StatReason

Split-Half Reliability Calculator

Estimate how internally consistent a multi-item scale is by splitting its items into odd and even halves, correlating the two half-totals, and correcting for the fact that each half is only half as long as the full scale.

One subject per line, comma-separated item scores. Items are split by column position: the 1st, 3rd, 5th, and so on go to the odd half; the 2nd, 4th, 6th, and so on go to the even half.

What this answers

This calculator answers "if I cut this scale into two halves, how consistently do those two halves agree with each other?" That agreement is a reliability estimate, standing in for how consistent the scale would be with itself if you could somehow re-measure the same subjects again.

How it is calculated

Items are assigned to an odd half or an even half by column position (a deterministic, reproducible split, not a random one), each subject's odd-item and even-item scores are summed separately, and the Pearson correlation between those two half-total columns is the raw half-correlation. Because each half is only half the length of the full scale, and longer scales are generally more reliable, that raw correlation understates the reliability of the full-length instrument. The Spearman-Brown formula, 2 times r divided by 1 plus r, corrects for this by estimating what the correlation would be at full test length.

Worked example

For a 4-item scale, items 1 and 3 form the odd half and items 2 and 4 form the even half. Each subject's two half-totals are correlated across all subjects, giving a raw correlation, then the Spearman-Brown correction raises that number toward what the full, uncut scale's reliability would be. A raw half-correlation of .60, for instance, corrects upward to .75 after applying the Spearman-Brown formula.

Assumption audit

Calculated from your data: the number of subjects and items, the resulting odd and even item counts, and whether either half's total has zero variance, which would make the correlation undefined.
Evidence to review: whether the odd-even split happens to separate items in a meaningful, unintended way, for example, if every reverse-worded item lands on the same half. A different split of the same items can give a noticeably different estimate.
You must verify: that every item is intended to measure the same underlying construct and that any reverse-keyed items were already re-scored before being entered here; this calculator does not detect or correct reverse-keyed items on its own.

What this result does not mean

A high split-half reliability does not prove the scale measures the right construct, only that its two halves are internally consistent with one another. It also does not represent every possible way the items could be split; one particular odd-even split is not the same as averaging across all possible splits, which is closer to what Cronbach's alpha approximates directly.

Limitations

This calculator uses only the fixed odd-even split, not a user-selectable custom split. Different splits of the same data can give somewhat different raw correlations, so a single split-half estimate should be read as one reasonable estimate, not the only possible answer; Cronbach's alpha, which effectively averages over many possible splits, is often reported alongside or instead of any single split-half result.