The Good Life

I. What is being measured

Research measures at least five aspects of a good life: life satisfaction (how a person rates their life overall, usually 0–10), daily affect (how their days actually feel, collected by sampling people during the day), meaning (whether life feels worthwhile), symptoms (depression, anxiety, pain), and health and longevity. These terms are used consistently throughout.

They overlap but are not interchangeable. Parenthood is associated with more meaning and harder days. Money moves life satisfaction more than daily affect. Treatment can reduce symptoms without addressing whether a life is meaningful.

There are also things research barely measures. A person may value freedom, knowledge, achievement, character, contribution, or having their informed preferences met, whether or not those improve mood. Evidence on meaning is thinner and more observational than evidence on the other four; evidence on the rest is close to absent.

This document therefore does not define a good life. It asks a narrower question: what changes the outcomes that can be measured, by how much, and with what confidence?

The conclusion. The evidence is strongest at identifying large harms and basic conditions, useful but less decisive for recurring life circumstances, and weakest at ranking versions of an already good life.

On self-report. Four of the five measures rely on it. Self-reports predict later suicide, match ratings by friends and family, and predict who leaves a job or a marriage. They are also affected by what the person was just asked about, liable to overweight the most intense and the final moments of a remembered episode, and used differently across languages. People also reset their own standards, so an unchanged score can mean unchanged conditions or a moved yardstick. Health and mortality outcomes appear throughout as a check.

II. How to read the evidence

The central problem is selection. If people who exercise are happier, exercise may cause happiness — or happier, healthier, wealthier people may exercise more. The same problem affects almost every study of marriage, work, neighbourhood, and money.

Study designs, roughly strongest first. Randomized trials assign the exposure independently of the person. Natural experiments use events that break the link between exposure and personal characteristics. Some come close to random assignment — lottery prize variation among people who all played, an airport changing flight paths. Others are better described as plausibly exogenous shocks: a plant closure removes worker choice from the timing, but which plants close, in which industries, staffed by whom, is not random. Studies following the same person over time remove stable differences between people, but not things changing alongside the exposure. Twin comparisons remove genes and shared upbringing, but not the non-shared reasons twins diverged. Adjusted observational studies depend entirely on what was measured. Cross-sectional comparisons are weakest, and are the source of most popular advice about how to live.

Twin comparisons for exercise and marriage typically come out much smaller than ordinary observational estimates, though not always. They reduce confounding without producing random assignment.

Effect sizes. Standardized effects are reported where studies use them. An effect of 0.2 moves a person from the median to roughly the 58th percentile, 0.5 to the 69th, 0.8 to the 79th. These are illustrations of magnitude, not predictions for an individual, and effects on different outcomes in different populations are not comparable.

Four ratings, kept separate. Collapsing these into one word loses information, so each finding carries whichever apply:

Causality — how well the evidence rules out alternative explanations. High: convergent evidence that strongly rules out the plausible alternatives, whether through repeated randomized trials, an unusually clean natural experiment, or triangulation across independent designs that fail in different ways. Smoking is the clearest case of the third route. Moderate: reasonably consistent longitudinal or quasi-experimental evidence with real residual uncertainty. Low: predominantly observational, selected, heterogeneous, or sparsely replicated.
Size — how large the effect is.
Generalizability — how far it travels beyond the studied population, exposure, and period. A clean study of an unusual group can rate high on causality and low here.
Duration — how long it has been shown to last, which is usually shorter than assumed.

A well-run study can find a small effect and a weak one can suggest a large effect, and a clean result may not transfer. These are different questions. Not every finding has enough evidence to rate all four; where only one rating appears it is causality, and the others are unstated because they are unknown. Where a study measured only the immediate effect, that is stated rather than assumed to persist.

Method is given only where methods have been compared directly. Knowing that exercise helps is less useful than knowing what dose, but the two are separate questions and the second is far less often studied. Where a course of action follows from a finding by reasoning rather than by test, it appears as Practical implication and is labelled as an inference. Many findings have neither, and that is stated rather than filled in with plausible advice.

Temperament. Life satisfaction is about 30–40% heritable at a given time. Genetic influence also accounts for much of why people’s relative positions stay stable — one recent estimate puts 65–79% of that persistence down to stable genetic effects. But estimates of how much of the stable level is genetic vary sharply: older studies reported 70–90%, while a 2026 study of over 10,000 people, separating lasting differences from real change across eight years, estimated 17–37%. Disposition matters substantially; long-run wellbeing is not mostly genetically fixed.

Heritability describes differences between people, not how much one person could change, and says nothing about whether changing a job or a location would help.

Adaptation. Most events move life satisfaction sharply, and much of that reverses. Panels show anticipation before a wedding, a spike, and substantial return within about two years. Bereavement produces one of the largest measured shocks: in a long German panel, losing a partner or child dropped satisfaction by more than a standard deviation, with one analysis finding near-complete recovery at five years and another a residual gap of roughly 0.4 ladder points. Adaptation is real, larger than intuition suggests, and highly variable by exposure. The best-measured cases: marriage, roughly two years back toward baseline; bereavement, several years with recovery largely but perhaps not entirely complete; unemployment, no full return even after re-employment. Some persistent conditions show much less adaptation than discrete events do. Chronic pain and noise exposure are the best-evidenced cases; persistent loneliness likewise does not show the rapid return seen after most single events, though it has been studied less carefully. Other exposures have been studied with less comparable methods, and the rates are not standardized enough to tabulate. Some people also change durably for reasons unconnected to any single event.

Predicting your own response. People often overestimate how long and how strongly a specific event will affect them, though the size of that error depends on how the questions are asked.

Averages hide people. Money does different things at different points in the distribution. The parenthood gap varies by country and childcare provision. Retirement depends on whether it was chosen. Averages are starting points, not forecasts.

Two kinds of uncertainty. Uncertainty about the average shrinks with better research. Uncertainty about your own response does not. Repeated personal experience can legitimately override an average — someone who has lived through several dark winters and several bright ones has evidence about themselves. Predictions about a state never experienced carry much less weight, since that is where forecasting errors concentrate.

III. Major determinants and recurring conditions

Grouped by domain, with the better-established findings emphasised within each.

Health and longevity

Smoking. Continuing smokers lose roughly a decade of life expectancy against never-smokers. Quitting at 30 recovers close to all of it, at 40 about nine years, at 50 about six.
Causality: high, despite the absence of randomized trials, from convergent evidence across designs, dose-response, and cessation studies. Size: the largest modifiable effect on lifespan available to most people. Generalizability: high.
Method: in a large Cochrane synthesis, roughly 6 in 100 attempts succeed without cessation aids, about 12 with two forms of nicotine replacement combined, and about 14 with varenicline or nicotine-containing e-cigarettes. Gradual reduction and abrupt quitting produce similar long-term quit rates. Repeated attempts are the norm rather than a sign of failure.

Physical activity. Moving from none to roughly 150 minutes a week of moderate activity is associated with 20–30% lower all-cause mortality, plus lower rates of cardiovascular disease, diabetes, several cancers, and later loss of function. Returns are steepest at the bottom.
Causality: high for several health and functional outcomes; moderate for the exact mortality effect, which rests on observational data. Size: large.
Practical implication: because benefits appear across many forms of activity and the steepest gain is from none to some, adherence is likely to matter more than fine optimisation of type for most people. This is an inference from the shape of the curve rather than a comparison of adherence strategies.

Alcohol and drugs. The finding that light drinking protects the heart has been substantially undermined by studies using inherited differences in alcohol metabolism, which find risk rising from low intake. Severe substance-use disorders are associated with large harms across health, functioning, and the subjective measures.
Causality: high for many of the direct health harms of dependence; moderate on the low-intake alcohol risk curve; lower for the size of the effect on subjective measures specifically.
Risk generally rises with consumption and no general threshold has been identified below which alcohol is risk-free; dose-response curves differ by outcome, with cancer risk beginning at low intake. The treatment evidence differs by substance and does not exist for all of them. For alcohol use disorder, adding medication to psychosocial treatment is supported; whether adding psychosocial treatment to medication helps is less clear. For opioid use disorder, medication is the central treatment and outperforms psychosocial treatment alone; psychosocial support can accompany it, but its absence should not delay medication.

Chronic pain. Among the larger negative associations across the subjective measures, with limited recovery over time.
Causality: moderate to high; longitudinal rather than randomized. Size: large.

Prevention and treatment. Age- and risk-appropriate medical care has far better evidence for health outcomes than generic wellbeing optimisation. These are separate literatures with different sizes and harms, grouped only by being well studied and unglamorous.

Mental illness

Depression and anxiety show large prospective associations with the other measures simultaneously. Their causal contribution to those downstream outcomes cannot be randomized. What is established is narrower: effective treatment reduces symptoms and improves functioning (Part V).
Causality: high for treatment effects; moderate for the disorder’s causal contribution to everything else. Size: large.

Time

A modest effect on an experience repeated every day may matter more across a year than a larger but short-lived one, where adaptation to the recurring exposure is incomplete. Standardized effects cannot literally be summed across days; this is a reason to attend to recurring conditions, not an arithmetic. Recurring conditions are not exempt from adaptation — annoyance at a commute tends to decline, annoyance at noise tends not to.

Sleep. Within-person studies link short or disrupted sleep with worse mood the next day. Experimental sleep restriction and insomnia treatment carry the causal claim.
Causality: high, resting on the experiments. Size: moderate, daily.
Method: a component analysis of insomnia trials identifies sleep restriction (compressing time in bed toward actual sleep time, then extending), stimulus control (bed used only for sleep, leaving it when awake), and cognitive approaches as the best-supported elements, with cognitive restructuring showing the largest incremental contribution. Sleep hygiene education — caffeine, screens, room temperature — contributes little on its own, despite being the advice most often given. Light exposure and consistent timing are tools for shifting the body clock rather than components of insomnia treatment.

Commuting. Longer commutes are consistently experienced as unpleasant and predict lower satisfaction with time and with the journey. Longitudinal estimates for overall life satisfaction are mixed, and people select into commutes alongside job and housing moves.
Causality: low to moderate for life satisfaction; higher for daily experience.
Practical implication: because commuting recurs daily, commute time is worth pricing explicitly in hours per year when comparing jobs or homes. Mode comparisons exist but are strongly selected, and no study establishes that shortening a commute beats improving it.

Work and autonomy. Job strain — high demands with low control — consistently predicts depression. Evidence for a causal effect on heart disease is weaker: a 2026 review estimated a 14% association and judged it insufficient for confident causal attribution.
Causality: moderate for mental health, low for cardiovascular disease. Size: moderate, across a third of waking life.

Caregiving. Sustained care of a dependent relative predicts reduced wellbeing and elevated depression, especially without relief.
Causality: moderate.
Method: multi-component programmes combining skills training, respite, and support have randomized evidence of reducing carer depression, where information alone does not.

Buying time. In a within-person experiment with 60 adults, each spending $40 on a time-saving purchase one weekend and a material purchase another, the time-saving weekend produced better mood, standardized effect about 0.33. Observational samples show the same association with life satisfaction.
Causality: low to moderate; one small study, no evidence on repeated use. Size: small to moderate on the day.

Relationships

Among the more consistent observational correlates of the subjective measures, and prospectively associated with mortality. Evidence that relationships matter is much stronger than evidence on how to create them.

Isolation and loneliness differ. Across about 90 cohorts and 2.2 million people, social isolation was associated with roughly 32% higher all-cause mortality and loneliness with roughly 14%.
Causality: moderate; ill and unhappy people also become isolated.

Quality, not status. Relationship quality is more consistently associated with wellbeing than marital status. Chosen solitude is not the exposure these studies measure.

How many. No optimum has been established, and the question is worse posed than it looks: loneliness is a gap between wanted and actual contact rather than a count, which is why people with many contacts report it and people with few often do not. Descriptive work puts a typical inner circle near five people and the wider active network near 150. These are observed regularities in particular populations, not targets, and neither number has been shown to matter for wellbeing.
Causality: low. No study supports a target number.

Maintaining existing relationships is a different question from curing loneliness, and better evidenced. Friendships decay measurably without periodic contact; distance predicts declining closeness; long hours predict reduced non-work contact; people who move away show reduced contact that does not fully recover.
Causality: low to moderate, observational throughout.
One study of recent movers and university students gives a rough scale for how much time friendship formation appeared to require in those samples: around 40–60 hours to casual friendship, 80–100 to friendship, and over 200 to close friendship. Observational, specific to those groups, and a scale rather than a formula.
Practical implication: the cost of a move or a demanding job includes network decay. That contact predicts closeness is well supported; no particular maintenance schedule has been tested.

Material resources

The relevant variable is the whole material position — income, wealth, debt, liquidity, income volatility, housing and job security — not income alone.

Income. Within a country, higher income predicts higher life satisfaction across the range studied, roughly logarithmically: each doubling of income adds a similar increment, so the same sum matters far more lower down.

Income and daily affect. An early finding put satiation around $75,000 a year in US incomes of the late 2000s; later in-the-moment sampling found no ceiling. A subsequent collaboration between both groups reconciled them: for the least happy 15–20% of people, wellbeing plateaus above roughly $100,000 on that study’s US income measure at that time; for the rest it continues rising. The figure is a feature of one sample, currency, and period, not a general threshold.

Wealth. Large-prize lottery winners in Sweden, followed 5 to 22 years, showed sustained higher life satisfaction that did not fade, with much smaller and often indistinguishable effects on daily affect and mental health. This concerns unearned wealth; earned income arrives bundled with hours.
Causality: high; among the cleanest designs in this document. Size: moderate and durable, larger for life satisfaction than for daily affect. Generalizability: moderate to low — Swedish lottery players receiving a windfall are not a general population receiving earned income. Duration: demonstrated 5 to 22 years after winning.

Security, separately from level. Income volatility, debt, illiquidity, and housing and job insecurity predict distress beyond average income, and financial strain often predicts mental health better than income does.
Causality: low; observational, and distress also worsens financial management. Size: unclear; associations can be substantial for severe strain.

Synthesis. Money reliably improves life satisfaction, can improve daily affect, and has diminishing value as resources rise. Evidence comparing uses of money — security, autonomy, time, consumption, status — is much thinner than evidence on income itself. It does not establish that how money is spent matters more than how much there is.
Practical implication: the only use with a direct experimental result behind it is buying time. The popular advice to buy experiences rather than possessions has replicated inconsistently and should not be relied on.

Environment

Country. Differences in life satisfaction are large — roughly 7.5–7.8 at the top, under 2 at the bottom — and persist after adjusting for income. Migrants converge partly toward destination levels.
Causality: moderate.

Climate. Probably overestimated. Students in Southern California and the Midwest reported essentially the same life satisfaction while both groups predicted Californians would be happier. That is good evidence that a salient climate difference gets overweighted when imagining another place, not evidence that climate has negligible effects generally.

Noise. Transportation noise causes persistent annoyance and sleep disruption, with little adaptation. Each additional 10 decibels is associated with roughly 2–3% higher heart-disease incidence.
Causality: moderate. Size: small for an individual, meaningful across a population. Implication: worth checking before choosing a home, though quieter locations may cost more or lengthen a commute.

Unemployment

Panels following people through job loss find life satisfaction falling by 0.5–0.7 ladder points. Most of that is recovered on re-employment, but not all: fifteen-year panels find people who have been unemployed sitting durably below their own prior level even once back in work, on the order of 0.2 ladder points. The fall during the spell exceeds what income loss explains, and does not shrink with its duration. Wage scarring runs alongside it for a decade or more. Plant-closure studies show the same pattern.
Causality: high. Size: large during the spell; small but apparently permanent afterwards.
Job insecurity among people still in work also predicts distress, so security and employment are not the same variable.

Leisure

Active against passive leisure. Time-sampling and day-reconstruction studies find large differences in how activities feel in the moment, with active and social leisure rating well above passive leisure. Sustained hobbies correlate with higher meaning and lower depression, and receptive arts engagement correlates with lower mortality in UK cohorts.
Causality: low. Activity choice is strongly selected, and people who are well take up hobbies. No particular hobby has been shown superior to another.

Travel and holidays. Wellbeing generally improves during a holiday. Older reviews concluded the benefit faded within two to four weeks of returning; a larger 2025 meta-analysis found bigger effects and slower fade-out, with holiday length moderating the result. Psychological detachment from work and physical activity during the trip were associated with larger benefits.
Causality: moderate; one of the better-measured before-and-after effects, though the literature has recently shifted.
Practical implication: holidays function as temporary recovery periods, and detaching from work appears to matter. The evidence does not establish that several short trips beat fewer long ones.

IV. The big choices

Partner. Panels show a rise around a wedding and substantial return toward baseline within about two years, with wide variation inside that average. Married people live longer, with a larger gap for men, though selection accounts for part of it. Chronic conflict predicts worse outcomes than being single. Causality: low, because happier and healthier people marry more. Marital status is the weaker variable: relationship quality is far better evidenced. The costs are autonomy, geographic flexibility, and exposure to a bad separation, among the larger negative events in the panel data.

What predicts relationship quality. The most informative study pooled 43 longitudinal datasets covering about 11,000 couples and used machine learning to rank predictors. A person’s own relationship-specific perceptions — perceived commitment from their partner, appreciation, sexual satisfaction, perceived partner satisfaction, low conflict — predicted quality far better than demographics or the partner’s reported traits. Once those perceptions were known, measured partner characteristics added little.

Two limits on what that shows. It does not establish that partner choice matters little; it establishes that the partner characteristics researchers measured did not improve prediction beyond a person’s own experience of the relationship. And the same data predicted change in relationship quality poorly, using every variable available — which is arguably the more striking result.
Causality: moderate. Longitudinal but not experimental, and self-reported perception carries much of the weight.

What predicts divorce. Contempt and its relatives — criticism, defensiveness, withdrawal during conflict — are the most consistently replicated behavioural predictors, though the widely cited accuracy figures for predicting divorce from short observations come from models fitted to the same couples they predict and should be discounted. More robust demographic predictors: marrying very young, short courtship, low income or education, a partner’s parents having divorced, and remarriage, which fails more often than first marriage. In US data, divorce risk falls steeply with age at marriage up to the late twenties and is roughly flat or slightly rising after the early thirties. The pattern is cohort- and country-specific.
Causality: moderate for the demographic predictors, low for behavioural ones as causes rather than symptoms.

What happens over time. Marital satisfaction tends to decline over the first years rather than holding steady, with the steepest drop after a first child, and steeper for mothers than fathers. Perceived unfairness in the division of domestic work is one of the more consistent correlates of that decline. In observational studies of North American couples, sexual frequency is associated with higher satisfaction up to roughly weekly and not beyond. This describes an average pattern across couples, not an optimum for any of them, and it is not a target.
Causality: low to moderate; all observational.

Whether anything helps. Couples therapy produces substantial improvements among distressed couples — a meta-analysis of 58 studies found large gains generally maintained at follow-up, though long-term data are thinner than immediate post-treatment data. Premarital and relationship-education programmes produce smaller effects, around 0.3, that fade. Both are better evidenced than most of the advice literature and much less used.

In observational work, what distinguishes stable from unstable couples is conflict conduct rather than conflict frequency: contempt and withdrawal separate them, disagreement itself does not. Expressed appreciation is among the strongest correlates of satisfaction, and perceived fairness of domestic work predicts the post-childbirth decline better than the actual division does. These are correlates rather than tested interventions.

Children. Parenthood is more consistently associated with meaning than with life satisfaction. Effects on life satisfaction and daily affect vary substantially with children’s age, income, partnership, gender, and country, and the gap is smaller where childcare and leave provision are stronger. Causality: low, since people who want children have them. The costs are money and a large number of hours during the intensive years. The central uncertainty is unusually deep, because having children can change the preferences by which the choice is later judged.

Career and hours. Autonomy, security, commute, and relationships at work each predict aspects of how work is experienced. Pay also matters; the independent role of prestige is much less clear. No single model has ranked these against one another. Unemployment risk is the largest work-related exposure in Part III. Additional income bought with substantially more hours trades one measured outcome against another rather than producing a net gain. Causality: moderate for unemployment and low for the rest, since people sort into jobs. Whether work that feels meaningful to a particular person can be predicted in advance is not answered by any of this.

Where to live. Location affects proximity to friends and family, commute, housing cost and quality, safety, pollution, noise, job opportunities, and amenities simultaneously. Cross-country differences are large and partly travel with migrants; causal evidence for moves within rich countries is weak. Ordinary climate differences appear less important than people predict. The main modifiers are the social network and the work. Moving away from established relationships reduces contact durably, while a move can improve housing, commute, or career. The evidence can price the components; it cannot say how a place will feel.

Housing. Crowding, noise, safety, commute, and financial strain are the housing variables with evidence attached. Square footage and status are not. Causality: low, though the noise and commute components inherit whatever rating those findings carry. Housing is usually the largest financial decision a person makes and often fixes the commute for years.
Practical implication: the variables with evidence attached can be checked cheaply before committing — noise measured rather than asked about, the commute priced in hours a year. This follows from which components have evidence, not from any study of how people should view houses.

Retirement. More control over time; less structure, identity, and daily contact. Causality: moderate where pension eligibility rules create quasi-random variation in timing, low otherwise. Voluntariness is the dominant modifier. The costs are income and a large recurring block of social contact that is rarely rebuilt deliberately.

Caregiving. Sustained reductions in wellbeing and elevated depression, especially without relief. The need arises exogenously; who absorbs it does not. Relief, duration, and whether the role was chosen or fell to one person by default are the main modifiers. The costs are hours, career, and often the carer’s own health, and the duration — the variable the decision most depends on — is usually unknown at the outset.

V. What you can deliberately do

Intervention estimates are larger against waiting lists than against active comparison groups, and larger in small or lower-quality studies. The figures below emphasise higher-quality estimates where available.

Psychotherapy for depression. A review of over 400 CBT trials estimated 0.79 overall, 0.60 among trials at low risk of bias, and 0.47 after adjusting for publication bias. Earlier work using the strictest filters produced figures near 0.20.
Causality: high that it works; moderate on the size.
Duration: roughly half of treated patients relapse within two years, though relapse is lower after therapy than after stopping medication, which is the main argument for it.
Method: different established therapies perform similarly in head-to-head trials, so the choice of school matters less than getting a trained practitioner and completing a course, typically 12–20 sessions. Behavioural activation — scheduling rewarding activity — performs comparably to full cognitive therapy and is simpler. Measuring symptoms session by session modestly improves outcomes, particularly for patients who are not improving. How long to persist before changing approach or practitioner is a matter of clinical judgement rather than an established rule.

Antidepressants. Around 0.30 against placebo. Whether the effect is larger in severe than mild depression is disputed rather than settled.
Causality: high. Size: small to moderate.
Duration: among people who respond, continuing medication substantially reduces relapse — roughly 21% relapsing on continuation against about 40% after discontinuation in pooled maintenance trials.
Method: an adequate trial is four to six weeks at a therapeutic dose before judging it. Persistent non-response warrants reassessment and a change may be appropriate, but switching has not been shown generally superior to continuing the same drug; dose optimisation, augmentation, and adding psychotherapy are alternatives, and augmentation may suit partial response better than switching. Stopping abruptly produces discontinuation symptoms.

Insomnia treatment. A 2024 network meta-analysis estimated long-term remission at roughly 41% for strategies beginning with CBT for insomnia, against 28% in arms starting with medication, at a median 24-week follow-up. The 41% is derived by applying the network’s relative effect to the medication baseline rather than being two directly observed proportions.
Causality: high. Size: large for sleep, moderate for mood.
Duration: gains largely hold at twelve months, unusually for this literature.

Exercise, mindfulness, yoga, and positive-psychology programmes in people without a diagnosis. A 2026 preregistered network meta-analysis of 183 trials found these clustered around 0.41–0.49, with 71% of comparisons rated moderate certainty. Most of the network anchors to inactive controls, so these figures include the non-specific effects of structured activity, expectation, and attention. For major depression, a 2024 analysis of 218 studies found moderate effects for walking, jogging, yoga, and strength training against active controls, with low certainty because of bias.
Causality: moderate. Size: small to moderate. Generalizability: moderate; most trials recruit volunteers willing to be assigned to a programme. Duration: mostly measured at or shortly after the programme ends, and evidence for benefits persisting afterwards is limited.

Gratitude exercises. A 2025 preregistered meta-analysis of 145 papers and about 24,800 participants found 0.19, with checks generally supporting a small positive effect. A daily fifteen-minute practice is roughly 90 hours a year, which is the relevant cost.
Causality: moderate. Size: small.
Duration: fades within months where it has been followed up.
Method: one study that varied frequency directly found weekly practice outperforming several times a week, apparently because the exercise loses effect as it becomes routine. That is a single comparison rather than a settled finding.

Couples therapy and relationship education. For distressed couples, structured therapies produce effects around 0.5 or larger at the end of treatment. Preventive education for non-distressed couples produces around 0.3.
Causality: moderate. Size: moderate for treatment, small for prevention.
Duration: meta-analytic evidence suggests therapy gains are often maintained at follow-up, though long-term data are limited; education effects fade.

Loneliness interventions. An older review of befriending schemes found essentially no pooled effect, on five comparisons. A 2025 preregistered meta-analysis of 280 studies found reductions around 0.5 in randomized trials, holding at one to six months, rated low or very low certainty. Psychological approaches showed larger pooled estimates than schemes that only increased opportunities for contact, though that compares across studies rather than head to head.
Causality: moderate that something works short-term; low on which, for whom, or for how long.
Duration: held at one to six months. Nothing is known beyond that.
Approaches targeting how a person interprets social situations showed larger pooled estimates than those supplying more contact. That is a comparison across studies rather than a head-to-head test.

Prosocial spending and volunteering. Small experimental effects on mood with mixed replication; larger observational associations with meaning that are heavily confounded.
Causality: low.

Time in nature. Improves short-run mood against doing nothing. Whether nature adds anything beyond the exercise, the change of scene, or the company is uncertain; comparisons of outdoor with identical indoor activity are much less impressive.
Causality: low.

Cash transfers in poor countries. Reliably improve reported wellbeing, but the averages are small: about 0.13 for wellbeing and 0.07 for mental health at roughly two years. A review restricted to randomized trials puts depression and anxiety near 0.10, with differences possibly undetectable two to nine years after a programme ends. Larger transfers do more.
Causality: high. Size: small on these measures, whatever the material value. Generalizability: low to moderate — these are low-income households in poor countries, and the result should not be transferred to transfers in rich ones.
Duration: attenuating, and possibly undetectable two to nine years after a programme ends. Money that clearly improves a household’s material position shifts reported feelings only modestly, which is a useful reference point for how small population-level effects on these measures tend to be.

VI. Where evidence cannot choose

Claims the evidence does not justify

That more income is always worth more work.
That prestige independently improves life, once income, autonomy, job quality, and comparison are separated. Unresolved rather than disproved.
That larger or more expensive housing improves wellbeing.
That climate alone is a major determinant.
That most self-help techniques have large durable effects.
That one-off events matter as much as forecast; adaptation is substantial.
That several short holidays beat fewer long ones. The fade-out evidence has recently shifted and no such rule is established.

What evidence cannot value

Meaning against pleasure, and whether a life high in one and low in the other goes well.
Health against enjoyment. Alcohol, food, sleep given up for social life, risky sport, long hours on work that matters, health optimisation that consumes the time it adds. Consequences on each side can be estimated. What a year of additional life is worth against the quality of the days spent obtaining it cannot be.
Money against time. Additional income usually costs hours, and hours are where daily affect lives.
Parenthood. More meaning, harder days, and life satisfaction that depends on circumstances. Both answers are correct about different things people want.
Duties to others. A person can become happier by behaving badly. Data about their wellbeing cannot settle whether the life is good.
Achievement, character, contribution, and understanding, which many accounts treat as valuable in themselves. Studies can ask whether these produce satisfaction; that does not settle whether their value consists in producing it.
Truth against pleasant illusion. Most people say a life resting on false beliefs went worse, which is a fact about preferences rather than an answer.
The present self against the future self, even holding preferences fixed.
Adaptation itself. People in poor conditions often report high satisfaction. Either conditions matter less than assumed, or self-report fails where it matters most. This is a disagreement about what wellbeing is.

What evidence cannot predict well

Individual responses. Averages conceal wide variation, and the characteristics that would let someone apply an average to themselves are usually unidentified.
Endogenous preferences. Parenthood, migration, religion, careers, and difficult projects change what a person values, so the future self who evaluates a decision is partly created by it.
Tail risks. A choice worth +0.1 on average with a 1% chance of disaster is not summarised by its average, and most consequential decisions have this shape.
Reversibility and option value. Two choices with identical expected effects are not equally attractive if one can be tested and undone. Renting before buying, trying a job before committing, preserving financial flexibility. Under high uncertainty a reversible option can dominate one with a better estimated average. Wellbeing studies do not incorporate this.
Cross-cultural comparability. Country rankings assume scales mean the same thing in every language.

Findings with no method attached

The asymmetry between the two questions is one of the document’s main results. For many outcomes, whether something matters is far better established than how to produce it. Comparative evidence on method exists for smoking cessation, insomnia treatment, depression treatment, and carer support. For most of the rest it does not, and filling that gap with plausible advice is where popular writing goes wrong.

Meaning. Strongly measured as an outcome, barely studied as something that can be produced.
Making new close relationships in adulthood. Loneliness can be reduced; building a confidant from nothing has almost no intervention evidence.
Finding work that feels worthwhile. Autonomy and security predict how work feels, but nothing predicts which work will feel meaningful to a particular person.
Choosing a partner. The best evidence says the relationship matters more than the partner’s traits, which is unhelpful advice at the point of choosing.
Which hobby, which place, which career. The evidence identifies features that tend to help and does not rank the options.

VII. What follows

Avoid large harms first. Serious untreated disease, smoking, addiction, chronic pain, severe financial insecurity, destructive relationships, unemployment, and involuntary isolation deserve priority over marginal optimisation.
Protect health and functioning. Physical and mental health affect both longevity and the subjective measures. Smoking is measured in years of life; nothing in the intervention literature operates on that scale, and the two should not be traded off by comparing numbers.
Attend to recurring time use. Sleep, work, commute, relationships, caregiving, and exercise occupy thousands of hours.
Secure enough money and flexibility. Money matters most at low resources. Beyond that, count what additional income costs in hours, stress, and freedom.
Protect existing relationships. Their value is better established than the ability to rebuild them once lost.
Expect adaptation. Many changes matter less over time than they do at first. Do not assume the same of conditions experienced daily.
Match the cost to the evidence. Small effects can justify cheap actions. They rarely justify expensive or irreversible ones.
Distinguish knowing that from knowing how. For most of what matters here, the evidence establishes that something counts without establishing how to obtain it. Where a tested method exists — quitting smoking, insomnia, depression, carer support — use it. Elsewhere, treat confident procedural advice with suspicion, including the inferences drawn above.
Use averages as starting points. Update with known modifiers and with repeated evidence from your own life.
Prefer reversible experiments under uncertainty. Preserve option value where the long-run effect is hard to predict.
Recognise where evidence ends. Data can estimate consequences. It cannot decide how much weight to give money, time, pleasure, meaning, achievement, freedom, responsibility, or risk.

The evidence is most useful for escaping bad states, moderately useful for designing recurring conditions, and least useful for choosing among already-good lives. Better research can improve the predicted consequences of those choices. It cannot supply the values used to rank them.