Split-Half Reliability

Split-half reliability separates the items of a scale into two sets, scores each set for every participant, and asks how closely those two scores travel together. The reasoning resembles test-retest reliability without requiring a second administration: the two halves stand in for parallel forms of one another. Because each half is only half the length of the instrument, the raw correlation understates the reliability of the full scale, and the Spearman-Brown correction compensates for that loss of length, which is why the corrected figure is the one to report. The Guttman coefficient answers the same question along a different route, working from the variances of the half scores, and it can depart from the corrected value when the two halves differ in variance. YouReply Analyze computes the association between the half scores with the pearsonr function from scipy, requiring at least four items and at least ten complete cases. How the halves are formed is a parameter: the default pairs odd-numbered items against even-numbered ones, a second option cuts the list into a first half and a last half, and a third assigns items at random, with a seed available so a random split can be reproduced. The output names the items in each half explicitly, because the coefficient depends on exactly that division.

• Do the two halves of a ten-item knowledge test yield consistent results?
• Do the odd-numbered and even-numbered items of a scale measure the same construct with equal strength?
• Can an internal consistency figure be corroborated by a second route that does not rest on alpha?
• How much does the reliability estimate move when the halves are formed in different ways?

• When a second line of evidence for internal consistency, independent of alpha, is wanted.
• When the items were administered in a single session and a second administration is out of reach.
• When the scale has four or more items, since fewer leaves each half too short to produce a meaningful score.
• When item order carries meaning and consistency between the opening and closing parts of the instrument is itself the question.

• Items: at least four numeric columns measuring the same construct. An even number of items makes it easier for the halves to come out the same length.
• All items must be coded in the same direction, since a reverse-worded item left untouched corrupts the score of whichever half it lands in.
• No grouping or predictor column enters; the method works on a single set of items.
• One row per participant, with at least ten complete cases required for the computation.

The halves can pass as parallel forms

The method takes the two halves to measure the same construct with equal strength and to have comparable score variances. Where the split produces sets that differ markedly in difficulty or content, the coefficient reflects a poor division rather than the reliability of the scale.

One common construct

The coefficient is interpreted on the premise that the items tap a single construct. Items drawn from two constructs can still produce a high value when they happen to be spread evenly across the halves.

Consistent item direction

A reverse-worded item left unrecoded drags down the score of its half and weakens the association between halves, and the size of the damage depends on which half it happens to fall in.

A linear relationship between the half scores

The association measure is the Pearson correlation, so the link between the two half scores is assumed linear and assumed not to be driven by extreme cases.

Enough observations

Since the estimate rests on a correlation, it swings widely in small samples, and both half scores have to be complete for a case to count.

• The halves can pass as parallel forms: Whether the halves are parallel goes untested, and no check comparing their variances is returned. The one indirect signal available is the gap between the Spearman-Brown corrected value and the Guttman coefficient: a clear divergence means the half variances are not equal.
• One common construct: This method extracts no factors and never inspects how many dimensions the scale has. Establishing the structure of the item set calls for a separate run of exploratory factor analysis.
• Consistent item direction: The engine never reverses an item for you and issues no warning about it. Recode on the Variable tab with derived columns and run the analysis again.
• A linear relationship between the half scores: The distribution of the half scores is not examined, no outlier screening is performed and no scatter plot is returned. In a test with a ceiling effect the half scores compress and the correlation falls, and spotting that means looking at the item distributions with the descriptive tools.
• Enough observations: The engine requires at least ten complete cases and halts below that. Ten is the computational floor rather than a standard of adequacy, and rows with a missing answer in either half drop out of the calculation.

1. 1Upload the file of item responses, laid out with one column per item and one row per participant.
2. 2Recode the reverse-worded items on the Variable tab and declare unanswered cells as missing, since the computation runs on complete cases.
3. 3Choose the method from the reliability category on the Analysis tab and confirm in the data requirements box that the four-item condition is met.
4. 4Tick the item columns in the parameter form, then pick how the halves are formed: odd-even comes as the default, with first-last and random also available.
5. 5If you choose the random split, fill in the seed, because the same seed reproduces the same division and therefore the same coefficient.
6. 6Run it and read which items fell into which half before you look at the Spearman-Brown corrected value. Tables export to Excel, and rerunning with a different split lets you compare the stability of the estimate from the History tab.

Correlation between halves with its p value

The Pearson association between the two half scores and the p value for it differing from zero. This raw figure is the reliability of two half-length forms rather than of the whole scale.

Spearman-Brown corrected coefficient

The raw correlation adjusted for scale length, and the figure this method exists to report. It compensates for the understatement caused by halving, so it always exceeds the raw value.

Guttman split-half coefficient

A second estimate derived from the variances of the half scores and of the total. Its distance from the corrected coefficient hints at how far the two half variances diverge.

Interpretation label

The label attached to the corrected coefficient: excellent at 0.90 and above, good from 0.80, acceptable from 0.70, questionable from 0.60, poor from 0.50 and unacceptable below. It is tied to the corrected value alone.

Items in each half

An explicit list of which items went into the first half and which into the second. Since the coefficient depends on that division, the list is part of what has to be reported.

Split method used

Whether the odd-even, first-last or random option was applied. Reproducing the result depends on this being stated, together with the seed where the split was random.

Number of items and observations

How many items entered the scale and how many complete cases the computation used. When the second figure falls short of your row count, look at how the unanswered items are distributed across the halves.

Spearman-Brown corrected coefficient

This is a reliability estimate rather than a significance test, and its interpretation follows disciplinary convention: 0.70 is a common floor for group-level comparisons, while measurement feeding decisions about individuals is expected to reach 0.90. Bear in mind that the value depends on how the halves were drawn, so presenting one run's number as a fixed property of the instrument misrepresents it.

Raw correlation between halves

It expresses the agreement between two half-length forms and sits systematically below the reliability of the full scale. The gap to the corrected value shows what length contributes to reliability: a raw figure near 0.65 rises to around 0.79 once corrected.

None of the coefficients comes with an interval: no 95% bounds are returned for the raw correlation, the Spearman-Brown corrected value or the Guttman coefficient. The output carries only the p value of the raw correlation, and that speaks to the correlation differing from zero rather than to the precision of the estimate. The gap has two layers in this method, since neither the sampling variability nor the movement caused by the choice of split is quantified anywhere. The practical response is to rerun the analysis with different splits and inspect the spread of the values yourself, which is no substitute for an interval but does give a concrete sense of how stable the estimate is.

The numbers below are a representative example, not data from a real study or a real user.

Research question

In an illustrative education study, is a twelve-item reading comprehension test usable on the evidence of its split-half reliability?

Variables

Items: twelve reading comprehension items scored 1 for correct and 0 for incorrect · Split method: odd-numbered against even-numbered items · Cases: 188 complete responses

Example result

The correlation between halves was 0.71 (p < .001), the Spearman-Brown corrected coefficient 0.83 (label: good) and the Guttman split-half coefficient 0.82. The first half consisted of the six odd-numbered items and the second of the six even-numbered ones. Items = 12, n = 188. Splitting the same items into a first half and a last half instead brought the corrected coefficient down to 0.74.

Interpretation

At 0.83 the odd-even split places the test in the good range and supports its use for group-level comparison. The corrected and Guttman coefficients sit close together, suggesting the half variances are well balanced. What stands out is the second split: cutting the items into the first six and the last six drops the coefficient to 0.74. A fall like that usually signals an order effect, with later items perhaps harder or respondents tiring towards the end. For such a test the odd-even split gives the fairer estimate and is the value to report, though the write-up has to name which split produced it. The distance between 0.74 and 0.83 across the two runs makes tangible how far an estimate can move in a measure for which the panel returns no interval.

Research question: Do the odd and even numbered items of the first construct measure the same thing?

Scale items

item_1,item_2,item_3,item_4,item_5,item_6,item_7,item_8,item_9,item_10

Split method

odd_even

Computation credits: scipy 1.18.0 · statsmodels 0.14.6 · scikit-learn 1.9.0 · numpy 2.5.1 · pandas 3.0.5 · semopy 2.3.11 · 89fc29a · Data seed: 20260915

Split-half reliability of the twelve-item reading comprehension test was adequate under an odd-even split (correlation between halves r = .71; Spearman-Brown = .83; Guttman = .82; n = 188).

An example sentence close to APA style; the numbers are representative.

• No confidence interval accompanies any of the coefficients, leaving the sampling variability of the estimate unknown.
• The result hinges on how the halves are drawn: the same instrument yields different coefficients under different splits, and the engine never averages across all possible divisions.
• Scales with fewer than four items cannot be analysed, and with an odd number of items the halves are unequal in length.
• Whether the halves are parallel goes untested, so a division unbalanced in content or difficulty depresses the coefficient for reasons having nothing to do with the scale.
• Dimensionality is never checked, because the method extracts no factors, and items from two constructs spread evenly across the halves can still produce a high value.
• Reverse-coded items are not recoded automatically, and the damage an untouched one does depends on which half receives it.

• Cronbach's AlphaWhy: When an internal consistency measure that does not hinge on a particular division is wanted, since alpha is conceptually the average over all possible splits.
• McDonald's OmegaWhy: When the items are thought to carry the construct with differing weights, it avoids the equal-loading premise and estimates reliability with less bias.
• Exploratory Factor Analysis (EFA)Why: When the number of dimensions in the item set is uncertain, the structure has to be established before halves are formed at all.
• Inter-Rater AgreementWhy: When the reliability question concerns the people judging the same material rather than the items of a scale.

• Cronbach's Alpha
• McDonald's Omega
• Exploratory Factor Analysis (EFA)

Why does the coefficient change when I change the split?

Because the figure is the product of one particular division rather than a fixed property of the instrument. A ten-item scale can be cut into two equal halves in a hundred and twenty-six distinct ways, each yielding a slightly different number, and the engine computes only the one you asked for rather than averaging over all of them. How much movement you see depends on how homogeneous the items are: closely similar items produce similar values under different splits, while items that differ in difficulty or content open the spread wide. Two things follow. State the split you used, with the seed if it was random. And run a few different splits to see the spread, because a wide one argues for moving to a split-independent measure such as alpha or omega.

Which split method should I choose?

Odd-even is the right default in most situations, since it alternates items along their order and thereby spreads rising difficulty and growing fatigue evenly between the halves. Reach for the first-last split only when you are asking a specific question, such as how consistently the opening and closing sections of an instrument behave. That split lowers the coefficient noticeably whenever an order effect is present, and the drop itself is an informative finding. The random split earns its place when you want to probe how much the division matters: several runs with different seeds, compared for spread, give the best available picture of stability in the absence of an interval.

Should I report the raw correlation or the corrected coefficient?

The Spearman-Brown corrected value is the one to report, because what interests you is the reliability of the full instrument rather than of a half-length form. The raw correlation is systematically lower, and passing it off as reliability makes the scale look weaker than it is. Giving both figures is nonetheless good practice: the reader sees how much the correction contributed and the arithmetic stays transparent. Reporting the Guttman coefficient helps too, particularly when it diverges clearly from the corrected value, since that divergence points to unequal variances between the halves and warns the reader about how parallel they really are.

If I can compute alpha, what does this method add?

It earns its place in three situations. The first is corroboration: alpha and the split-half coefficient come from the same data by different routes, and agreement between them strengthens confidence in either. The second is detecting order effects, because a first-last split reveals something alpha cannot show at all, namely inconsistency between the beginning and the end of an instrument. The third is convention, since the literature on knowledge and attainment testing has long reported this coefficient and comparability argues for continuing to do so. Relying on it alone as the primary reliability evidence in a scale development study is unwise, though, given that the value depends on the split and arrives without an interval.

• Nunnally, J. C., & Bernstein, I. H. (1994). Psychometric Theory
• Spearman, C. (1910). Correlation Calculated from Faulty Data
• Field, A. (2018). Discovering Statistics Using IBM SPSS Statistics
• scipy.stats.pearsonr documentation