Short answer: most Canadian supervisors will accept a Cronbach’s alpha of .70 or above for a research scale, and the published range of “acceptable” values runs from .70 to .95. But alpha is not a quality score. It rises automatically as you add items, it assumes every item measures one thing on the same scale, and an alpha above .90 usually means your questionnaire is repeating itself.
You ran the reliability analysis, you got .58, and you have three weeks. Below is what the number actually means, the four reasons it comes out low, what you can legitimately do about each one, and how to report the result so that an examiner reads it as competence.
What does Cronbach’s alpha actually measure?
Alpha is an index of internal consistency: the extent to which the items in a scale are inter-related. Tavakol and Dennick, in a widely cited short explanation of the coefficient published in the International Journal of Medical Education, define it as a number between 0 and 1 describing “the extent to which all the items in a test measure the same concept or construct.”
The psych package documentation for R, which is where a great many reliability analyses are actually computed, puts the mechanics more plainly still: “Alpha is the mean of all possible split half reliabilities (corrected for test length).” Imagine splitting your ten-item scale into two halves in every possible way, correlating the halves each time, and averaging the result. That is alpha.
The consequence students discover too late is that reliability is not a property of your instrument. It is a property of the scores your instrument produced in your sample. Tavakol and Dennick are explicit: “alpha is a property of the scores on a test from a specific sample of testees. Therefore investigators should not rely on published alpha estimates and should measure alpha each time the test is administered.” The .89 printed in the paper you took the scale from is evidence about that author’s respondents, not about yours.
And reliability caps validity without creating it: a scale can be perfectly consistent and consistently measure the wrong thing.
What number is acceptable?
The convention is real, and worth stating precisely rather than repeating as folklore.
Tavakol and Dennick report that “there are different reports about the acceptable values of alpha, ranging from 0.70 to 0.95,” and that at the top end “a maximum alpha value of 0.90 has been recommended.” So the defensible statement in a methods chapter is not “alpha should be above .7” but something closer to: the literature converges on .70 as a working floor for research scales, with values above .90 treated as a signal of item redundancy rather than as a better result.
| Alpha | Conventional reading | What it usually means in a thesis |
|---|---|---|
| Below .60 | Not acceptable as a summed scale | The items are not measuring one thing, or there are too few of them |
| .60 – .69 | Marginal; reportable with justification | Common for short subscales and newly adapted instruments |
| .70 – .79 | Acceptable for research use | The usual target for a thesis subscale |
| .80 – .89 | Good | Typical of a well-established multi-item scale |
| .90 and above | Check for redundancy | Often several items asking the same question in different words |
One caveat decides whether your committee accepts a .68. Thresholds are stricter when a score drives a decision about an individual — a clinical cut-off, a placement, a diagnosis — and looser when a score is one variable in a group-level model. A .68 on a four-item subscale used as a predictor is normal to report with a note. The same .68 on an instrument you propose to screen patients with is not.
Why is my alpha low?
There are four causes and they need four different responses. Diagnosing which one you have is the whole task, because three of them are fixable and one of them is not fixable by anything you do to the numbers.
Cause 1: too few items
This is the most common and the most misdiagnosed. Alpha is arithmetically bound to scale length. As the psych manual states, “Alpha and G6 are both positive functions of the number of items in a test as well as the average intercorrelation of the items in the test.” Tavakol and Dennick say the same from the other direction: “If the test length is too short, the value of alpha is reduced.”
A three-item subscale with well-correlated items can still land around .60. That is the mathematics of the coefficient, not a defect in your measure — which is why every alpha you report needs its item count beside it.
Cause 2: the scale is measuring more than one thing
This is the interesting failure, and the one that turns a problem into a finding. Alpha assumes what psychometricians call essential tau-equivalence: every item measures the same latent trait, on the same scale. When two constructs are hiding inside one set of items, that assumption breaks.
Tavakol and Dennick set out the consequence directly: “if a test has more than one concept or construct, it may not make sense to report alpha for the test as a whole as the larger number of questions will inevitable inflate the value of alpha. In principle therefore, alpha should be calculated for each of the concepts rather than for the entire test or scale.”
Note the direction of that argument, because it is counter-intuitive. A multidimensional scale does not reliably produce a low alpha — the extra items can push it up. The authors are blunt about the implication: “Internal consistency is a necessary but not sufficient condition for measuring homogeneity or unidimensionality,” and “a high coefficient alpha does not always mean a high degree of internal consistency.”
The diagnostic is a factor analysis, not a bigger alpha. If your factor solution shows two components, report two alphas.
Cause 3: one or two bad items
A reverse-coded item that nobody reverse-coded. A double-barrelled question. An item that everyone answered the same way, leaving almost no variance to correlate with anything.
The instrument for finding these is the item–total correlation column your software already prints. In R’s psych, the useful one is r.drop — “the correlation of the item with the scale composed of the remaining items.” SPSS calls the equivalent the corrected item–total correlation and prints “Cronbach’s Alpha if Item Deleted” beside it. An item whose corrected item–total correlation is near zero, or negative, is not contributing; a negative one is usually a coding error rather than a bad item, and it is worth checking the recode before deleting anything.
Cause 4: the sample is too homogeneous
Alpha depends on variance. If every respondent in your study is a second-year nursing student at one faculty, and they all answer a job-satisfaction item within a narrow band, the correlations between items shrink and alpha falls with them. Nothing is wrong with the items. The restriction is in your sampling frame, and the honest response is to say so in your limitations rather than to keep deleting items until the number improves.
Rule this cause out before the others, because it is the one you cannot fix at the analysis stage. It is also why a secondary analysis of a national survey often produces healthier coefficients than a convenience sample of eighty classmates: the sampling frame carries the variance. If your design could plausibly run on existing national files, our guide to getting Statistics Canada microdata for your thesis sets out which access tier gives you which variables.
What can I legitimately do about a low alpha?
In descending order of defensibility.
- Check your recoding first. Reverse-scored items that were never reversed are a frequent cause of an inexplicably low alpha, and they announce themselves as a negative item–total correlation. This costs five minutes and fixes the problem completely when it is the problem.
- Run a factor analysis and report subscale alphas. If the instrument has published subscales, you should have been doing this from the start. Reporting three subscale alphas of .81, .77 and .74 is a better result than one whole-scale alpha of .69, and it is more informative.
- Delete an item on evidence, and say that you did. Dropping the one item with a near-zero corrected item–total correlation is standard practice. Dropping four items in sequence until the number crosses .70 is not analysis; it is fitting your instrument to your desired output, and an examiner who asks how many items you removed and in what order will find it.
- Report McDonald’s omega alongside alpha. Because alpha assumes tau-equivalence, it is, in Tavakol and Dennick’s words, “a lower-bound estimate of reliability” whenever that assumption fails — which is most of the time. The model-based alternative is omega, and the
psychdocumentation states plainly that “ωt (see omega) is a better estimate of the reliability of the total test.” It is free to compute:psych::omega()in R, and both jamovi and JASP run on R underneath. Which package you are working in is a separate decision, covered in our comparison of SPSS, R, jamovi and Stata for a Canadian thesis. - Keep the low alpha and interpret it. Sometimes this is correct. A short subscale in a homogeneous sample produces a low coefficient for known reasons; naming those reasons is a stronger methods chapter than a laundered number.
Does translating a scale change its reliability?
Yes, and this matters more in Canada than in most places, because a great many Canadian graduate projects run in both official languages or adapt an English instrument for a francophone sample.
The published alpha attaches to the original instrument in its original language with its original respondents. A translated version is, psychometrically, a new instrument: you have to establish its reliability yourself, in your own sample, and report that value rather than the one in the source paper. This follows directly from the rule quoted above — investigators “should not rely on published alpha estimates.”
Two practical duties come with it. Get written permission from the copyright holder before translating a scale, because most published instruments are not in the public domain and permission is normally granted quickly when asked and never granted retroactively. And treat the translation as part of your methods: who translated it, whether it was back-translated, who reconciled the versions. If your study involves human participants at all, this sits alongside your ethics application rather than after it — see our guide to REB approval and TCPS 2 in Canada.
How do I report reliability in my thesis?
Five elements, in one or two sentences per scale:
- The coefficient you used and its value to two decimal places.
- The number of items it was computed on.
- The sample it was computed in — yours, with its n.
- Any item you removed, and the evidence for removing it.
- The published value from the source study, clearly labelled as the source study’s.
Internal consistency for the six-item engagement subscale was acceptable in the present sample (α = .78, n = 214). One item (item 4) was removed prior to analysis because its corrected item–total correlation was .04; the original nine-item scale reported α = .86 (Author, 2021). McDonald’s ω for the retained items was .80.
That sentence reports the number, its basis, the intervention, the evidence for it, and the comparison — leaving no hostile question it has not already answered. The reference formatting for the source instrument follows the ordinary rules, which our complete APA 7 guide for a Canadian thesis covers, including the Canadian source types the manual itself does not.
Once the coefficients are settled, the results chapter still has to be written so that a reader can follow it. Draft your results chapter in Tesify and keep tables, references and structure consistent while every interpretation stays 100% written by you.
Frequently asked questions
What is an acceptable Cronbach’s alpha?
The published range of acceptable values runs from .70 to .95, with .70 the usual working floor for a research scale. Values above .90 are conventionally read as a sign of redundant items rather than as a better result; a maximum of .90 has been recommended.
Is .65 too low to report?
Not automatically. On a short subscale, in a homogeneous sample, .65 is a normal value with a known explanation. Report it, report the number of items, and explain it in your limitations rather than deleting items until it moves.
Why did my alpha go up when I added items?
Because it is designed to. Alpha is a positive function of both the number of items and their average intercorrelation, so lengthening a scale raises the coefficient regardless of whether the scale improved.
Can Cronbach’s alpha be too high?
Yes. An alpha above .90 typically indicates that several items are asking the same question in different words, which wastes respondent effort and adds no information.
Does a high alpha mean my scale is unidimensional?
No. Internal consistency is a necessary but not sufficient condition for unidimensionality, and a multidimensional test does not necessarily have a lower alpha than a unidimensional one. Use factor analysis to answer that question.
Should I report one alpha or one per subscale?
One per subscale where the instrument has them. Where a scale carries more than one construct, computing alpha for each concept rather than for the whole instrument is the recommended practice.
What is the difference between raw and standardised alpha in SPSS?
Raw alpha is computed from item variances and covariances; standardised alpha is computed from the inter-item correlations, so the two differ when your items are on different scales or have very different variances. If standardised alpha is noticeably higher, that is a cue to look at whether the tau-equivalence assumption holds.
What is McDonald’s omega and should I use it?
Omega is a model-based reliability estimate that does not assume every item relates identically to the underlying construct. Because alpha is a lower-bound estimate when that assumption fails, omega is generally the better estimate of total-scale reliability, and it is available free in R’s psych package and in the R-based interfaces built on it.
Can I use the alpha reported in the paper I took the scale from?
No. Reliability is a property of scores in a particular sample, so you must compute and report your own. Cite theirs as a comparison, clearly labelled.
My scale has reverse-worded items. Do I do anything special?
Recode them before computing alpha. An un-reversed item will show a negative corrected item–total correlation and can drag a healthy scale below .50 on its own, which is why the recode is the first thing to check.