# Mechanism × Phase × Failure-Mode Matrix — v0.2 Status: **working Bridge Object** Source: unified PA001–PA015 manifest. This matrix is descriptive and hypothesis-generating; it is not a risk-of-bias tool, causal model, or evidence ranking. ## Why this matrix exists The 15-study manifest shows that the same digital component can mean different things depending on: 1. **phase** — acquisition, continuous treatment, maintenance, post-intervention follow-up; 2. **incremental mechanism** — what the digital layer adds beyond the comparator; 3. **exposure** — whether participants actually continue to use/read/wear it; 4. **descriptive failure/constraint state** — what is observed when durable incremental value weakens, disappears, or cannot be cleanly interpreted. The matrix therefore asks: > When a digital PA intervention does not show durable incremental value, **which implementation/exposure, outcome, comparator, or attribution state best describes the evidence?** ## Controlled failure-mode vocabulary These are operational descriptive categories, not claims that a psychological or behavioral mechanism itself failed. ### F1 — No maintenance test The study ends during active exposure, so durability after withdrawal cannot be assessed. ### F2 — Effect attenuation An early between-group advantage weakens or disappears later. ### F3 — Comparator saturation The control/active-comparator already contains self-monitoring, feedback or another strong behavioral backbone, leaving little incremental headroom. ### F4 — Burden / habituation High frequency, repetition or automation may reduce attention to the intervention. ### F5 — Exposure uncertainty / attrition Delivery is known, but continued reading/use/wear/data completeness is weak or uncertain. ### F6 — Maintenance add-on redundancy Continuing an acquisition component during maintenance adds no measurable value beyond a simpler maintenance backbone. ### F7 — Preservation rather than gain Success is better understood as prevention of decline rather than additional improvement. ### F8 — Behavior-outcome decoupling Behavior changes without corresponding improvement in a downstream physiological/clinical outcome. ### F9 — Response/relevance heterogeneity Average effects obscure meaningful differences by responder status, perceived usefulness or other treatment-response structure. ### F10 — Multi-component attribution ambiguity The intervention works at package level but the active ingredient cannot be isolated. --- ## Matrix | ID | Phase | Incremental mechanism | Adaptation | Main failure mode | Structural reading | |---|---|---|---|---|---| | PA001 | Active only | Goal-setting + points feedback | Fixed factorial | F1 / F10 | Good acquisition component test; no post-exposure durability evidence | | PA002 | Continuous active | Tailored web intervention | Static tailoring | F2 | Early advantage disappears as active comparator catches up | | PA003 | Active only | Adaptive goals | Dynamic algorithmic | F3 | Relative benefit despite absolute decline in both arms | | PA004 | Active + post-follow-up | PAM + tailored advice | Static tailoring | F2 / F5 | No PA benefit at intervention end or later follow-up | | PA005 | Active only | Automated multi-component support | Static tailoring | F1 / F10 | Package-level benefit, no post-intervention test | | PA006 | Active only | SMS on top of Fitbit | Static timing preference | F4 | Heavy automated contact gives only transient benefit | | PA007 | Continuous active | Targeted SMS | Static targeted | F2 / F7 | Apparent benefit is mainly less decline, then attenuates | | PA008 | Continuous active | mHealth cardiac-rehab package | Static personalization | F8 | PA improves without peak-VO₂ advantage | | PA009 | Active only | Social comparison added to personal feedback | Fixed | F3 | Added social layer does not beat an already active feedback comparator | | PA010 | Acquisition → maintenance | Continued app diary/self-monitoring | Static tailored/graded | F6 / F5 | Continuing acquisition-era app support does not improve maintenance | | PA011 | Multistage adaptive + follow-up | Response-triggered escalation | SMART | F9 / F10 | Adaptive sequence looks promising but component attribution is complex | | PA012 | 18-month continuous | Training + peer support added to tracker/PAI | Static personalization | F2 / F5 | Extra layers produce only temporary separation; attrition is substantial | | PA013 | Long-term continuous | JITAI coaching on top of step counter | JITAI | F9 / F5 / F3 | Average effect null; usefulness matters; exposure is uncertain | | PA014 | Post-program maintenance | Tracker + professional support vs telephone counseling | Personalized support | F7 | Maintenance success appears as preserved behavior, not more gain | | PA015 | 32-week continuous | Step/rank app on top of shared program | Static | F8 / F3 | PA benefit without extra weight loss; shared backbone limits attribution | --- # Cross-era pattern 1 — The limiting evidence moved downstream Earlier studies often show limitations because: - the package is fixed; - extra prompts lose effect; - no post-intervention phase is tested. Recent studies increasingly fail **after** technically sophisticated support is already in place: - JITAI is delivered but not experienced as useful enough; - extra online layers add little beyond an already strong tracker/app backbone; - adaptive sequencing helps some responders but creates attribution complexity. The frontier therefore moved from: > Can digital delivery change behavior? to: > Can digital support remain incrementally useful after self-monitoring and feedback are already present? --- # Cross-era pattern 2 — Null or attenuated maintenance findings arise from distinct states A single label such as “did not maintain effect” hides structurally different observations: ### A. Attenuation Early effect weakens with time. Examples: PA002, PA006, PA007, PA012. ### B. Redundant maintenance support Continued digital exposure adds no value over a simpler maintenance condition. Example: PA010. ### C. Engagement/exposure decay Participants stop reading, using, wearing, or providing usable data. Examples: PA006, PA010, PA012, PA013. ### D. Comparator saturation The control already includes strong self-monitoring/feedback, reducing incremental headroom. Examples: PA003, PA009, PA013, PA015. These states and signals should not be pooled conceptually. --- # Cross-era pattern 3 — “More personalized” is too crude a descriptor The 15 rows now distinguish: 1. fixed component assignment; 2. static tailoring; 3. dynamic algorithmic goal adaptation; 4. contextual JITAI; 5. response-adaptive SMART sequencing. Those levels answer different behavioral questions. The strongest unresolved question is therefore not: > Is personalization effective? It is: > **Which form of adaptation changes a decision that still matters to the participant at that phase of behavior change?** --- # Cross-era pattern 4 — Behavioral salience is a useful bridge construct The matrix suggests a common pathway behind several apparently different failures: ``` support available → support noticed → support still relevant → support actionable now → action repeated → behavior preserved ``` Failure can occur between any two links. This gives a more precise working definition: ## Behavioral salience A digital mechanism remains behaviorally salient when it is still **noticed, relevant, actionable and connected to a meaningful behavioral decision** at the current phase. This is a hypothesis-generating construct, not yet a validated measure. --- # External expert-check questions generated by the matrix These are better network-edge questions than “What do you think of our project?” ## Q1 — Maintenance architecture When initial behavior change has already occurred, which components should be **withdrawn**, which should remain, and which should become adaptive? ## Q2 — Comparator saturation When both groups already receive self-monitoring and feedback, what should count as a meaningful incremental digital mechanism? ## Q3 — Implementation/exposure vs incremental behavioral effect Is our operational mapping of PA006, PA010, PA012 and PA013 onto implementation/exposure versus incremental behavioral-effect failure defensible, and are we using the most appropriate established terminology? ## Q4 — Preservation as outcome For maintenance trials, should preventing decline be treated as a distinct success state rather than a weaker version of continued improvement? ## Q5 — Adaptive relevance Is perceived usefulness best treated as an engagement construct, an effect modifier, or part of the intervention mechanism itself? --- # Candidate second Bridge Object A compact public derivative can be built from this matrix: **Digital Physical-Activity Maintenance Failure Map v0.1** Suggested structure: - 15 study IDs; - phase; - incremental mechanism; - engagement/exposure signal; - maintenance outcome; - primary failure mode; - one-line interpretation. Its purpose would be different from the original Micro Map: - **Micro Map:** what mechanisms/intervention structures were used? - **Failure Map:** where does durable incremental value break down? Do not publish this second object yet. First obtain at least one real external researcher/method-expert response and use that response to decide whether the descriptive failure/constraint vocabulary is defensible.