Draft item creation

The initial step in creating draft items involves understanding the concept thoroughly through reading related literature, talking to experts, sharpening one's own thinking and observation. Attention must also be paid to whether it is a uniform concept or if there are sub-aspects to it (Irvine Reference Irvine and Kyllonen2002). For illustrative purposes, let us take the example of anxiety. Anxiety involves not only the feeling aspects of being fearful, feeling discouraged and anxious, but also cognitive aspects such as constantly thinking of failures, problems and negative consequences, as well as physiological aspects such as tremors, palpitations, feeling faint, cold clammy hands, dry mouth and upset stomach. A pool of items representing all the sub-aspects of the concept must be prepared by the test constructor, ideally with a team of experts. Shortcomings at this stage with regard to the relevance and wording of the test items can contribute to the error variance. Error variance is variance due to random, irrelevant sources, as opposed to the true variance, which is variance due to true difference (Cohen Reference Cohen and Swerdlik2009).

Test format

The test format also has to be decided. Will the test be open-ended (also called free response or constructed response) or will it be closed (objective or forced choice)? Open-ended responses may involve writing samples (e.g. in an ability test), free oral responses (e.g. in projective tests) or even performance of some kind of task. They are good for exploring concepts/situations/phenomena and provide a wider range of possibilities, richer samples of the individual's behaviour and allow for their unique characteristics to emerge. However, scoring is more complex and time-consuming, and the reliability and validity of such tests are lower than in closed-response (forced-choice) tests (Urbina Reference Urbina2004).

The forced-choice format presents a limited number of alternatives from which the respondent must choose. The choice can be on the basis of multiple choice, true/false statement, rating, ranking and matching, as well as rearrangement of the options provided. Well-known among them is the Likert scale, where the items are listed and the respondent expresses their degree of agreement or acceptance or frequency of occurrence by choosing one of an odd number (3, 5, 7 or even 9) of response options, with the midpoint usually signifying a neutral or middle-of-the-road position (Dawes Reference Dawes2008). The sum of the responses to all the items is taken as the score. The 5-point Likert scale is particularly popular, with response options such as ‘strongly agree’, ‘agree’, ‘neither agree or disagree’, ‘disagree’ and ‘strongly disagree’ (Derrick & White Reference Derrick and White2017). Seven or more response options are used if there is just a single item for measurement, as in the Net Promoter Score, a single-item customer loyalty measure where customers are asked how likely it is that they will recommend a company or product to others (Sauro Reference Sauro and Dumas2009).

The forced-choice format is the most frequently used option for test items because of the ease and speed of scoring. The objectivity of the responses makes it more reliable. It can be applied in individual or group testing. However, errors are possible if the respondent guesses in ability tests, or answers manipulatively or carelessly in personality tests. These can reduce the reliability and validity of the test. Also, preparing a forced-choice test requires thorough familiarity with the concept and specialised test development and item-writing skills. Most of this article focuses on this type of test construction.

Qualitative item analysis

The order, type and wording of the measurement items, the introduction, the instructions, the transitions and the closure/ending of a test ideally should encourage respondents to provide accurate and adequate information. They should also discourage them from discontinuation or refusing to answer specific questions. Test item wording also has to be sensitive to the target group. The following tips help meet these aims.

• • Use simple language, short sentences and terms that the participant understands (e.g. in the item ‘I dread panic attacks’, the respondent might not know what ‘panic attacks’ are).

• • Ensure that all possible response alternatives are provided for forced-choice (closed-response) items (e.g. in asking about ‘Frequency of panic attacks’, if the response options range from ‘once in a year’ to ‘several times a day’, someone who has never experienced a panic attack ever is forced to chose an option that does not apply to them).

• • Ensure that the wording of the item matches the response options (e.g. if the item reads ‘Do you feel fearful for no reason?’, the response options ‘strongly agree’, ‘agree’, ‘disagree’, ‘strongly disagree’ will not be suitable as they are meant for a statement and not a question).

• • Ensure that all the items are relevant and needed.

• • Avoid double-barrelled items (e.g. ‘Do you have digestive problems and shortness of breath?’ – if the patient has only one and not the other, how will they answer?).

• • Avoid leading, biased wording (e.g. ‘Anxiety is often accompanied by irritability; rate your level of irritability on a scale of 1 to 10’).

• • Avoid ambiguous words with broad or abstract meaning. For example, words such as ‘generally’, ‘on average’ and ‘regularly’ can mean different things to different people.

• • Avoid items that require effort and distant recall (e.g. ‘Did you experience anxiety attacks 5 years ago?’)

• • Avoid asking for ego-threatening, embarrassing or private information if it is not essential (e.g. income, moral stand, private habits).

A test maker should also be aware of ‘response set’ while creating the final form of the test. The response a person makes to a test is a function not just of the item content but also of the form of the items and the test direction. Going for speed rather than accuracy, tendency to opt for the neutral category, tendency to guess when in doubt, tendency to mark extreme categories, tendency to agree or like, tendency to respond desirably are all examples of an individual's response set when completing a test. A person's response set is the consistency in the manner of their response to the test's form rather than its specific content (Cronbach Reference Cronbach and Lee1950). For example, if all the items of the test refer to the anxiety symptoms such as trembling, fearfulness, thoughts of failure, this might induce a temporary preference to respond in a set way to all the items. To avoid this response set, some of the items could be worded in the reverse direction (‘reverse wording’), with corresponding reverse scoring. For example, ‘I am usually calm most of the time’ instead of ‘I am often fearful’.

Administering the draft version of the test to a small sample of 15–30 respondents, representative of the population the test is intended for, can help to ensure that the test items are understandable and answerable. Thus in this qualitative item analysis phase, the content coverage, wording and sentence structure of the item pool are fine tuned by the test constructor, then submitted to reviewers for their comments and for further revision, if any.

After the refinement at these three levels by the test constructor, using information from the pilot study and the reviewers’ comments, the test is ready for the important stage of quantitative item analysis, which is carried out on a much larger sample of 200+ respondents.

Quantitative item analysis

This involves a variety of statistical procedures used for the final selection of the items of a test on the basis of the responses obtained from the samples used in the process of test development. Two important concepts used for selecting items from the pool are their difficulty level and discriminative power.

Difficulty level is a measure of the proportion or percentage of respondents who answered the item correctly; for this reason it is frequently called the P-value (Anastasi Reference Anastasi1954; Urbina Reference Urbina2004). It can range between 0.0 and 1.0, where 1.0 indicates that 100% of examinees responded to the item correctly, indicating that it is an easy item. In criterion-referenced tests, where each respondent's performance is compared directly with the standard, without considering how others perform on it (e.g. classroom-achievement tests), the item difficulty is kept in the range 0.7–0.8, as it has to be within the reach of the majority of respondents. On the other hand, norm-referenced tests, where the purpose is usually to rank and compare respondents, are designed to be harder and to spread out the examinees’ scores. To achieve this purpose, an item difficulty index between 0.4 and 0.6 is chosen for such tests (Urbina Reference Urbina2004). Difficulty level is critical in ability tests, as it differentiates between people of high and low ability, but in personality tests, difficulty level is applicable only to ensure that the respondents are in a position to understand and respond as required.

Discriminative power refers to the extent to which items elicit responses that accurately differentiate test takers in terms of the behaviours, knowledge or other characteristics that the test is designed to evaluate (Urbina Reference Urbina2004). For every type of test, be it an ability test, an achievement test, an attitude or a personality test, this is one of the most important, as well as basic, qualities an item must possess: individuals high on what the test aims to measure should score high on the item and those low on that variable should score low on the item.

A simple way to compute the index of discrimination (D) using the classical test theory (CCT) approach is to arrange the respondents’ total scores (sum or average of all the items) for the test in descending order and classify the respondents into three distinct groups: those scoring the highest 27% of marks, those scoring lowest 27% and those in the middle. For each item, the percentage of respondents in the upper and lower groups who answer correctly or answer in the intended direction is calculated. The difference is one measure of item discrimination. The formula is:

(1)$$\eqalign{& D = \left( {{\rm upper group percentage}} \right) \cr & \,\,\,\,\,\,\,\,\,-{\mkern 1mu} \left( {{\rm lower group percentage}} \right).} $$

The possible range of the discrimination index is −1.0 to 1.0. A negative discrimination index may indicate that the item is measuring something other than what the rest of the test is measuring. More often, it is a sign that the item score has been wrongly entered. This can also happen due to carelessless or when the items are written in reverse direction.

Another method of determining discriminative power for questions with a right and wrong answer is the point-biserial correlation (PBC), which measures the correlation between the correct answer (viewed as 1 = right and 0 = wrong) on an item and the total test score of all students (Wright Reference Wright1992). The PBC is sometimes preferred because it identifies items that correctly discriminate between high- and low-scoring groups, as defined by the test as a whole, instead of the upper and lower 27% of a group.

Summary

Item analysis will thus help the test constructor to decide on the items to select for the final test by choosing those with levels of difficulty and discriminative power suited to their purpose.

Test–retest reliability

The most straightforward method is where the test is administered once and then a second time to the same or a similar group after a suitable gap (not too short that they remember the items, and not too long that respondents could have changed with respect to the variable being measured). This is called test–retest reliability, and it measures the temporal stability. The correlation of the scores between the two administrations will give the test–retest reliability. Scores above 0.70 indicate reasonable reliability.

Parallel (or alternate) forms reliability

To overcome the possible practice effect of the test–retest method, instead of the same test, a similar version of the test is administered to the same group. This is called parallel or alternate forms reliability (it is also called equivalence). A high correlation of above 0.70 between the two test scores indicates reliability.

Split-half reliability

A third type of reliability, which requires only a single test administration is split-half reliability, checks the reliability by splitting the test into two comparable halves and finding the correlation between the data for each half.

Internal consistency reliability

A single administration is also enough to determine internal consistency reliability, which measures how well each item measures the content or construct under consideration. The logic here is that different items that measure the same construct should display high relation with each other. There are a variety of internal consistency measures. Usually, they involve determining how highly these items are correlated and how well they predict each other. Cronbach's alpha is one commonly used measure if the variable is unidimensional, and it can be quickly calculated using statistical software.

Test length

It was believed that the longer a test is, the more reliable it is because more items reduce the error of measurement. Indeed, a sufficient number of items must be included to cover the content areas tested; however, the quality of items can contribute to how efficiently a test measures and separates respondent's ability. Thus, if the quality of the items is high, a shorter test can have higher reliability than a longer one (Urbina Reference Urbina2004).

Content validity

Content validity is the extent to which test items are relevant to and representative of the concept being measured (Urbina Reference Urbina2004). Psychological tests designed to aid in the diagnosis of psychiatric disorders often include, or may even be entirely composed of, items that reflect critical symptomatic aspects of the syndromes they are designed to diagnose. Here again, the relevance and representativeness of the items sampled by these instruments are of crucial importance in determining their usefulness for diagnostic purposes. One way to provide evidence for content validity is by using subject matter experts to review a measure for any construct deficiency or contamination. For example, a test of anxiety can be relevant if it adequately covers the emotional, cognitive and behavioural symptoms of anxiety and also does not inadvertantly cover symptoms of depression, which is another concept.

Criterion validity

Criterion validity uses various strategies that focus on the correlation between the score obtained from the test being matched on a known outcome indicator (i.e. criterion) for the construct being tested. Criterion validity can be manifested by concurrent validity and predictive validity. Concurrent validity refers to the extent to which the score (of the test being validated) is related to an outcome indicator score that is measured at the same time, whereas predictive validity refers to how well the score can predict a criterion obtained later (Hubley Reference Hubley, Zumbo, Geisinger, Bracken and Carlson2013; Messick Reference Messick1995). For example, the scores of a newly constructed test of intelligence are matched to students' current grades in class (concurrent) and are also matched to their final grade point average a year later.

Construct validity

Construct validity is ‘the degree to which a test measures what it claims, or purports, to be measuring’ (Cronbach Reference Cronbach and Meehl1955). In the classic model of test validity, construct validity is one of three main types of validity evidence, alongside content validity and criterion validity. Modern validity theory defines construct validity as the overarching concern of validity research, subsuming all other types of validity evidence (Messick Reference Messick1998). Key to construct validity are the theoretical ideas behind the concept under consideration. There is no single way to measure it: construct validity should be demonstrated from a number of perspectives, by an accumulation of evidence (Brown Reference Brown1996).

Convergent validity and discriminant validity are two subtypes of construct validity. Convergent validity represents the correlation between the score obtained from the target test (e.g. a newly developed scale for anxiety) and the score on a test for an existing highly related construct (e.g. emotional stability) or the score derived from another well-validated test that measures the same construct (e.g. another scale to measure anxiety). Discriminant validity indicates the relationship between the scores of anxiety and of an unrelated construct (e.g. intelligence).

Face validity

Face validity refers to the superficial appearance of what a test measures, from the perspective of a test taker or any other naive observer. Improving face validity does not improve the test's objective validity, but it is needed for cooperation and public relations – for test users, for those who select and decide on which tests to be used, for the observer, and for judicial and legislative purposes. For example, in measuring the numerical aptitude of engineers, having test items that deal with machines and tools might elicit more involvement than items about flowers and oranges.

Factor structure (factors to be retained)

Scree plots and eigenvalues are the two widely used indicators to determine the number of factors to be retained. A scree plot is a curve that shows the eigenvalues in a downward direction (Fig. 1). The number of factors to be retained is determined by referring to the left of the point where the ‘elbow’ of the graph seems to level off. Factors with eigenvalue greater than 1.0 should be retained. For instance, if the EFA shows five factors but only the first three factors’ eigenvalues exceed 1, then a three-factor, instead of five-factor, solution is recommended.

FIG 1 A scree plot shows the eigenvalues for a 12-item test. Factors sitting to the left of the point where the ‘elbow’ of the graph levels off are those to be retained. Here, three factors should be retained.

It is important to note that that the eigenvalue method is not without limitations. Researchers can fix the number of factors according to the underlying theory of the test.

Factor loading

After identifying the number of factors to be retained, the next step is to examine the loading pattern of the items. Loading shows the variance explained by the item on that particular factor. Theoretically speaking, items that are designed for a specific factor should load onto the target factor. Moreover, such items are expected to demonstrate (a) a high factor loading (e.g. ≥0.40) on the target factor and (b) lower factor loading on non-target factors. For example, in the anxiety scale, items that measure biological responses such as hand tremors and palpitations should have a high factor loading (e.g. ≥0.40) on the physiological response factor and lower factor loading on the emotional and cognitive factors.

If the factor-loading results are different from the expected, for example the factor loading of an item is consistently low on all factors, researchers can remove such item(s) and submit the remaining items to another EFA. The procedure can be repeated until all the items show satisfactory factor loading. When the results find more than one item that deserves deletion, it is advisable to remove them one item at a time (starting with the item with the lowest factor loading) and then to re-run EFA. The sequence of removing items does matter and should be reported. In addition, it is possible that an item may have acceptable factor loading on both target and non-target factors. Two solutions are suggested to deal with this cross-loading problem. Some researchers suggest removing such items to enhance clarity of the structure and ease interpretation of results; others argue that cross-loading is not uncommon and should be allowed, especially when it is theoretically adequate.

Explained variance

Finally, after identifying a satisfactory factor structure with acceptable factor loadings, it is critical to examine whether the structure has a high percentage of total explained variance, which is the part of the model's total variance that is explained by factors that are actually present (e.g. >50%). Higher percentages of explained variance indicate low discrepancy between a structure (or model) and actual data (i.e. low error variance) and hence better predictions can be made (Rosenthal Reference Rosenthal and Rosenthal2011).

Summary

Taken together, a EFA result consists of an interpretable factor structure, clear-cut factor loading and adequate explained total variance. Note, however, that EFA may generate three types of result: exactly identical to the structure of the original version of the test, slightly different or totally different from the original version. Regardless of the results, it is necessary to collect another set of data and further examine the qualities of the test using CFA.

CFA for translated tests

CFA is also conducted for tests that have been translated into another language to ensure that the translated version is true to the original (translation is discussed in more detail below). It is noteworthy that the best model of a translated test that is supported by both EFA and CFA could be the same or different from the structure of the parent version. When the same structure is found, it implies that translation has had very little impact, if any, on the structure of the test. Therefore, one can confidently believe that the same respondents would perceive items of the parent and translated versions equally. On the other hand, minor or major differences imply either that translation has distorted the meaning of the items or that the theoretical concept or construct of the test is cross-culturally different. In the latter case, researchers may want to examine the target population's perception of the construct assessed by the test using a qualitative approach. Alternatively, if possible, researchers may collect data using both the parent and the translated versions and evaluate the variance of the structure of the two tests to identify the sources (e.g. different structure, factor loading) of any differences.