives you, you will consider it valid. If you don’t, you won’t.” That’s not how scientists think. A problem with the CSEPT view
Borsboom et al reject CSEPT’s view “Validity… is a very basic concept and was correctly formulated, for instance, by Kelley (1927, p. 14) when he stated that a test is valid if it measures what it purports to measure.” (p. 1061) Borsboom et al. (2004)
“a test is valid for measuring an attribute if and only if (a) the attribute exists and (b) variations in the attribute causally produce variations in the outcomes of the measurement procedure.” Variations in what you are measuring cause variations in your measurements. E.g., variations across people in intelligence cause variations in their IQ scores This is not a correlational model of validity Borsboom et al. (2004)
You don’t create a test and then do the analysis necessary to establish its validity Rather, you begin by doing the theoretical work necessary to create a valid test in the first place. On this view, validity is not a big issue. Borsboom et al. (2004)
Who is right? Each scientist has to make up his or her own mind on that question I find Borsboom et al.’s arguments compelling. Other psychologists may disagree Borsboom et al. vs. CSEPT
CSEPT recognizes 3 types of evidence for test validity: Content-related Criterion-related Construct-related Boundaries not clearly defined Cronbach (1980): Construct is basic, while Content & Criterion are subtypes. The CSEPT view
Face validity refers to the appearance that a test measures what it is intended to measure. Face validity has P.R. value – test-takers may have better motivation if the test appears to be a sensible way to measure what it measures. Parenthetical Point – Face Validity
Content-related evidence considers coverage of the conceptual domain tested. Important in educational settings Like face validity, it is determined by logic rather than statistics Typically assessed by expert judges CSEPT: Content validity
Content-related evidence considers coverage of the conceptual domain tested. Construct-irrelevant variance Construct under-representation Is each item relevant to domain? Is domain adequately covered or are parts of it left out? But if you are going to ask these questions, why not do it when creating the test? CSEPT: Content validity
Borsboom et al. would say that content validity is not something to be established after the test has been created. Rather, you build it into your test by having a good theory of what you are testing E.g., for a test in this course to have content validity, it should test your understanding of content validity! Borsboom et al.: Content validity
Criterion-related evidence tells us how well a test score corresponds to a particular criterion measure. A criterion is a standard against which a test is compared. The test score should tell us something about the criterion score. CSEPT: Criterion validity
A criterion is a standard against which a test is compared. E.g., we could compare GPAs to SAT scores to produce evidence of validity of conclusions drawn on basis of SAT scores Two basic types: Predictive Concurrent CSEPT: Criterion validity
Predictive validity Test scores used to predict future performance – how good is the prediction? E.g., SAT is used to predict final undergraduate GPA SAT – GPA are moderately correlated CSEPT: Criterion validity
Predictive validity Concurrent validity Correlation between test scores and criterion when the two are measured at same time. Test illuminates current performance rather than predicting future performance (e.g., why does patient have a temperature? Why can’t student do math?) CSEPT: Criterion validity
“Criterion validity” involves a correlation, of test scores with some criterion such as GPA That does not establish the test’s validity, only its utility. E.g., height and weight are correlated, but a test of height is not a test of what bathroom scales measure. Borsboom et al.: Criterion validity
SAT is valid because it was developed on the sensible theory that “past academic achievement” is a good guide to “future academic achievement” Validity is built into the test, not established after the test has been created Borsboom et al.: Criterion validity
Validation research aims at showing how variation in the attribute causes variation in the test score This requires a “theory of the task”: how does the test-taker do the mental operations needed to respond to test items? Borsboom et al.: Criterion validity
Note: no point in developing a test if you already have a criterion – unless impracticality or expense makes use of the criterion difficult. Criterion measure only available in the future? Criterion too expensive to use? CSEPT: Criterion validity
Validity Coefficient Compute correlation (r) between test score and criterion. r = .30 or .40 would be considered normal. r > .60 is rare Note: r varies between -1.0 and +1.0 CSEPT: Criterion validity
Validity Coefficient r2 gives proportion of variance in criterion explained by test score. E.g., if rxy = .30, r2 = .09, so 9% of variability in Y “can be explained by variation in X” CSEPT: Criterion validity
Interpreting Validity Coefficients – watch out for: Changes in causal relationships What does criterion mean? Is it valid, reliable? Is subject population for validity study appropriate? Sample size CSEPT: Criterion validity
Interpreting Validity Coefficients – watch out for: Criterion/predictor confusion Range restrictions Do validity study results generalize? Differential predictions CSEPT: Criterion validity
Problem: for many psychological characteristics of interest there is no agreed-upon “universe” of content and no clear criterion We cannot assess content or criterion validity for such characteristics These characteristics involve constructs: something built by mental synthesis. CSEPT: Construct validity
Examples of constructs: Intelligence Love Curiosity Mental health CSEPT: We obtain evidence of validity by simultaneously defining the construct and developing instruments to measure it. This is ‘bootstrapping.’ CSEPT: Construct validity
assemble evidence about what a test “means” – in other words, about the characteristic it is testing. CSEPT: this process is never finished Borsboom: this is part of the process of creating a test in the first place, not something done after the fact Bootstrapping construct validity
assemble evidence show relationships between a test and other tests none of the other tests is a criterion Borsboom: these relationships do not tell us what a test score means (e.g., age is correlated with annual income but a measure of age is not a measure of annual income). Bootstrapping construct validity
assemble evidence show relationships each new relationship adds meaning to the test test’s meaning is gradually clarified over time Borsboom would say, why all the mystery? The meaning of many tests (e.g., WAIS, academic exams, Piaget’s tests) is clear right from the start Bootstrapping construct validity
Example from text: Rubin’s work on Love. Rubin collected a set of items for a Love scale He read poetry, novels; asked people for definitions created a scale of Love and one of Liking CSEPT: Construct validity
Rubin gave scale to many subjects & factor-analyzed results Love integrates Attachment, Caring, & Intimacy Liking integrates Adjustment, Maturity, Good Judgment, and Intelligence The two are independent: you can love someone you don’t like (as song-writers know) CSEPT: Construct validity
Two types of Construct-related Evidence Convergent evidence When a test correlates well with other tests believed to measure the same construct Campbell & Fiske (1959)
Two types of Construct-related Evidence Convergent evidence Discriminant evidence When a test does not correlate with other tests believed to measure some other construct. Campbell & Fiske (1959)
Example – Health Index Scores correlated with age, number of symptoms, chronic medical conditions, physiological measures Treatments designed to improve health should increase Health Index scores. They do. Convergent validity
low correlations between new test and tests believed to tap unrelated constructs. evidence that the new test measures something unique Discriminant validity
CSEPT: No point in trying to establish validity of an unreliable test. It’s possible to have a reliable test that has no meaning (is not valid). Logically impossible to produce evidence of validity for an unreliable test. CSEPT: Validity & Reliability
Borsboom et al: what does it mean to say that a test is reliable but not valid? What is it a test of? It isn’t a test at all, just a collection of items Borsboom: Validity & Reliability
Borsboom et al: validity is a necessary condition for reliability Reliability of a test of X estimates precision of measurement of X – but how could you estimate the precision of measurement of X for a test that does not measure X? Thus, validity is presumed when you assess reliability Borsboom: Validity & Reliability
We observe a behavior in order to learn about the underlying psychological characteristic A person’s test score represents their standing on that underlying dimension Such scores form an arbitrary metric That is, we do not know how the observed scores are related to the true scores on the underlying dimension Blanton & Jaccard – arbitrary metrics
Person A Person B Underlying dimension Neutral Test 1 0 1 2 3 4 5 6 Test 2 0 1 2 3 4 5 6 Adapted from Blanton & Jaccard (2006) Figure 1, p. 29
Implicit Association Test (IAT) – claimed to diagnose implicit attitudinal preferences – or racist attitudes IAT authors say you may have prejudices you don’t know you have. Are these claims true? Arbitrary metrics – the IAT
Task: categorize stimuli using two pairs of categories Two buttons to press, two assignments of categories to buttons, used in sequence Arbitrary metrics – the IAT
Assignment pattern A Button 1 – press if stimulus refers to the category White or the category Pleasant Button 2 – press if stimulus refers to the category Black or the category Unpleasant Assignment pattern B Button 1 – press if stimulus refers to the category White or the category Unpleasant Button 2 – press if stimulus refers to the category Black or the category Pleasant Arbitrary metrics – the IAT
IAT authors claim that if responses are faster to Pattern A than to Pattern B, that indicates a “preference” for Whites over Blacks – in other words, a racist attitude IAT authors also give test-takers feedback about how strong their preferences are, based on how much faster their responses are to Pattern A than to Pattern B This is inappropriate Arbitrary metrics – the IAT
Blanton & Jaccard: The IAT does not tell us about racist attitudes IAT authors take a dimension which is non-arbitrary when used by physicists – time – and use it in an arbitrary way in psychology Arbitrary metrics – the IAT
The function relating the response dimension (time) to the underlying dimension (attitudes) is unknown Zero on the (Pattern A – Pattern B) difference may not be zero on the underlying attitude preference dimension There are alternative models of how that (Pattern A – Pattern B) difference could arise Arbitrary metrics – the IAT
CSEPT: Validity is a characteristic of evidence, not of tests. Valid evidence supports conclusions drawn using test results Validity is determined by social consequences of test use Borsboom et al. Validity is not a methodological issue, but a substantive (theoretical) issue A test of an attribute is valid if (a) the attribute exists, and (b) variation in the attribute causes variation in test scores Review