MBTI Test-Retest Reliability: What 0.5 Actually Means (And Why Your Type Might Change On Retest)

Six things to know before reading further:

Test-retest reliability is a psychometric statistic that captures how consistent a measurement instrument is when applied to the same people across multiple sessions. Mathematically, it is the correlation between scores from session 1 and scores from session 2 (typically with a 1-5 week interval between sessions). The correlation is bounded between 0.0 (scores are random across sessions) and 1.0 (scores reproduce perfectly).

The interpretation: if you take a personality test twice within a few weeks, and the test has reliability 0.9, your two scores will be very close to each other. If the test has reliability 0.5, your two scores will share about half their variance — meaning they will correlate but with substantial drift. The drift is what produces the famous "50% of people get a different MBTI type on retest" finding: when categorical cutoffs are applied to dimension scores with reliability 0.5-0.6, the binary letter assignments flip easily for anyone whose dimension score is near the midpoint.

Reliability is independent of validity. A test can be highly reliable (reproduces itself) without being valid (measuring what it claims to measure); a test can be valid (measuring something real) without being highly reliable (the measurement is noisy). The two properties are separately important. MBTI's reliability problem is specifically about consistency across sessions — what your type code is depends partly on which session you happened to take the test in. The validity question (whether MBTI is actually measuring what it claims) is a separate axis.

Per psychometric convention, reliability above 0.7 is considered acceptable for research purposes; above 0.8 for clinical or selection contexts. MBTI's 0.5-0.6 falls below both thresholds. Big Five trait measures typically clear the 0.7 bar and often clear the 0.8 bar. The reliability gap between the two frameworks is the technical foundation of why one is preferred for measurement contexts.

If you took the MBTI yesterday and got INFP, and you take it again next week, what should you expect? At reliability 0.5-0.6, here is the rough breakdown:

**Most likely outcome (about 50% probability)**: you get INFP again. Your dimension scores will not be identical between sessions, but they will land on the same side of the midpoint cutoff for all four dichotomies. This is the modal case.

**Second most likely outcome (about 35% probability)**: you get a single-letter different type, typically with the flipped letter being on the dimension where you originally scored closest to 50%. If your J/P score on session 1 was 52% J / 48% P, the J/P letter is the most likely to flip on retest.

**Third most likely outcome (about 12% probability)**: you get a two-letter different type. This is rarer and more often indicates that the testing conditions varied (one session in a focused mood, another in a tired or distracted mood) than that your underlying personality changed.

**Rarest outcome (about 3% probability)**: you get a three- or four-letter different type. When this happens, it usually means either the test was rushed in one of the sessions, or the question framing was interpreted differently across sessions, or there was a substantial life-circumstance change between sessions that shifted self-perception.

These probability estimates are approximate (derived from Pittenger 2005's review of multiple studies; the exact numbers vary across study samples). The pattern is the structurally important point: most type changes are single-dimension flips at the cutoff, not personality reorganization. If your type changed from INFP to INFJ, you have not become a different person; you've answered slightly differently on items that map to the J/P dimension.

Across the four MBTI dichotomies, the reliability is not uniform. Capraro and Capraro's 2002 meta-analysis (DOI 10.1177/00131640221102234) aggregated reliability studies and found per-dimension test-retest reliability roughly as follows: E/I ≈ 0.66, S/N ≈ 0.65, T/F ≈ 0.59, J/P ≈ 0.61. The J/P dimension is the lowest reliability, with T/F close behind, while E/I and S/N are slightly more stable.

This is consistent with empirical reports of which letter changes most often: in retest studies, J/P is the most-frequently-flipped letter, followed by T/F. E/I and S/N change less often. The reason is partly methodological — the J/P items measure preferences that are more context-dependent (e.g., "Do you prefer planning or going with the flow?" can be answered differently depending on what's currently going on in your life). The reason is partly conceptual — J/P combines structure-preference with closure-seeking, two correlated but distinct constructs that the dichotomy treats as one.

The practical implication: if you took MBTI multiple times and your type changed, check which letter flipped. If it's J/P, you are in the most common pattern, and the appropriate read is that you have a near-midpoint J/P preference rather than that your personality changed. If it's S/N or E/I, the change is rarer and worth more reflection — those letters tend to be more stable, so a flip in them is more likely to indicate a meaningful self-perception shift or a testing-condition difference.

None of this is a critique of MBTI's design — it is descriptive of how the framework actually behaves under measurement. The framework's creators (Isabel Myers and Katharine Briggs) did not have the modern psychometric tools to optimize for reliability when they built the assessment in the 1940s, and the assessment has retained its original four-dichotomy structure for backward-compatibility reasons across decades of practice. Modern Big Five assessments were designed with reliability as a primary engineering target and consequently achieve higher reliability scores.

MBTI's modest reliability is not primarily a function of how the items are written or how the test is administered — it is primarily a function of the architectural choice to convert continuous dimensional scores into categorical type codes via binary cutoffs. The categorical conversion is the source of most of the reliability problem.

Consider what happens in scoring. A test-taker answers 93 items (the Form M MBTI). The items map onto four dimensions, producing four continuous scores: E/I score, S/N score, T/F score, J/P score. Each continuous score is then converted to a binary letter (E or I, S or N, T or F, J or P) by comparing it against a cutoff (usually 50%). Two test-takers with continuous scores of 51% J and 49% J get assigned to different MBTI types (one J, one P) despite being functionally identical on that dimension. The categorical conversion throws away information.

Big Five test scoring keeps the continuous scores. A Big Five report says "73rd percentile Conscientiousness, 41st percentile Openness, ..." rather than "Conscientious type, Open type, ...." When the scores are continuous, retest drift produces small numerical changes ("73rd percentile session 1, 71st percentile session 2") that don't trigger categorical re-classification. When the scores are categorical, the same retest drift can produce a different letter ("J session 1, P session 2") that looks like a major change.

This is the architectural reason MBTI's apparent reliability looks worse than its underlying dimensional reliability. The dimensional reliability (the consistency of the continuous score before binary cutoff) is closer to 0.6-0.7 — still below Big Five's 0.7-0.9 but not as bad as the categorical reliability suggests. The categorical conversion artificially inflates the appearance of instability for any test-taker whose dimension score is near the midpoint.

Pittenger's 2005 review (DOI 10.1037/1065-9293.57.3.210) discusses this explicitly: most of MBTI's famous "50% retest type change" finding is single-dimension flips at the cutoff, not deep personality reorganization. The categorical framing magnifies what would be a small numerical drift into what appears to be a major identity shift. This is the strongest single technical reason researchers prefer Big Five for measurement contexts.

In 1991, the U.S. National Research Council convened a panel to review the scientific evidence behind various human-performance enhancement techniques being marketed to the U.S. military and federal agencies. The resulting volume, edited by Daniel Druckman and Robert Bjork ("In the Mind's Eye: Enhancing Human Performance," National Academy Press, 1991), included a chapter on personality assessment that directly addressed MBTI.

The NRC panel's conclusion: "At this time, there is not sufficient, well-designed research to justify the use of the MBTI in career counseling programs." The panel reviewed the available reliability and validity evidence and found that the assessment did not meet the standards required for selection or career-direction use cases. The reliability evidence reviewed was consistent with the 0.5-0.6 range later confirmed by Capraro & Capraro 2002 and Pittenger 2005.

The NRC review is historically significant because it represented an independent, government-commissioned evaluation by a panel of senior psychometricians. The panel had no commercial relationship with the MBTI publisher and no incentive to either support or undermine the framework. Its conclusion aligned with the Pittenger 2005 academic-research consensus and with the Myers-Briggs Foundation's own "not for selection" Ethical Use Guidelines.

The NRC review is sometimes cited in popular discussions of MBTI as the canonical "NRC concluded MBTI is invalid" finding. That gloss is too strong — the NRC panel did not conclude MBTI is invalid in all uses; it concluded the evidence was insufficient to justify selection or career-counseling use. This is a narrower technical claim than the popular gloss and is consistent with the publisher's own narrowing of legitimate use to development contexts.

If you have your MBTI dimension scores (the percentages or numerical values for each of the four dichotomies), here is the practical reading framework that takes the reliability findings seriously.

A common practical question: I tested as INFP last year and as INFJ this year — which is my real type? The reliability framework gives a structured answer.

**Step 1: Identify which letter flipped.** Compare the two type codes. If only one letter is different, you flipped on that single dimension. If two or more letters differ, the testing-condition or framing differences between sessions are likely playing a larger role and the answer is harder to read.

**Step 2: Look at the dimension score for the flipped letter (if you have it).** If you have access to the dimension percentages from at least one of the test sessions, check whether your score on the flipped dimension was near 50%. If yes, the flip is a near-midpoint flip and means you have a weak preference on that axis — neither type is fully wrong, you're just close to the cutoff.

**Step 3: Don't try to pick "the real type" between the two.** If the flipped letter is on a near-midpoint dimension, neither type code is more correct than the other. Instead, read both type descriptions, identify the dimensions where they agree (those are your reliable letters), and treat the dimensions where they disagree as your near-midpoint axes.

**Step 4: Behavioral test for the flipped axis.** If the flipped letter is J/P, ask yourself behaviorally: do you actually plan ahead and prefer closure, or do you actually keep options open and decide near the deadline? If your real-life behavior pattern is consistent across contexts, that is more reliable signal than either of the two test sessions. If your behavior is genuinely mixed (J in some contexts, P in others), then your preference is mixed and the type code is hiding that nuance.

If you've worked through these four steps and still feel split between two types, the honest read is that MBTI's categorical framework does not carve nature at the joints for you on the contested axis. The Big Five continuous-dimension framework (which doesn't force a binary cutoff) is more honest about borderline cases. For more on this trade-off, see /blog/mbti-vs-big-five.

Three honest takeaways from the reliability evidence.

**Takeaway 1: MBTI carries real signal at the dimension-score level.** Reliability of 0.5-0.6 is modest but not zero. Your specific I/E, S/N, T/F, J/P scores are produced by a deterministic algorithm from your specific answers, and they carry information about your underlying preferences. The signal is just not strong enough to make near-midpoint dimensions stable across retests.

**Takeaway 2: The categorical type code is less reliable than the dimensional scores underneath.** When you read "I'm INFP" or "I'm ENTJ," treat that as compressed information about a four-dimensional measurement. The compression is convenient (the letters travel in conversation) but loses the percentile information that distinguishes a strong INFP (90% I, 80% N, 75% F, 70% P) from a borderline INFP (52% I, 53% N, 51% F, 51% P). The borderline cases are where retest flips happen.

**Takeaway 3: For measurement contexts (research, hiring, longitudinal self-tracking), Big Five is the better instrument because of higher reliability.** For vocabulary contexts (team-building, self-reflection, conversational shorthand), MBTI is fine because the categorical labels work as memorable anchors. The two frameworks serve different purposes and are not in zero-sum competition.

For the long-form treatment of MBTI's measurement properties beyond reliability, see /blog/mbti-common-misconceptions-and-data. For the framework comparison with Big Five, see /blog/mbti-vs-big-five. For the specific case of why MBTI is not a fit for hiring, see /blog/mbti-for-hiring.

Three caveats to keep the reliability findings calibrated.

**Caveat 1: Reliability of 0.5-0.6 is modest, not catastrophic.** The number sometimes gets quoted with the implication that MBTI is essentially random — "50% of people get a different type on retest, so it's a coin flip." That overshoots. Most retest changes are single-dimension flips at the cutoff; multi-dimension flips are much rarer. The framework is producing real measurement, just with substantial noise on the categorical letters.

**Caveat 2: Reliability and validity are independent.** A test can be reliable (reproduces itself) without being valid (measuring what it claims). A test can be valid (measuring something real) without being highly reliable (the measurement is noisy). The reliability evidence reviewed here does not, by itself, settle the validity question. The validity question is covered separately in Pittenger 2005 and in /blog/mbti-common-misconceptions-and-data.

**Caveat 3: The reliability findings do not invalidate all uses of MBTI.** The Myers-Briggs Foundation explicitly endorses development, team-building, and individual-coaching uses where the categorical labels work as vocabulary. Those uses do not require high test-retest reliability the way selection contexts do. The reliability evidence is most consequential for uses where consistency matters (hiring, research, performance evaluation), and least consequential for uses where the labels function as conversational anchors.