---
name: Self-determination theory
slug: self-determination-theory
category: Motivation
categories: [Onboarding, Retention]
evidence: mixed
page: /principles/self-determination-theory/
updated: 2026-10-15
---

# Self-determination theory

> People keep doing things that feel self-chosen for longer than things that feel required.

## What it is

The theory holds that motivation depends on three needs — autonomy, competence and relatedness. The product-relevant consequence is that motivation running on the user's own reasons survives the withdrawal of the ones you supplied, and motivation you supplied does not.

## Evidence

This is a framework rather than an effect, which is the first thing to be honest about. Ryan and Deci (2000) postulate three innate psychological needs — competence, autonomy and relatedness — which when satisfied yield self-motivation and well-being and when thwarted diminish both. No screen can falsify that sentence. What can be tested are the specific moves the theory licenses: letting the user choose, telling them what they are getting better at, connecting them to other people, and not paying them for things they already wanted to do. The last of those has its own entry. The best-quantified of the rest is choice.

Patall, Cooper and Robinson (2008) meta-analysed 41 studies on choice. Choice raised intrinsic motivation at d = 0.30 under a fixed-effects model and 0.36 under random effects, and it also raised perceived competence (0.59), task performance (0.32) and effort (0.22) while doing nothing measurable for how much people learned (0.10, interval crossing zero). Two numbers set the planning figure: published studies gave 0.41 and unpublished ones 0.20, and trim-and-fill drops the average to 0.21–0.22, about a third smaller by the authors' own description. Measurement type did not matter at all — behavioural measures gave 0.29 and self-report 0.31 — so this is an effect you can test on behaviour rather than on a survey.

The moderators are the design spec, and they are where this principle stops contradicting choice overload. The largest effect came from what the authors call instructionally irrelevant choices — what colour pen to write with, what name to give a game character, what music to play while exercising — at d = 0.59, against 0.27 between versions of a task, 0.24 for choices with real consequences for the outcome and 0.16 for choosing between activities. Three to five options beat both two and more than five (0.38 against 0.20 and 0.26). Two to four choices beat one and beat five or more (0.61 against 0.21 and 0.32), and choices made one at a time beat several picks off a single list (0.54 against 0.18). Patall and colleagues read this as autonomy gain against self-regulatory cost: every extra decision buys less and costs more. Autonomy in an interface is a few cheap picks from short lists, not a long settings page.

Two things in the same meta-analysis switch the effect off. A reward external to the choice reduced it to d = −0.01, against 0.35 with no reward and 0.35 when the user chose the reward themselves, so an incentive attached to a setup flow cancels the autonomy it was meant to support. Pressure costs most of the rest: yoked designs, where the participant received what they picked, gave 0.37, while designs in which choice was reduced by persuasion or by differentially attractive alternatives gave 0.20. The effect was also larger when the control group was explicitly denied a choice (0.43) or at least knew the alternatives existed (0.36 against 0.24 when unaware), which is the uncomfortable half of the finding: part of what is being measured is the cost of visibly withholding a choice rather than the gift of granting one.

Outside the lab the intervention record is real, small and entirely off-screen. Sheeran and colleagues (2020) pooled 65 independent tests from randomised trials of self-determination interventions for health behaviour change and found d+ = .23, with significant effects for physical activity, sedentary behaviour, diet, alcohol and smoking cessation; after correcting for publication and small-sample bias the estimate held at .15 or above, mediated by autonomous motivation and perceived competence. Ntoumanis and colleagues (2021) pooled 73 studies, N = 30,088, and reported g = 0.64 for need support, 0.37 for autonomy satisfaction and 0.45 for health behaviour at the end of the intervention, falling to 0.28 at follow-up. Three results there matter more than the averages: relatedness satisfaction did not move (g = 0.20), quasi-experimental trials produced larger behaviour effects than randomised ones, and the two techniques that produced reliably larger effects on psychological health were using non-controlling language and providing a meaningful rationale. Those two are copy decisions.

On actual interfaces the behavioural evidence is thin enough to name study by study. Peng, Lin, Pfeiffer and Winn (2012) switched autonomy-supportive and competence-supportive features on and off in a 2 × 2 exergame experiment: the features produced the matching need satisfaction and moved most motivation and engagement outcomes, all of which were self-reported. Birk, Atkins, Bowey and Mandryk (2016) found that identification with the avatar predicted autonomy, enjoyment and time spent in an unending version of a runner game — a mediation path, not a difference between arms. The one result shaped like a product metric is Birk and Mandryk (2018): 250 participants asked to do a one-minute breathing exercise daily for three weeks, where the arm that customised its avatar showed significantly less attrition and more sustained engagement in login counts. That is a single study, and the published abstract reports no effect size.

The measurement story is worse than the reputation. Need satisfaction in games is usually indexed by the Player Experience of Need Satisfaction scale, introduced alongside Ryan, Rigby and Przybylski's (2006) studies but never validated in print by its authors; Johnson, Gardner and Perry (2018) supplied the first factor-analytic test, n = 571, largely supporting the structure while recommending that two subscales be merged, and the Player Experience Inventory (Vanden Abeele and colleagues, 2020) has the better psychometrics. On the question a product team actually asks — does a need score predict behaviour — the evidence is one correlational dataset: Azadvar and Canossa (2018) report each UPEQ subscale independently predicting engine-tracked playtime, money spent and group play, and Kayser, Perrig and Brühlmann (2021) then found weaknesses in that scale's three-factor structure with 369 players. Whether a need score predicts churn in your product is unanswered, which makes it a diagnostic and not a KPI. Tyack and Mekler's audits are the honest summary of the applied literature: of 110 CHI and CHI PLAY papers citing the theory, few engaged with it beyond descriptive accounts, and across 259 papers the dominant use was selecting a self-report scale, in 76 cases with no theoretical rationale. They also point out that the claim a high degree of choice satisfies autonomy has never been tested there, that one study found high in-game choice perceived as controlling, and that recent work found no support for the claim that optimal challenge raises competence satisfaction. Build from the choice moderators, measure behaviour, and treat the three needs as a checklist for what to try rather than as a finding.

## When it holds

- Low-stakes, expressive choices carry it. Instructionally irrelevant choices — pen colour, a character's name, the music playing — gave d = 0.59, against 0.24 for choices with real consequences for the outcome and 0.16 for choosing between activities.
- Three to five options, not two and not a dozen. Two options gave 0.20, three to five gave 0.38, more than five gave 0.26.
- Two to four choices, made one at a time. One choice gave 0.21, two to four gave 0.61, five or more gave 0.32; successive single picks gave 0.54 against 0.18 for several picks off one list.
- No unrelated reward in the same flow. A reward external to the choice reduced the effect to d = −0.01, against 0.35 with no reward at all and 0.35 when the user picked the reward.
- The user has to get what they picked. Yoked designs gave 0.37; designs where choice was reduced by persuasion or by differentially attractive alternatives gave 0.20.
- The alternative has to be visible. Where control participants knew the options existed the effect was 0.36, against 0.24 where they did not, and 0.43 where they were explicitly denied a choice.
- Expect about a fifth of a standard deviation. Published studies gave 0.41 and unpublished ones 0.20, and trim-and-fill brings the average to 0.21–0.22.
- Relatedness is the weakest leg. Across 73 intervention studies, need support moved g = 0.64 and autonomy 0.37, while relatedness satisfaction did not move at all (g = 0.20).
- It is not a comprehension aid. Choice did nothing measurable for how much people subsequently learned, d = 0.10 with an interval crossing zero, so framing a tutorial as autonomy support will not make it land better.

## How to apply

- Build setup out of two to four separate picks, one per screen, three to five options each. That is the shape of the strongest cell in the meta-analysis, and the opposite of a single long preferences page.
- Spend the autonomy on expressive, reversible things: display name, avatar, theme, the reminder's wording, the order of the home list. Cosmetic choices measured d = 0.59, cost nothing to honour, and do not fork your product.
- Carry each pick into the first screen after setup, visibly. Receiving what you chose is the yoked condition (0.37); quietly overriding it is the reduced-choice condition (0.20).
- Put the edit control for a choice where its consequence appears, not in a settings tree. A choice the user cannot revisit is a configuration you took from them.
- Attach no reward, discount or bonus to completing the setup choices. An external reward in the same flow measured d = −0.01 — it does not dilute the autonomy effect, it removes it.
- Where you have to constrain something, give the reason in non-controlling language: “we start everyone on weekly digests because daily mail buries the useful one”, not “you should keep this on”. Non-controlling language and a meaningful rationale were the two techniques with reliably larger effects.
- Report competence informationally: what the user did, and what changed because of it, rather than a score to chase.
- Make difficulty a control the user sets, not an engine that adapts it for them. The claim that optimal challenge raises competence satisfaction has no empirical support in this literature.

## What to avoid

- Choices the product does not honour: a toggle that changes nothing, a stated preference the recommender overrides, a frequency the notification service ignores. Every measured benefit of choice comes from conditions where the participant received what they chose.
- A preselected recommendation surrounded by visibly worse alternatives. Reducing choice by persuasion or by differentially attractive options cut the effect from 0.37 to 0.20, and where the choice is a consent it is also a legal problem: GDPR Article 4(11) requires consent to be freely given, and Article 7(4) tells you to take utmost account of whether a service was made conditional on processing it does not need.
- Counting the size of your settings screen as autonomy support. More than five choices measured 0.32 against 0.61 for two to four, and once the options get complex and the user does not yet know what they want, you are in choice-overload territory instead.
- Reporting a need-satisfaction score as a success metric. The PENS was published without a validation by its authors, the first independent factor analysis recommended merging two of its subscales, and the three-factor structure of the game-industry alternative did not hold up with 369 players.
- Launching a community because the theory has three legs. Relatedness satisfaction was the one construct 73 intervention studies failed to move, so a forum, a feed or a friends list has to earn its place on its own evidence.
- Selling the customisation. Choice effects were larger for children (0.55) than adults (0.25), and a paid personalisation aimed at the group the effect works hardest on is a sharper instrument than it looks.
- The ethical line: the user's choice has to survive contact with your funnel. If a pick can be reversed by a growth experiment, a default reset on update or a dialog that asks again until it gets the other answer, you are running the illusion of autonomy, which the evidence says is worth about half as much and costs you the trust as well.

## Measuring it

- Primary metric: return rate, sessions or adherence over the following weeks for a cohort that made the setup choices against a cohort assigned sensible defaults. Behavioural and self-report measures of the choice effect were the same size (0.29 and 0.31), so there is no reason to settle for the survey; power the test for roughly 0.2 of a standard deviation.
- Secondary: the share of users who change at least one default, and the share of their picks still in force 30 days later. A customisation path nobody walks is not autonomy support, and a setting that silently reverts is worse than one you never offered.
- Backfire metric: setup completion and time-in-setup, per added choice step. Every extra decision carries the self-regulatory cost that flattens the effect past four or five choices, and a step that loses signups is not paid for by d = 0.2 downstream.
- Backfire metric: overrides. Count the cases where a stored preference was ignored by a recommender, a reset, a migration or an experiment. This is the metric that distinguishes the principle from its anti-pattern, and no standard dashboard has it.
- Backfire metric: for any needs survey you run, the correlation between the score and the next period's behaviour in your own data. Near zero means you built a satisfaction metric, and should stop calling it motivation.

## Live examples

- UNIQLO makes the first of three setup questions a single short list — Women, Men, Kids, Baby — with the counter reading 1 / 3, a Skip in the corner and the reason printed on the screen: “We will recommend items that best match with your preferences.”
- Reddit hands over a pick with nothing riding on it — a swipeable carousel of ready-made avatars, captioned “Select your first Avatar. You can always change your style later.”, with Skip in the top corner.
- MyFitnessPal leaves the difficulty to the user rather than fitting it to them: 12:12 marked Easiest, 14:10 Moderate and 16:8 Challenging, on the first of three steps, under the line “New to this? Begin your journey with a shorter fast. You can switch any time”.
- Instagram puts Not interested and Interested as two buttons of equal size directly under the post they would act on, so the control sits against the thing it changes instead of in a settings tree.
- Beli shows the anti-pattern on the same kind of screen: “Allow notifications” is a filled teal button and “Not now” is grey text with no button around it, so the two answers to one question are not equally easy to find.

## References

- Ryan & Deci (2000). Self-Determination Theory and the Facilitation of Intrinsic Motivation, Social Development, and Well-Being. American Psychologist. — https://doi.org/10.1037/0003-066X.55.1.68
- Patall, Cooper & Robinson (2008). The Effects of Choice on Intrinsic Motivation and Related Outcomes: A Meta-Analysis of Research Findings. Psychological Bulletin. — https://doi.org/10.1037/0033-2909.134.2.270
- Sheeran, Wright, Avishai, Villegas, Lindemans, Klein, Rothman, Miles & Ntoumanis (2020). Self-Determination Theory Interventions for Health Behavior Change: Meta-Analysis and Meta-Analytic Structural Equation Modeling of Randomized Controlled Trials. Journal of Consulting and Clinical Psychology. — https://doi.org/10.1037/ccp0000501
- Ntoumanis, Ng, Prestwich, Quested, Hancox, Thøgersen-Ntoumani, Deci, Ryan, Lonsdale & Williams (2021). A Meta-Analysis of Self-Determination Theory-Informed Intervention Studies in the Health Domain. Health Psychology Review. — https://doi.org/10.1080/17437199.2020.1718529
- Peng, Lin, Pfeiffer & Winn (2012). Need Satisfaction Supportive Game Features as Motivational Determinants: An Experimental Study of a Self-Determination Theory Guided Exergame. Media Psychology. — https://doi.org/10.1080/15213269.2012.673850
- Birk, Atkins, Bowey & Mandryk (2016). Fostering Intrinsic Motivation through Avatar Identification in Digital Games. Proceedings of the 2016 CHI Conference on Human Factors in Computing Systems. — https://doi.org/10.1145/2858036.2858062
- Birk & Mandryk (2018). Combating Attrition in Digital Self-Improvement Programs using Avatar Customization. Proceedings of the 2018 CHI Conference on Human Factors in Computing Systems. — https://doi.org/10.1145/3173574.3174234
- Ryan, Rigby & Przybylski (2006). The Motivational Pull of Video Games: A Self-Determination Theory Approach. Motivation and Emotion. — https://doi.org/10.1007/s11031-006-9051-8
- Johnson, Gardner & Perry (2018). Validation of Two Game Experience Scales: The Player Experience of Need Satisfaction (PENS) and Game Experience Questionnaire (GEQ). International Journal of Human-Computer Studies. — https://doi.org/10.1016/j.ijhcs.2018.05.003
- Vanden Abeele, Spiel, Nacke, Johnson & Gerling (2020). Development and Validation of the Player Experience Inventory: A Scale to Measure Player Experiences at the Level of Functional and Psychosocial Consequences. International Journal of Human-Computer Studies. — https://doi.org/10.1016/j.ijhcs.2019.102370
- Azadvar & Canossa (2018). UPEQ: Ubisoft Perceived Experience Questionnaire. Proceedings of the 13th International Conference on the Foundations of Digital Games. — https://doi.org/10.1145/3235765.3235780
- Kayser, Perrig & Brühlmann (2021). Measuring Players' Experience of Need Satisfaction in Digital Games: An Analysis of the Factor Structure of the UPEQ. Extended Abstracts of the 2021 Annual Symposium on Computer-Human Interaction in Play. — https://doi.org/10.1145/3450337.3483499
- Tyack & Mekler (2020). Self-Determination Theory in HCI Games Research: Current Uses and Open Questions. Proceedings of the 2020 CHI Conference on Human Factors in Computing Systems. — https://doi.org/10.1145/3313831.3376723
- Tyack & Mekler (2024). Self-Determination Theory and HCI Games Research: Unfulfilled Promises and Unquestioned Paradigms. ACM Transactions on Computer-Human Interaction. — https://doi.org/10.1145/3673230
- Regulation (EU) 2016/679 (GDPR), Article 4(11): consent means any freely given, specific, informed and unambiguous indication of the data subject's wishes. — https://www.legislation.gov.uk/eur/2016/679/article/4
- Regulation (EU) 2016/679 (GDPR), Article 7(4): in assessing whether consent is freely given, utmost account shall be taken of whether a service is conditional on unnecessary processing. — https://www.legislation.gov.uk/eur/2016/679/article/7

## Similar principles

- Overjustification effect — /agents/overjustification-effect.md
- Choice overload — /agents/choice-overload.md
- Variable reward — /agents/variable-reward.md