Contribution
The 2006 meta-analysis (LIT-1497) pooled 94 tests of goal attainment. This one pools 642 tests, widens the outcomes to cognition, affect and social judgement, corrects for publication bias, and asks two new questions: for which outcomes and study features planning works (its scope), and which plan formats, formation procedures and contents work (its components). Its new product is a taxonomy of the cues and responses that plans have specified.
Key insight
The format matters more than the content. A plan that links a cue to a response (“if X arises – then I do Y”) works better than one that schedules when, where and what without the link, and almost any cue and any response can be put into the link. The authors’ reading: the contingent format “causes participants not only to become perceptually ready to meet the cue but also mentally link that cue with the prepared response” (p. 165), and anything that invites deliberation about the plan, such as keeping a written copy, undermines it.
Assumptions
- Inclusion (p. 167). Experimental designs only (unlike 2006); planning isolated from other intervention components; the content of the plans retrievable, so menu-driven planning and mental contrasting with implementation intentions were excluded; an outcome with a computable effect size.
- Search (pp. 167–168). Web of Science and PsycINFO, January 1994 to September 2019, with citation and hand searches. Not updated: “Several factors precluded updating the literature search” (p. 168), among them funding and personnel; year of publication did not moderate any effect (all ps > .061). Not preregistered, because the taxonomy was built inductively (p. 167).
- Units (p. 169). Each treatment condition is a separate test, with the control group’s N divided among them; the longest follow-up and intention-to-treat data were used. Random-effects models in Stata 17.
- Plan schedules count as implementation intentions here. 269 of the 642 tests used a schedule format (Table 3, p. 174): more than two in five of the “tests of implementation intentions” are of plans the authors say should not automate responding (p. 165).
Key results
- Sample (pp. 169–170). Median 30 treated and 28 controls per test (range 6 to 33,995); 557 tests with adults, 363 with university students; mean age 28.6, 60.4% female.
- Overall and bias (p. 170). Pooled d = .36 [.33, .40]. Egger’s b = 1.06, “substantial” publication bias; trim and fill gave .35. RoBMA: strong evidence of an effect (BF10 = 319.3) at a model-averaged d = .15 [.08, .22], strong evidence of publication bias and of heterogeneity (tau = 0.27). “Thus, implementation intentions appear to have a robust impact on outcomes though the effect size is small according to conventional guidelines” (p. 170).
- Scope by outcome (Table 1, p. 171). Behaviour d = .27 (k = 301), cognitive performance .47 (k = 204), affect regulation .66 (k = 26; fear 1.22, disgust .77, anxiety and stress .44), social judgement .52 (k = 20). Physical activity .14 (k = 64); vaccination .11 and voting .14, neither reliably above zero; psychotherapy attendance .19.
- Scope by study feature (Table 2, pp. 172–173). Similar for children, adolescents, adults and older adults. Lab .49 against field .27; in person .53 against online .31. Objective outcomes .43, larger than self-report .28. Follow-up under a week .48, 1–6 months .19, over 6 months .51 (k = 12). Control conditions did not differ (p = .053). High against low motivation, d = .79 (k = 8).
- Motivation explains the weak physical-activity effect (p. 177). Among 31 tests reporting intention strength, planning worked when intentions were strong (d = .31) and “had no effect when intentions were weak”.
- Components (Table 3, pp. 173–174). Contingent format .43 (k = 373) against schedule .29 (k = 269), Q(1) = 20.81. Rehearsing the plan at least once .45 against .33, and more rehearsal added nothing. Recording the plan .31 against .44 without. Visualising the response made no difference. Number of plans: 3–5 plans did worst, and one plan or 7–8 did best.
- Recording (p. 180). “when participants had the opportunity to revisit their plans (e.g., because they retained copies), those plans were less effective.” Read by the authors as deliberation undermining automaticity. Adding how or for how long to a time-and-place cue roughly halved the effect (.24 against .46).
- The taxonomy (Table 4, pp. 175–176, 180–183). Six kinds of cue (time and place, a specific event, a task juncture, a critical stimulus, a feeling, a thought) and six kinds of response (approach or avoidance behaviour, cognitive procedure, ignoring, inner speech, ignoring plus inner speech, a goal reminder). Nearly all were effective. Cues the participants chose themselves did worst (d = .16, k = 20). “Wait before responding” gave .66 (k = 7); “I will keep calm” .69 (k = 14), larger than the authors’ cited figures for reappraisal (.36) or distraction (.27) from another meta-analysis (p. 182).
- The mechanism is cited, not tested here. The introduction states that cue accessibility and cue–response links “serve to delegate action control from the self to prespecified cues” (p. 163), from earlier studies. This review tests no mediator; it reads the rehearsal and recording results as consistent with the mechanism and asks for direct tests (p. 180).
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | Forming implementation intentions has a reliable beneficial effect on outcomes | strong for existence (BF10 = 319) | RoBMA, p. 170 |
| C2 | The bias-corrected effect is small, d ≈ .15, and highly heterogeneous | moderate: one correction method; trim and fill gives .35 | p. 170 |
| C3 | Plans help cognitive performance, affect regulation and social judgement as well as behaviour | moderate; affect regulation k = 26 | Table 1 |
| C4 | Contingent if-then plans outperform schedule plans | moderate: between-test comparison, not experimental | Table 3, p. 173 |
| C5 | Rehearsing a plan helps and recording it hurts | moderate, between-test; the authors’ explanation untested | Table 3; p. 180 |
| C6 | Planning works when motivation to reach the goal is high, and not when it is weak | moderate: k = 8, and 31 physical-activity tests | Table 2; p. 177 |
| C7 | Nearly every cue and response type is effective; self-chosen cues are weakest | weak to moderate: many categories have k < 10 | Table 4 |
Concepts
- Implementation intention: an if-then plan, “If [opportunity/obstacle] arises, then I will [respond in this way]!” (p. 179).
- Contingent against schedule format: a plan that links cue to response, against one that lists when, where and what with no conditional link (p. 165).
- Scope and components: for which outcomes and study features plans work; which format, formation process and contents (pp. 163–166).
- Cue and response taxonomy: Table 4’s inductive categories.
Connections
It updates Gollwitzer & Sheeran 2006 (LIT-1497, NOTE-1124) and corrects its picture of the effect’s size: the 2006 estimate was d = .65 from 94 tests with no correction for bias. It draws on Webb & Sheeran’s mediation studies for the mechanism, on Webb, Schweiger Gallo et al.’s action control model of affect regulation for the affect results (p. 178), and on the RoBMA method of Maier, Bartoš and Wagenmakers.
Bearing on the record
- CLAIM-443 (a learned prior over which processes govern). The
claim cites this review for implementation intentions as pre-arranged
if–then links of situation to response, and for nothing more. The reading
supports that and adds four things.
- The link, not the schedule, is what works. Plans with the contingent format beat plans that fix the same when, where and what without the conditional link (THEORY-908). That fits the claim’s interest in learned relations rather than fixed behaviour, though the link here is cue to action.
- Deliberating over the plan weakens it. Keeping a copy to revisit, or specifying how and for how long, cut the effect. That is evidence, between tests, that the plan’s value lies in removing deliberation in the situation, which is the claim’s amortization side. It is not a measure of computation saved.
- Motivation gates it. Plans worked when the goal was wanted and not otherwise, as in 2006’s goal-dependence (THEORY-904).
- The effect is small once corrected (THEORY-918). The record should not cite implementation intentions as a large effect; d = .65 is the 2006 figure, before bias correction.
- Nothing here tests the claim’s distinguishing case, a change in which response is appropriate.
- CLAIM-349 (access, authority and action are separable). The taxonomy holds response types that act on access and on authority, not only on action: ignore responses (d = .48), waiting before responding (.66) and goal reminders (.79, k = 3). “Wait before responding” is the same operation as DBT’s STOP, as the claim’s note of 2026-10-23 reads it: holding the transition from urge to action open. A parallel, from small categories.
- CLAIM-358 and QUESTION-095 (amortized cost; caching). The between-test findings on recording and elaboration are consistent with a cached decision losing value when re-opened. Not tested directly.
- No instruction for machine-learning practice; nothing for the anthology.
Limitations
- Substantial publication bias, by the authors’ own analysis; “Until then, caution is warranted in interpreting findings from the present review” (p. 183).
- The search ends in September 2019, for a paper published online in March 2024.
- Moderator results are between-test comparisons, confounded with outcome type and lab: the authors say so for the number of plans (p. 180).
- Many categories rest on fewer than 10 tests, some on 2 or 3 (p. 183).
- No mechanism is tested in this review.
- Race, education and socioeconomic status could not be summarised (pp. 169, 183).
- Reporting slips. “(d = 45. vs. 33)” for .45 against .33 (p. 173); “if-than plan” in a heading (p. 179); seven response categories in the limitations (p. 184) against six in the results (p. 176).
Open questions
- Does the contingent format beat scheduling in an experiment that assigns the format, outcome and sample held fixed?
- Does rehearsal work through cue accessibility and link strength, as the authors suggest? A direct test is what they ask for (p. 180).
- What happens to an if-then plan when its response stops being the right one? Neither this review nor the 2006 one reports such a test.
Note of 2026-10-29: what d = .79 measures
The summary’s “effects were large only when motivation was high”, the results line “High against low motivation, d = .79 (k = 8)” and claim C6 read Table 2’s motivation row as a plan effect under high motivation. It is the difference: forming implementation intentions “had a larger effect when intentions were strong, activated, or self-concordant as compared to weak, low in activation, or self-discordant (d = .79)” (p. 173), “a large moderating effect” (p. 177), over eight tests. An effect absent under weak intentions was found only in the 31 physical-activity tests split at the median of intention (d ≈ 0 against .31, p. 177). Checked against the version of record on 2026-10-29; THEORY-904 was corrected the same day.