How to Report an Item Analysis in APA 7

You run the reliability analysis on your 12-item scale, look at the output and find three columns nobody ever explained properly: a mean per item, a corrected item-total correlation and an alpha if item deleted. One item has its correlation in the negative, another has a mean of 4.62 out of 5, and alpha rises by one hundredth if you drop either one. The question is not which numbers to copy, it is which of them mean anything and which of those two items has a real problem. Here are the three columns, what you do with each and how the paragraph gets written. If what you were after was the coefficient for the whole scale, that lives in how to report Cronbach's alpha in APA 7.

What numbers an item analysis carries

Four per item and one for the scale. Per item: difficulty (the mean and standard deviation, or the proportion correct if the item has a right answer), discrimination (the corrected item-total correlation), alpha if that item is deleted and, where the scale calls for it, the percentage of responses in the extreme categories. For the scale: total alpha with its confidence interval and the number of items. Beyond five or six items that is a table, not a paragraph; APA 7 does not force you to include one, but if your paper validates an instrument a reviewer will ask for it.

Symbol formatting is where half the manuscripts come apart. M and SD are italicized and they do take a leading zero, because they can exceed 1: M = 3.24, SD = 0.97. The item-total correlation is an italic r and it does not: .48, never 0.48. Alpha is a Greek letter, so it stays upright, unitalicised and without a leading zero: α = .89. And classical difficulty, the proportion correct, is written with a p that collides head-on with the p of a p value: define the symbol in the table note or spell out “proportion correct”, which is what most journals now do.

What an item analysis does not carry is p values. Nobody tests whether a corrected item-total correlation differs from zero, not least because with N = 312 anything above .11 comes out significant and that tells you nothing at all about whether the item is any good. What you read is the magnitude, and what you compare is each item against the others in its own scale. Do state the N each statistic was computed on and what you did with missing data, because complete cases and pairwise deletion do not give you the same table.

Where it goes in the manuscript depends on what the paper is for. If your study uses an already-validated scale and you only want to show it behaved well in your sample, this belongs in the Method, inside the instruments section, summarized in two sentences. If your paper is the validation, it belongs in the Results with the full table and with every deletion explained. Dropping a 40-item table into the Method of a study about something else is the fast route to being asked to move it into the supplementary material.

Difficulty: the item mean, and who ran out of variance

In an item with a correct answer, difficulty is the proportion of people who get it right, and the name misleads: a high value means an easy item. On a 20-item statistics test taken by 214 students, mean difficulty was .68 and item 12 came out at .96, which is to say almost everybody got it. An item like that does not discriminate and cannot: its discrimination index was D = .04, the difference between the top 27% and the bottom 27% of the distribution. The convention from Ebel and Frisbie (1991) is that D ≥ .40 is good, .30 to .39 acceptable, and anything below .20 gets revised. Maximum theoretical discrimination sits around a difficulty of .50, and rises to .625 on four-option multiple choice once you allow for guessing.

On a Likert scale there is no correct answer, so difficulty is the item mean, and it is read the same way: where the item sits on the continuum. If you want the number on the same 0–1 scale as classical difficulty, rescale it with (M − minimum) / (maximum − minimum). Item 9 of our statistics anxiety scale, with M = 4.62 and SD = 0.58 on a 1 to 5 format, has a rescaled difficulty of .91: nearly everyone marks it at the top. Seventy-one per cent picked the highest category. That percentage in the extreme categories is as informative as the mean and fits in a column, so put it there.

An item with a ceiling or floor effect cannot discriminate because it has no variance left to correlate. Item 9 took the lowest corrected item-total correlation in the scale, .19, and not because it measures something else: because 71% gave the same answer and there is nothing left to rank. Before accusing an item of measuring badly, look at its standard deviation: if it is 0.58 across a four-point range, the problem is distributional. The item analysis calculator returns the mean, the standard deviation, the corrected correlation and alpha-if-deleted in one table, which is exactly the comparison you need to make.

Difficulty is neither good nor bad on its own; it depends on what the instrument is for. A depression screening scale in the general population is full of “hard” items — few people endorse them — and that is correct, because severe symptoms are rare. A classroom climate questionnaire wants items spread across the whole continuum. What must not happen is that all 12 of your items bunch into the same zone: that gives you a scale which separates people around the average and distinguishes nothing at the extremes, which is usually exactly where the cases you care about are.

Discrimination: the corrected item-total correlation

“Corrected” means the item is taken out of the total before being correlated with it. Uncorrected, you are correlating an item with a sum that contains it, and that correlation is inflated by construction: on a five-item scale each item is 20% of its own total, and the values rise so far they stop telling a good item from a mediocre one. SPSS labels it “Corrected Item-Total Correlation” and R returns it as r.drop in psych::alpha(). The uncorrected one is the raw.r column of that same output, and it is not what you report; watch out for r.cor too, which is not the raw one but a different correction, for item overlap and scale reliability. If every value in your table for a short scale sits above .70, check which column you copied.

The threshold always quoted is .30, and it is quoted as a convention rather than a law: items below .30 are candidates for revision (Nunnally and Bernstein, 1994). On short scales or deliberately heterogeneous constructs people drop to .20 without scandalizing anyone, and on a six-item scale measuring something tightly defined a .35 is weak. Clark and Watson (1995) suggest also watching the mean inter-item correlation and keeping it between .15 and .50: below that the scale is not a scale, above it the items are paraphrases of one another. That second number is what catches redundancy, which the item-total correlation cannot see.

The corrected item-total correlation assumes there is one meaningful total. If your scale has three factors and you compute every item against the global score, the items on the small factor will come out low by construction and you will delete them for a problem that does not exist. Each item is correlated with the total of its own subscale, which means you have to know how many dimensions there are first, and that is the job of exploratory factor analysis or of a confirmatory model, not of the reliability table. Item analysis comes after the structure, never before it.

One detail that saves arguments with reviewers: the corrected item-total correlation is a Pearson correlation between an ordinal five-category variable and a continuous sum. It behaves reasonably with five or more categories and distributions that are not badly skewed, and it degrades when the item is dichotomous or heavily skewed. With binary items the right statistic is the point-biserial correlation, r_pb, which is mathematically the same Pearson but read against different benchmarks, or you move to item response modeling outright if the instrument deserves it.

A negative correlation is almost always an unrecoded reverse item

This is the first thing to look at and the thing most often reported wrongly. In our 12-item scale with N = 312, item 7 came out with a corrected item-total correlation of −.31 while the rest ranged from .19 to .62. Item 7 is “I feel calm when I work with data”, the one positively worded item in an anxiety scale. It is not measuring something else: it measures the same thing backwards and nobody turned it round. Check whether the items with negative correlations are exactly the reverse-worded ones in your questionnaire, and the investigation almost always ends right there.

The arithmetic check takes two seconds and needs nothing rerun. When you recode an item, its new mean has to equal (maximum + minimum) − the old mean. On a 1 to 5 format, item 7 had M = 2.41, so after recoding it has to come out at exactly 3.59. If you get anything else, you have recoded it wrongly, or recoded it twice — the classic, because the second pass returns you to the starting point and everything looks normal except the correlation, which is negative again — or you have written the result over the original variable and can no longer check. Always recode into a new variable with a different name: item7r, not item7.

With item 7 recoded, its corrected item-total correlation went from −.31 to .46 and the scale alpha rose from .82 to .89. Notice what would have happened had you obeyed the column beside it: alpha if item 7 deleted was .88, so the output was telling you to drop it, and dropping it would have left you at .88 with eleven items, worse than .89 with twelve and without the content that item contributed. The alpha-if-deleted column cannot tell a bad item from a badly coded one. You can.

If the correlation stays negative after recoding, then it is time to think. It can be a translation error that flipped the sense of the sentence, an item with a double negative that people read backwards, or an item that means something different in your population from what it meant in the original sample. There is also a documented phenomenon worth not mistaking for a finding: reverse-worded items, even properly recoded, tend to correlate with each other more than they should and to form a factor of their own that is a method artifact rather than content (Marsh, 1996). If your factor analysis throws up a second factor and it turns out to be made of exactly the reverse-worded items, you have not found a new dimension, you have found the wording.

Alpha if item deleted, and why you should rarely obey it

That column tells you what the scale's alpha would be if you erased that item and kept the rest. In a healthy scale every value in the column sits below the total alpha: each item contributes something, and removing any of them makes things worse. When a value sits above it, that item is subtracting, and that is where people start deleting. The trouble is that the column does not say why it subtracts. Our item 9, the ceiling one, has an alpha-if-deleted of .90 against the scale's .89. One hundredth. Deleting it for that is trading content for decoration.

There is a mathematical reason to distrust that column: alpha grows with the number of items by construction, on the same logic as the Spearman-Brown formula. A 20-item mediocre scale can post a higher alpha than a 6-item excellent one, so pruning to raise the coefficient works against itself unless the item is genuinely harmful. If you want to see how far yours moves, the Cronbach's alpha calculator gives you the coefficient with its confidence interval, and that interval is usually the answer: if the 95% CI around .89 is [.87, .91], the difference from .90 does not exist.

The attenuation paradox has been described since Loevinger (1954) and has not aged: past a certain point, raising internal consistency lowers validity. It happens because the most efficient way to inflate alpha is to add items that paraphrase one another, and a scale with eight ways of asking the same question posts a splendid alpha and a ridiculous coverage of the construct. The other side of it is that alpha assumes tau-equivalence, meaning every item carries the same weight, and that almost never holds; when it does not, McDonald's omega is the honest estimate, and the difference between the two coefficients is worked through in alpha, omega and the ICC.

Most important of all: alpha if item deleted describes your sample, not the instrument. With N = 312 that column shifts by several hundredths from one sample to the next with nothing real changing. Deleting items in pursuit of the maximum in the same sample where you then report reliability is capitalizing on chance: the alpha you publish is optimized for your data and will not replicate. If you are going to trim the scale, trim it in one subsample and confirm the final version in the other, and say so.

When an item genuinely gets deleted, and how you write it up

The working rule is short: you delete an item when you can say in words what is wrong with it, and that sentence cannot be “it raised alpha”. Four reasons survive review. One, the item is badly built and you can see it by reading it: double-barrelled (“I find it useful and enjoyable”), ambiguous, one negative too many, a translation slip. Two, it has no usable variance: everybody answers the same. Three, the corrected item-total correlation is still negative or near zero after the coding problem has been ruled out. Four, in the factor analysis it does not load where it should, or it loads in two places at once. In all four the criterion is fixed before you look at the results, and applied whatever comes out.

Deleting an item carries a cost that appears in no column: the scale stops being the published scale. Your scores are no longer comparable with studies that used the full version, published norms do not apply, and the validity evidence for the original instrument does not automatically transfer to your eleven-item version. That gets said in the manuscript in those words, in the limitations, and it is not disguised by calling your version by the original's name. If the scale has subscales, check as well that you have not gutted one: removing two of four items from a factor does not shrink the scale, it deletes it.

The whole paragraph, with the before and the after, is written like this: “All 12 items were examined in the full sample (N = 312). Item 7, which is reverse-worded, initially showed a negative corrected item-total correlation (−.31); after recoding it rose to .46, and all subsequent analyses used the recoded version. Corrected item-total correlations ranged from .19 to .62 (Mdn = .48) and item means from 2.14 to 4.62 (SDs from 0.58 to 1.24). Item 9 showed a ceiling effect (M = 4.62, SD = 0.58; 71% responded in the highest category) and the lowest discrimination (.19); it was retained because deleting it raised alpha only from .89 to .90 while reducing coverage of the somatic content of the construct. Internal consistency of the 12-item version was α = .89, 95% CI [.87, .91].” The table's numbers are not repeated in the text: you summarize the range, the median and whatever is exceptional.

Five errors send this analysis back, and none of them is arithmetic. Reporting the uncorrected item-total correlation as if it were the corrected one, which shows up because every value is suspiciously high. Deleting an item over one hundredth of alpha. Correlating every item against the global total in a multidimensional scale. Accepting a reverse item because “it looked fine once recoded” without checking that the mean matches (maximum + minimum) − the original mean. And the costliest of the lot: trimming the scale and then reporting the trimmed version's alpha as if it belonged to the original instrument, with nothing anywhere saying that items were removed or which ones.

Already drafted it? Paste your Results section into Review one section before it goes to your supervisor or the journal. You get the first thing a reviewer would object to, with your own sentence quoted.

Frequently asked questions

How do you report an item analysis in APA 7?

With a table carrying, per item, the mean and standard deviation, the corrected item-total correlation and alpha if that item is deleted, plus a note defining the symbols and stating the N. The text summarizes the range and median of the correlations, the items that stand out and the scale alpha with its confidence interval: “corrected item-total correlations ranged from .19 to .62 (Mdn = .48) and internal consistency was α = .89, 95% CI [.87, .91]”. M and SD take a leading zero; correlations and alpha do not.

What is an acceptable corrected item-total correlation?

The most quoted convention is .30: below that the item is a candidate for revision (Nunnally and Bernstein, 1994). It is a convention rather than a mathematically derived threshold, and it is applied with judgment: on short scales or heterogeneous constructs people drop to .20, and on a six-item scale measuring something tightly defined a .35 is already weak. Watch the mean inter-item correlation as well and keep it between .15 and .50 (Clark and Watson, 1995), because very high values mean your items are paraphrases of one another.

Why is my item-total correlation negative?

Almost always because it is a reverse-worded item that was never recoded. Check whether the items with negative correlations are exactly the reverse-worded ones in the questionnaire, and recode them into a new variable. The check is arithmetic: the recoded mean has to equal (maximum + minimum) − the original mean, so an item with M = 2.41 on a 1 to 5 format has to end up at 3.59. If it does not, you recoded it twice or wrote over the original variable. If the correlation is still negative afterwards, then you do have a content problem: a double negative, a translation error, or an item that means something else in your population.

Should I delete an item if alpha goes up when I remove it?

Not on that basis alone. Going from .89 to .90 by dropping an item is not an improvement: it is one hundredth, sitting inside the coefficient's own confidence interval. Alpha grows with the number of items by construction, so pruning to inflate it works against you, and taken to the limit it produces scales with superb consistency and painfully narrow content, the attenuation paradox Loevinger (1954) described. Delete an item when you can say what is wrong with it — bad wording, no variance, loading in the wrong place — and then the rise in alpha is a consequence rather than the reason.

What is item difficulty on a Likert scale?

The item mean, which tells you where the item sits on the response continuum. Classical difficulty, the proportion correct, only makes sense for items with a right answer; on Likert the equivalent is the mean, and if you want it on a 0–1 scale you rescale with (M − minimum) / (maximum − minimum). An item with M = 4.62 and SD = 0.58 on a 1 to 5 format gives .91, a ceiling effect, and its poor discrimination is a variance problem rather than a content one. Always report it alongside the percentage of responses in the extreme categories.

What is the difference between corrected and uncorrected item-total correlations?

The corrected one removes the item from the total before correlating; the uncorrected one leaves it in, so the item is partly correlating with itself and the value comes out inflated. On a five-item scale each item is 20% of its own total and the gap between the two versions is large. The one you report is the corrected one, which SPSS calls “Corrected Item-Total Correlation” and psych::alpha() returns as r.drop. If every value in your table for a short scale exceeds .70, you have very probably copied the wrong column.

With each item's mean and standard deviation, the corrected item-total correlation, alpha-if-deleted and a sentence explaining every deletion, your item analysis is reported the way APA 7 asks; and if any item comes out negative, start with the coding rather than the theory. When the hard decision is not the format but how many dimensions the scale has and which items survive validation, that is what my statistical consulting is for.