14 Reliability and Validity of Measurement

From the theoretical to the empirical plane

In a previous chapter, we discussed that doing research involves moving from the theoretical plane to the empirical plane. In order to do so, we need to become very specific about how we manipulate (in experiments) and measure (in most research) our independent and dependent variables.

But how do researchers know that the scores actually represent the underlying construct, especially for multidimensional and perhaps non-obvious or non-intuitive constructs? The answer is that they conduct research using the measure to confirm that the scores make sense based on their understanding of the construct being measured. This is an extremely important point. Business researchers do not simply assume that their measures work. Instead, they collect data to demonstrate that they work. If their research does not demonstrate that a measure works, they stop using it.

As an informal example, imagine that you have been dieting for a month. Your clothes seem to be fitting more loosely, and several friends have asked if you have lost weight. If at this point your bathroom scale indicated that you had lost 10 pounds, this would make sense and you would continue to use the scale. But if it indicated that you had gained 10 pounds, you would rightly conclude that it was broken and either fix it or get rid of it. In evaluating a measurement method, business researchers consider two general dimensions: reliability and validity.

Reliability

Reliability refers to the consistency of a measure. Business researchers consider three types of consistency: over time (test-retest reliability), across items (internal consistency), and across different researchers (inter-rater reliability).

Test-Retest Reliability

When researchers measure a construct that they assume to be consistent across time, then the scores they obtain should also be consistent across time. Test-retest reliability is the extent to which this is actually the case. For example, intelligence is generally thought to be consistent across time. A person who is highly intelligent today will be highly intelligent next week. This means that any good measure of intelligence should produce roughly the same scores for this individual next week as it does today. Clearly, a measure that produces highly inconsistent scores over time cannot be a very good measure of a construct that is supposed to be consistent.

Assessing test-retest reliability requires using the measure on a group of people at one time, using it again on the same group of people at a later time, and then looking at the test-retest correlation between the two sets of scores. This is typically done by graphing the data in a scatterplot and computing the correlation coefficient. Figure 15.1 shows the correlation between two sets of scores of several university students on the Rosenberg Self-Esteem Scale, administered two times, a week apart, with the score in week 1 on the x axis, and the score in week 2 on the y axis. The correlation coefficient for these data is +.95. In general, a test-retest correlation of +.80 or greater is considered to indicate good reliability.

Again, high test-retest correlations make sense when the construct being measured is assumed to be consistent over time, which is the case for many different constructs in business research, mainly if they are specific to a respondent. However, there are also many constructs which may change over long time periods (e.g. manager competence; managers may learn), or even over short periods (e.g. customer satisfaction). A low test-retest correlation for such measures would not be immediate cause for concern.

Internal Consistency

Another kind of reliability is internal consistency, which is the consistency of people’s responses across the items on a multiple-item measure. In general, all the items on such measures are supposed to reflect the same underlying construct, so people’s scores on those items should be correlated with each other. For example, the Rosenberg Self-Esteem Scale has multiple items (survey questions) trying to measure self-esteem. If people’s responses to the different items are not correlated with each other, then it would no longer make sense to claim that they are all measuring the same underlying construct. This is as true for behavioral measures and for self-report measures. For example, it would be odd if respondents would respond very differently to different items trying to measure service quality using the SERVQUAL scale we discussed in an earlier chapter. Internal consistency is not limited to survey questions. For example, in experiments in which we’d ask respondents to donate or perform some other behavior, we would also expect consistent behavior if we’d ask them to donate in multiple instances.

Like test-retest reliability, internal consistency can only be assessed by collecting and analyzing data. One approach is to look at a split-half correlation. This involves splitting the items into two sets, such as the first and second halves of the items or the even- and odd-numbered items. Then a score is computed for each set of items, and the relationship between the two sets of scores is examined. For example, Figure 15.2 shows the split-half correlation between several university students’ scores on the even-numbered items and their scores on the odd-numbered items of the Rosenberg Self-Esteem Scale. The correlation coefficient for these data is +.88. A split-half correlation of +.80 or greater is generally considered good internal consistency.

Perhaps the most common measure of internal consistency used by researchers in psychology is a statistic called Cronbach’s α (the Greek letter alpha). Conceptually, α is the mean of all possible split-half correlations for a set of items. For example, there are 252 ways to split a set of 10 items into two sets of five. Cronbach’s α would be the mean of the 252 split-half correlations. Note that this is not how α is actually computed, but it is a correct way of interpreting the meaning of this statistic. Again, a value of +.80 or greater is generally taken to indicate good internal consistency.

Interrater Reliability

Many behavioral measures involve significant judgment on the part of an observer or a rater. Interrater reliability is the extent to which different observers are consistent in their judgments. For example, if you were interested in measuring university students’ social skills, you could make video recordings of them as they interacted with another student whom they are meeting for the first time. Then you could have two or more observers watch the videos and rate each student’s level of social skills. To the extent that each participant does, in fact, have some level of social skills that can be detected by an attentive observer, different observers’ ratings should be highly correlated with each other. Interrater reliability is useful in many contexts. One context in which it often appears in business research is in meta-analyses. Meta-analyses code and analyze results from prior work. Before running analyses, it’s important to be certain that this coding was done correctly, which is why often at least two independent raters code the same studies. Subsequently, interrater reliability is calculated, to ensure both raters sufficiently agree. Interrater reliability is often assessed using Cronbach’s α when the judgments are quantitative or an analogous statistic called Cohen’s κ (the Greek letter kappa) when they are categorical.

Validity

Validity is the extent to which the scores from a measure represent the variable they are intended to. But how do researchers make this judgment? We have already considered one factor that they take into account—reliability. When a measure has good test-retest reliability and internal consistency, researchers should be more confident that the scores represent what they are supposed to. There has to be more to it, however, because a measure can be extremely reliable but have no validity whatsoever. As an absurd example, imagine someone who believes that people’s index finger length reflects their self-esteem and therefore tries to measure self-esteem by holding a ruler up to people’s index fingers. Although this measure would have extremely good test-retest reliability (index fingers don’t grow that quickly!), it would have absolutely no validity. The fact that one person’s index finger is a centimeter longer than another’s would indicate nothing about which one had higher self-esteem.

Discussions of validity usually divide it into several distinct types But a good way to interpret these types is that they are other kinds of evidence—in addition to reliability—that should be taken into account when judging the validity of a measure. Here we consider three basic kinds: face validity, content validity, and criterion validity.

Face Validity

Face validity is the extent to which a measurement method appears “on its face” to measure the construct of interest. Most people would expect a self-esteem questionnaire to include items about whether they see themselves as a person of worth and whether they think they have good qualities. So a questionnaire that included these kinds of items would have good face validity. The finger-length method of measuring self-esteem, on the other hand, seems to have nothing to do with self-esteem and therefore has poor face validity. Although face validity can be assessed quantitatively—for example, by having a large sample of people rate a measure in terms of whether it appears to measure what it is intended to—it is usually assessed informally and -to some extent- subjectively. Different experts might disagree about whether a measure is valid ‘at first glance’.

Face validity is at best a very weak kind of evidence that a measurement method is measuring what it is supposed to. One reason is that it is based on people’s intuitions about human behavior, which are frequently wrong

Content Validity

Content validity is the extent to which a measure “covers” the construct of interest. For example, consider that attitudes are usually defined as involving thoughts, feelings, and actions toward something. By this conceptual definition, a person has a positive attitude toward a firm to the extent that they think positive thoughts about this firm, would consider recommending the firm to others, and so on. So to have good content validity, a measure of people’s attitudes toward a firm would have to reflect all important aspects. Like face validity, content validity is not usually assessed quantitatively. Instead, it is assessed by carefully checking the measurement method against the conceptual definition of the construct, especially to see if all relevant underlying dimensions from the conceptual definitions are actually present in the measurement.

Criterion Validity

Criterion validity is the extent to which people’s scores on a measure are correlated with other variables (known as criteria) that one would expect them to be correlated with. For example, people’s scores on a a customer loyalty measure should be positively correlated with whether they actually stay at this firm. A measure of employee competence should (hopefully) be correlated with their end-of-year performance assessments, and so forth. If researchers would find that employees’ end-of-year assessments would be negatively correlated -or not correlated- to an employee competence assessment, this would cast doubt on the validity of the measure (or, in this specific case, perhaps on the performance assessments…).

A criterion can be any variable that one has reason to think should be correlated with the construct being measured, and there will usually be many of them. For example, one would expect a measure of physical risk taking to be correlated with their participation in “extreme” activities such as snowboarding and rock climbing, the number of speeding tickets they have received, and even the number of broken bones they have had over the years. When the criterion is measured at the same time as the construct, criterion validity is referred to as concurrent validity; however, when the criterion is measured at some point in the future (after the construct has been measured), it is referred to as predictive validity (because scores on the measure have “predicted” a future outcome). Criteria can also include other measures of the same construct. For example, one would expect new measures of test anxiety or physical risk taking to be positively correlated with existing established measures of the same constructs. This is known as convergent validity.

Assessing convergent validity requires collecting data using the measure. Researchers John Cacioppo and Richard Petty did this when they created their self-report Need for Cognition Scale to measure how much people value and engage in thinking (Cacioppo & Petty, 1982)^[1]. In a series of studies, they showed that people’s scores were positively correlated with their scores on a standardized academic achievement test, and that their scores were negatively correlated with their scores on a measure of dogmatism (which represents a tendency toward obedience). In the years since it was created, the Need for Cognition Scale has been used in hundreds of studies and has been shown to be correlated with a wide variety of other variables, including the effectiveness of an advertisement, interest in politics, and juror decisions (Petty, Briñol, Loersch, & McCaslin, 2009)^[2].

Discriminant Validity

Discriminant validity, on the other hand, is the extent to which scores on a measure are not correlated with measures of variables that are conceptually distinct. For example, self-esteem is a general attitude toward the self that is fairly stable over time. It is not the same as someone’s mood, which is how good or bad one happens to be feeling right now. So people’s scores on a new measure of self-esteem should not be very highly correlated with their moods. If the new measure of self-esteem were highly correlated with a measure of mood, it could be argued that the new measure is not really measuring self-esteem; it is measuring mood instead.

When they created the Need for Cognition Scale, Cacioppo and Petty also provided evidence of discriminant validity by showing that people’s scores were not correlated with certain other variables. For example, they found only a weak correlation between people’s need for cognition and a measure of their cognitive style—the extent to which they tend to think analytically by breaking ideas into smaller parts or holistically in terms of “the big picture.” They also found no correlation between people’s need for cognition and measures of their test anxiety and their tendency to respond in socially desirable ways. All these low correlations provide evidence that the measure is reflecting a conceptually distinct construct.