Parenting

Scope and thesis

This document is about one question: how much can ordinary differences in parenting change how a child turns out, and at what cost? “How a child turns out” covers several different things, and the evidence is very different for each:

Health and safety in childhood: injury, infection, poisoning, harm before birth.
Childhood experience itself: whether daily life is safe, warm, and interesting, or frightening.
Specific skills and behaviours in childhood: language, reading, self-control, tantrums, sleep.
Circumstances and opportunity: neighbourhood, school quality, income, treatment of problems.
Broad adult outcomes: personality, intelligence, happiness, mental illness, education, earnings.
The adult relationship between parent and child.

Evidence is generally stronger, and effects larger, for health and safety, childhood experience, specific childhood skills, and circumstances than for broad adult outcomes. This is partly because the chain from a parenting decision to an adult outcome is long and passes through many later influences. Something can reliably change a child next month and barely predict their life at 35. That does not make the near-term change unimportant; it means near-term and long-term claims need separate evidence. The adult relationship is a case of its own: it is influenced throughout life, and the evidence on it is consistent but almost entirely based on people looking back.

The conclusion the evidence supports is this. Parents have considerable power over children’s health, safety, daily experience, circumstances, treatment of problems, and the relationship, and some power over specific childhood skills and behaviours. They have much less demonstrated power to engineer broad adult traits or to guarantee success. Once the major conditions are adequate, evidence for expensive further optimisation is weak.

What the evidence does not support is the stronger claim that ordinary parenting has little causal importance. That claim is sometimes drawn from twin studies, and section 1 explains why it does not follow.

How numbers are reported. Two ratings are given separately for each item: how confident we can be that the effect is causal (high, moderate, low), and how big it is. A well-designed study can find a small effect and a weak study can suggest a large one, so the two should not be collapsed.

Sizes follow one of two conventions, chosen by the kind of outcome:

Rare or yes-or-no outcomes, such as a disease, an injury, or a diagnosis: the rate without the exposure and the rate with it, so the actual change in risk is visible. A halving from 2% to 1% and a halving from 40% to 20% are very different things and should not both be reported as “50% lower.”
Continuous outcomes, such as test scores, language, or behaviour ratings: where a child who would otherwise be exactly average, at the 50th percentile, ends up instead. This is an approximate way of picturing an effect size, not a claim about a real scale. Effects combined across studies are usually measured on different instruments, so the percentile is a translation of the average, not a prediction for any individual child. Each conversion comes from a single estimate, not a blend of studies.

Where a claim cannot be quantified honestly, it says so rather than borrowing a number from a different outcome.

One limitation applies throughout. Almost all of this evidence comes from rich Western countries, and much of it from the United States and northern Europe. Where a finding depends on a particular school system, health system, or level of inequality, it may not transfer.

Sections 2 to 6 are ordered from what parents can most clearly change to what they least can.

1. What twin, adoption, and genetic studies show, and what they do not

The main finding. An analysis pooling 2,748 twin studies, covering 17,804 traits and 14.5 million partly overlapping twin pairs, found that on average about 49% of the variation in a trait was associated with genetic differences. The shared-environment contribution was on average small, and for about two-thirds of traits the pattern of twin resemblance was consistent with little or no shared-environment effect at all.

What the three components are. These studies split the differences between people on a trait into three parts: genes; shared environment, meaning whatever non-genetic influence makes siblings raised together more alike than genes alone would predict; and non-shared environment, meaning whatever non-genetic influence makes them differ, plus measurement error. The parts are estimated from how much identical twins resemble each other compared with non-identical twins, and how much adopted children resemble their birth and adoptive families. Nothing is measured directly.

What the shared part is not. It is not “parenting,” and it is not “the home.” It is whatever siblings share that makes them alike, which includes the school they both attended, the neighbourhood, the family’s income and social position, the local culture, and the parents’ behaviour toward both children. The model cannot tell these apart. Everything below therefore calls it shared environment, not shared home, and a large figure for it is not by itself evidence that parents did anything.

Parenting can also fall into any of the three parts. If parents treat two children the same and it makes them more alike, that is shared. If parents respond differently to different children, or the same treatment affects two children differently because of their temperaments, that is non-shared. If parents provide an environment partly because of their own genes, and their children inherit those genes, some of the effect is counted as genetic. Studies of twins find that children’s genetically influenced characteristics partly shape how they are parented. So a small shared-environment figure does not mean parenting has a small causal effect. It means the ways ordinary families differ from each other do not, on average, make siblings more alike on that trait.

Parents pass on genes and environment together. The genes a parent has and the environment they provide are correlated. Studies using the parental genes a child did not inherit find that those genes still predict the child’s educational attainment, through the environment the parents provide. Pooled across studies, this indirect route is roughly half the size of the direct genetic effect for education. Some of it reflects the wider family’s social position rather than the household itself. The question “genes or home?” is partly artificial: the home is itself partly a product of the parents’ genes.

What the estimates assume. The split depends on assumptions: that identical and non-identical twins experience environments that are equally similar in the ways that matter; that who parents choose as partners is accounted for; that genes and environments do not interact in ways the model cannot represent; and that the trait is measured reliably. Estimates differ by sample, age, measurement method, and country. They are model-based estimates, not physical measurements, and rounded figures conceal real disagreement.

The genetic share depends on the population. Some US studies find much lower genetic variation in cognitive ability among children raised in low-income, low-education families; this pattern is not evident in Western European and Australian studies.

One possible interpretation: where access to developmentally important environments is more equal, there is less environmental variation left to explain differences between children, so a larger share of what remains is attributed to genes. Where access varies a lot, more is attributed to environment. Countries differ in how much a poor child’s circumstances vary — in school funding, healthcare access, and income support — and that is a plausible reason the pattern might appear in one country and not another.

The cross-national studies are consistent with this interpretation. They do not establish that any particular social policy causes the difference, and other explanations remain live: how socioeconomic status is measured, differences in the ages and composition of samples, and the possibility that the US finding is itself unstable. Note also that a high genetic share never means the remainder is trivial; non-shared environmental differences, developmental chance, and measurement error can all still be large.

This should be held loosely. The pattern is not found in every US study, socioeconomic status is measured differently across countries, and the samples differ in age and composition. It illustrates that the genetic share describes the population studied rather than a fixed property of the trait. It does not by itself show that the home explains more among poor families.

A high genetic share is not predictability. The best current genetic predictors of educational attainment explain about 12–16% of the differences between people. Comparing siblings within the same family, the direct effect of those same genes is roughly half that, because part of the between-family association is the family environment travelling with the genes. DNA therefore permits statements about groups, not precise prediction of an individual child’s eventual attainment.

A high genetic share is not a probability. Schizophrenia has a genetic share of about 80% and affects about 1% of people. A child of one affected parent has roughly a 10% risk; an identical twin of an affected person roughly 40–50%. About two-thirds of people who develop it have no affected parent, sibling, grandparent, aunt, or uncle. A high genetic share means genes explain most of the difference between who gets a condition and who does not; it does not make the condition likely for any given child, or mean it visibly runs in families.

Approximate estimates for common outcomes. Ranges across large studies; genes / shared environment. For diagnoses, the figures describe variation in underlying susceptibility to the condition, not the condition itself. “Little detected” means studies have not reliably found a shared-environment effect, not that one has been proved absent.

Continuous traits (adults unless stated):
Cognitive ability, childhood: 20–45% / 25–40%
Cognitive ability, adulthood: 60–80% / 0–10%
Years of education: 30–50% / 25–40%
Personality, five main traits: 40–50% / little detected
Life satisfaction: 30–40% / little detected
Religiousness, adolescence: 10–20% / 40–60%
Religiousness, adulthood: 35–50% / 10–25%

Diagnoses and behaviours (susceptibility):
Autism: 65–90% / 0–15%
ADHD: 70–80% / little detected
Schizophrenia: 65–85% / 0–15%
Bipolar disorder: 60–85% / little detected
Antisocial and criminal behaviour: 35–50% / 10–20%
Major depression: 30–40% / little detected
Anxiety disorders: 30–45% / little detected
Eating disorders: 40–60% / 0–15%

Substances. Whether a person ever tries something and whether they become dependent are different outcomes with different answers:
Ever trying a substance (any of the below): 30–55% / 20–40%
Nicotine dependence: 55–75% / 0–10%
Cocaine use disorder: 60–80% / 0–10%
Opioid use disorder: 45–70% / 0–15%
Cannabis use disorder: 45–60% / 5–15%
Alcohol use disorder: 45–55% / 5–15%

Three comparisons worth making across the whole list.

Same outcome, different age. Cognitive ability: the shared component explains 25–40% of the differences between children and close to nothing between adults. Religiousness: 40–60% in adolescence, 10–25% in adulthood. Shared environmental influences are larger earlier in life and shrink as people age and increasingly select their own environments. Whether the shrinking part is parents, schools, or something else is not something this model can say.

Same domain, different stage. The shared-environment component is two to four times larger for whether a child ever tries a substance than for whether they become dependent on it. Whatever siblings share — availability at home, family rules, local norms, what their school and friends do — shows up in first use and largely stops mattering for dependence, which looks far more like individual susceptibility. The decomposition cannot say which of those shared factors is responsible, but the contrast between the two stages is sharp and consistent.

Same person, different kind of outcome. The shared-environment component is substantially larger for years of education (25–40%) than for adult personality, life satisfaction, depression, or anxiety, where it is not reliably detected at all. On its own this says only that families make siblings more alike in schooling than in temperament; it does not identify parents as the cause. But read alongside the direct evidence in sections 3 and 4 on schools, neighbourhoods, and family resources, the pattern is that family circumstances have more demonstrated leverage over a child’s education and opportunities than over how they end up feeling as adults. Section 8 is about what follows from that.

Non-shared environment. For most adult outcomes this is the larger environmental part. Researchers have had limited success identifying the specific experiences that make up this part; in a well-known review, each measured candidate explained on average only about 1–2% of the variation. That may mean the influences are many and small, or it may reflect poor measurement of environments, measurement error in outcomes, interactions with the child’s characteristics, or developmental chance. What follows is that deliberately predicting or engineering these differences is difficult. What does not follow is that they lie outside parenting or cannot be changed.

A few have been identified, using a design that gets around genetic confounding: comparing identical twins who differed in one experience. On that design, the twin who was bullied, who had a serious head injury, or who was sexually abused does measurably worse than their genetically identical sibling. These are the clearest examples of what non-shared environment actually contains, and several of them are things parents can act on. They are covered in section 5.

Adoption. Children adopted out of deprived homes into well-off ones end up with more education, better health habits, and higher incomes than siblings who stayed; in some small French studies, IQ was 10 or more points higher. Large changes in circumstances change outcomes, even where variation among ordinary families explains little.

Average effects and individual effects. Every figure in this document is an average. A small average effect can contain many children who do not respond and some who respond a great deal; a near-zero average does not prove the effect is zero for a particular child. But unless the characteristics that predict response are known in advance, this cannot be used to justify a decision for a specific child. It is a reason for humility about what the averages rule out, not a licence to assume one’s own case is the exception.

What this section supports. Within the range of ordinary, adequately functioning homes, there is no demonstrated method by which parents can reliably engineer a particular adult personality, intelligence level, degree of happiness, or susceptibility to depression or anxiety. The same holds for a particular career or level of success: parents can change opportunity and specific skills, but not reliably convert either into a chosen adult outcome. This does not mean these things are immovable — severe deprivation, abuse, and lead all move them — and it does not show that parenting has no causal effects. Sections 2 to 5 show that it does.

2. What parents can clearly change

Exposures before birth.
Evidence: folic acid before and during early pregnancy. In a randomised trial among women who had already had an affected pregnancy, supplements reduced the recurrence of neural tube defects by about 72%. In the general population, where the baseline risk is roughly 1 in 1,000 pregnancies, adding folic acid to flour was followed by reductions of roughly a quarter to a half. Smoking in pregnancy roughly doubles the risk of low birth weight, from around 7% to around 13% in US data. Heavy drinking in pregnancy is associated with IQ deficits of several points and, at the extreme, fetal alcohol spectrum disorders.
Causal confidence: high for folate and smoking; high for heavy alcohol, less certain for light drinking.
Size: large reductions in uncommon but serious outcomes.

Lead.
Evidence: across seven long-term studies of children, a rise in blood lead from a low level (2.4 micrograms per decilitre) to the level once treated as acceptable (10) was associated with about 3.9 fewer IQ points, with further losses at higher levels; the relationship is continuous and no safe level has been found. Where lead was removed from the environment, later crime fell.
Causal confidence: high.
Size: several IQ points across the range of exposure common in older housing.
Where: in the United States, lead paint in homes was banned in 1978, so the relevant housing is pre-1978. In the United Kingdom, the main risk is lead water pipes in homes built before about 1970. Testing is cheap in both.

Abuse and neglect.
Evidence: physical abuse, emotional abuse, and neglect are associated with roughly 1.5 to 2 times the rate of adult depression, anxiety, and drug use, and with substantially less education. Comparing siblings and identical twins who differed in their exposure gives the same direction, which rules out the explanation that it is simply families with worse genes.
Causal confidence: high.
Size: large and lasting.
Implication: preventing abuse and neglect is among the most consequential things under parental control.

Health and safety. These sit outside the main scope of this document and belong ahead of nearly everything below it. Most are cheap, most are one-off, and they act on rare but severe outcomes. This is a checklist, not a substitute for medical advice.

Any age.

Vaccination. Modelling estimates about 154 million deaths prevented worldwide over 50 years, most of them infants. Before vaccination, measles alone killed roughly 2.6 million people a year.
Car seats. A correctly fitted, age-appropriate seat cuts serious injury risk by about 70–80% relative to seat belts alone. Fitting errors are found in a large share of seats when checked, and checks are usually free.
Second-hand smoke. Roughly 20–40% higher rates of asthma and ear infections, and roughly double the rate of sudden infant death.
Sun. Five or more blistering sunburns in adolescence is associated with roughly double the lifetime risk of melanoma, against a baseline of about 2% in fair-skinned populations.

Infancy.

Safe sleep. On the back, firm flat surface, no soft bedding, no shared sleeping surface. After the back-sleeping campaigns, sudden infant death in the US fell from about 1.3 to about 0.4 per 1,000 live births, a fall of roughly two-thirds.
Introducing allergens early. In a randomised trial of infants at high risk, introducing peanut rather than avoiding it cut peanut allergy at age five from 17% to 3%. Guidance in many countries changed as a result. Ask a doctor first if the child already has eczema or a known allergy.
Feeding. Breastfeeding roughly halves gastrointestinal infections and reduces respiratory infections by roughly a third in infancy. On cognition the evidence is weaker than usually claimed: the largest randomised trial assigned mothers who already intended to breastfeed to a promotion programme, not babies to breast or formula, and found a verbal advantage of about 7 IQ points at age six with a wide margin of error and unblinded testers, most of which had gone by sixteen. The case for breastfeeding rests on infection, not intelligence. Formula is not a developmental risk.
Newborn screening, hearing and vision checks. One to three infants per 1,000 are born with permanent hearing loss. Those identified and treated before about six months score far closer to normal on later language than those identified late, with differences of 20 percentile points or more. Uncorrected long-sightedness in a school-age child depresses reading directly and is fixed with glasses.

Toddlers and young children.

Drowning. Among the leading causes of death at ages one to four. Four-sided isolation fencing around pools reduces drowning risk by roughly 70–80%; formal swimming lessons reduce it further in young children. Neither replaces supervision within arm’s reach. Baths and garden ponds count.
Button batteries. A swallowed battery can cause serious internal burns within about two hours. Deaths are rare in absolute terms but have risen with the spread of the devices; the injuries are catastrophic and entirely preventable by storage.
Choking. Food causes over half of choking deaths in under-fours. Whole grapes, nuts, popcorn, and hard sweets are the main culprits.
Household hazards. Anchor tall furniture and televisions: tip-overs kill roughly one child a fortnight in the US alone. Window restrictors, blind cords out of reach. Set the hot water thermostat to about 50°C rather than 60°C: at 60°C a serious scald takes about five seconds, at 50°C about five minutes.
Poisons. Medicines, cleaning products, alcohol, cannabis, and nicotine liquid locked away rather than merely high up. Keep the poison helpline number to hand.
Speech and development. About 7% of children have a language disorder that will not resolve on its own. Referral is what a parent controls; waiting lists are long everywhere, so a delay of six months in asking becomes a delay of a year or more in help.
Teeth. Roughly a fifth to a third of children have visible decay by age five. Fluoride toothpaste and fluoridated water each reduce decay by roughly a quarter.

School age.

Iron, vitamin D, iodine. Iron-deficiency anaemia in infancy is associated with cognitive deficits of roughly 5–10 IQ points that persist after treatment, which is why prevention matters more than correction. Deficiency is common in some populations and cheap to prevent. Beyond avoiding deficiency, no particular diet has been shown to change long-term development.
Time outdoors. In a randomised trial in Chinese schools, adding 40 minutes of outdoor time a day cut the onset of short-sightedness over three years from about 30% to about 24%. The effect appears to come from daylight, not from what the child is doing outside.
Helmets. Roughly 60–70% lower risk of serious head injury in cycling. Repeated concussion in youth sport has measurable cognitive effects.

Adolescence.

Driving. Road crashes are among the leading causes of death at this age. Carrying one teenage passenger roughly doubles a young driver’s crash risk and two or more roughly quadruples it; graduated licensing schemes with night and passenger limits have reduced fatal crashes among sixteen-year-olds by roughly 20–40%.
Access to means. Where firearms are present, storing them locked and unloaded is associated with roughly 70–80% lower odds of adolescent shooting deaths, including suicide. The same logic applies to large quantities of medication kept at home. Restricting access to means is among the best-evidenced suicide-prevention measures there is, because adolescent crises are typically short-lived.

These rest on different kinds of evidence, and the confidence differs accordingly. Randomised trials: folate, early peanut introduction, outdoor time, and the breastfeeding-promotion trial. Strong observational and case-control evidence with a clear mechanism: car seats, helmets, pool fencing, safe storage, scald temperatures, choking, button batteries. Before-and-after population comparisons, which are compelling but cannot isolate which part of a campaign worked: safe infant sleep, graduated driver licensing. Modelling rather than measurement: the vaccination death toll. The first group has direct randomised evidence. Several items in the other groups nevertheless have strong causal support from converging observational studies, an understood mechanism, natural experiments, or effects far too large to be explained by confounding: nobody needs a trial randomising children to wear helmets or swallow batteries.

Responding to mental illness in adolescence.
What it is: noticing, and acting on, depression, anxiety, eating disorders, and self-harm, which typically first appear between about 12 and 25.
Evidence: about half of lifetime mental illness has begun by the mid-teens, and roughly one adolescent in seven meets criteria for a disorder in a given year. Treatment works: across randomised trials, cognitive behavioural therapy moves an average treated adolescent with depression or anxiety to roughly the 65th–70th percentile relative to untreated controls, and combined therapy and medication does better still for moderate to severe depression. Detection is one constraint among several: median delays between the onset of symptoms and any treatment run to several years and most affected adolescents never receive any, though availability of clinicians, cost, severity, and whether a young person sticks with treatment all matter too.
Causal confidence: high that treatment works; moderate for the value of early detection specifically.
Implication: the parental contribution is noticing changes in sleep, eating, withdrawal, and functioning; asking directly, which does not increase risk; and getting professional help rather than waiting for it to pass.

Treatment of a diagnosed condition.
Evidence: a diagnosed developmental, psychiatric, or learning condition changes the decision problem. Condition-specific treatments often have considerably stronger evidence than generic parenting advice. For example, behavioural parent training for children with ADHD shows small-to-moderate benefits that persist when trials are followed up years later. Evidence differs substantially between conditions; for autism, early-intervention trials are inconsistent and many have design weaknesses.
Causal confidence: varies by condition; moderate to high for the best-studied treatments.
Implication: the evidence summarised here for typically developing children should not be used to argue against assessment or treatment.

3. What parents can change modestly, mainly in childhood

Why effects here usually shrink, and why that is not the whole story. Almost everything in this section has been measured shortly after it happened. Where the same children have been followed for decades, the pattern is consistent and two-sided. Head Start’s advantage on test scores had largely gone by third grade, yet children who attended went on to more education and less adult poverty. Perry Preschool’s IQ gains disappeared within a few years, while its effects on graduation, earnings, and crime were still visible in middle age. So a faded test score does not prove nothing happened, and a large early gain should not be projected forward as though it will hold. The right expectation for anything below is that the measured effect shrinks substantially, and that whatever survives is unlikely to be the thing that was measured.

Parenting programmes in early childhood.
What they are: usually home visits or group sessions over several months, coaching parents to talk, respond, and play with a baby or toddler in specific ways, often with books or toys provided.
Evidence: across 102 randomised trials, average effects were large enough to move a child at the middle to roughly the 63rd percentile in cognitive development, the 61st in language, the 59th in motor development, the 58th in social and emotional development, and the 61st in secure attachment to caregivers, with a small reduction in behaviour problems. Most trials were in poorer countries or with disadvantaged families; follow-ups were mostly short.
Causal confidence: high for near-term effects; low for adult outcomes.
Size: small to moderate.

Parent training for disruptive behaviour.
What it is: typically eight to sixteen weekly sessions, in a group or with a therapist, teaching parents to give clear instructions, praise specific behaviours, ignore attention-seeking ones, and apply consistent non-physical consequences. Widely available; often free through health services.
Evidence: randomised trials consistently reduce disruptive behaviour straight after treatment. Across trials that followed families afterwards, improvements were generally maintained for up to three years. A 2026 analysis combining 64 trials found that immediately after treatment, an average preschooler whose parents took a parent-only programme had less disruptive behaviour than roughly 73% of untreated children, and an average school-age child less than roughly 65%, with much weaker evidence that this holds over time. Praise and non-violent logical consequences are the components most associated with larger effects. Where outcomes are reported by the parent who received the training, they can be inflated: in ADHD, effects that looked large on parent report shrank substantially when the people rating the child did not know who had been treated.
Causal confidence: high for effects immediately after treatment; moderate for persistence over the next few years; low for adult outcomes.
Size: moderate immediately, smaller thereafter.

Shared reading.
What it is: reading picture books with a young child interactively, asking questions and prompting the child to talk rather than reading straight through. Trials usually train parents over a few sessions and supply books.
Evidence: controlled trials improve vocabulary and language, with effects large enough to move an average child to roughly the 58th–66th percentile shortly afterwards.
Causal confidence: high for near-term language; no evidence on adult outcomes.
Size: small to moderate.

Self-control training.
What it is: classroom curricula, mindfulness sessions, or structured games that practise waiting, switching attention, and following rules; usually delivered at school over weeks or months.
Evidence: across 50 randomised trials, programmes improved measures of self-control enough to move an average child to roughly the 66th percentile. Follow-ups were short and many outcomes were reported by parents or teachers who knew what the child had received.
Causal confidence: moderate for near-term; low for lasting change.
Size: moderate in the short term.

Sleep.
Evidence: two different questions. First, whether enough sleep matters: school-age children need 9–12 hours and adolescents 8–10, and where schools moved start times later, adolescents gained roughly 30–60 minutes, with effects on grades positive in some studies and absent in others. Reported academic gains, where found, are on the order of 2–5 percentile points. Second, whether parents managing sleep changes outcomes: randomised trials of structured settling routines for infants and toddlers, in which parents are taught to put a child down awake and respond on a set schedule, resolve sleep problems in roughly 60–70% of cases against 40–50% without, and roughly halve rates of maternal depression at four months; evidence of lasting effects on the child is weak.
Causal confidence: high that sleep matters; moderate for short-term effects of parental sleep programmes; low for long-term child outcomes.

Organised activities.
What they are: school programmes teaching emotional skills, sports and physical activity, and after-school clubs.
Evidence: well-run school programmes for social and emotional skills raise achievement by about 11 percentile points; physical activity produces small gains in maths and reading; after-school programmes help when well run, and reviews find quality varies widely.
Causal confidence: moderate.
Size: small. Nothing shows that more activities help more.

Praise, consequences, routines.
Evidence: when parenting programmes are broken down into their parts, praise and non-violent logical consequences are the parts associated with better behaviour outcomes (above). There is no evidence that any of these produce large differences in adult outcomes.
Causal confidence: moderate for childhood behaviour; none for adult outcomes.

4. What changes circumstances and opportunity

Neighbourhood.
Evidence: a US experiment gave families in high-poverty public housing vouchers, assigned by lottery, to move to lower-poverty areas. Children who moved before age 13 earned about $3,477 more a year in their mid-twenties, 31% above the control group’s roughly $11,300, and attended college at higher rates. Children who moved as teenagers did slightly worse. A separate study of five million families who moved between areas estimated that each year of childhood in a better area closes about 4% of the gap between areas.
Causal confidence: high for this population.
Size: large for children leaving severe disadvantage, and dependent on years of exposure.
Implication: for families leaving severely disadvantaged neighbourhoods, earlier moves are substantially more beneficial than moves in adolescence. There is no equivalent experimental estimate for moves between generally good neighbourhoods.

Income.
Evidence: studies of tax credits and casino payments to poor families generally find children benefit. A randomised trial giving low-income mothers in four US cities $333 a month rather than $20 found no statistically significant effect on any of four outcomes chosen in advance, measured at age four: language, self-control and planning, social and emotional problems, and brain activity. It did reduce measured poverty and change some spending.
Causal confidence: mixed.
Size: modest on average; cash alone appears to shift less than correlations suggest. This is one size and design of payment, not a general result.

Childcare.
Evidence: Quebec’s cheap universal childcare expansion in 1997 was followed by lasting increases in anxiety, aggression, and hyperactivity, lower health and life satisfaction in young adulthood, and higher crime rates, concentrated among boys; about 5% of providers were rated high quality at launch. Norway’s expansion raised educational attainment and adult earnings, most for children of less-educated mothers. The programmes differed in ages covered, hours, what care they replaced, and rollout, so these studies do not isolate quality as the cause. A large US study following children over time found no advantage to full-time parental care over childcare; higher-quality care went with slightly better language and academic results, and very long hours with slightly more behaviour problems.
Causal confidence: moderate to high that the Quebec expansion produced adverse effects; moderate for why.
Implication: a childcare expansion can produce lasting adverse outcomes; quality is a plausible but unproven explanation for the contrast with more successful programmes. There is no robust evidence of a large average advantage of parental care over competent childcare.

Teachers.
Evidence: in a study of over a million children, having a teacher whose measured effect on achievement was well above average (in the top sixth) for one school year predicted about 1.3% higher earnings at age 28.
Causal confidence: high, given the study’s design.
Size: individual-year effects are modest but can accumulate across a classroom and a school career; the authors regard highly effective teachers as economically important.

Schools, including private schools.
Evidence: in PISA 2022 mathematics, averaged across OECD countries, private-school pupils scored about 24 points higher than state-school pupils, which is roughly a year of schooling on that scale; after adjusting for the socioeconomic profile of pupils and schools, state-school pupils scored about 11 points higher. Adjusted comparisons depend on what is controlled for and can over- or under-adjust. US voucher lotteries have produced test-score effects from strongly negative (Louisiana, a loss of roughly a third of a year in maths in the first year, partially recovering) to none (Washington DC), with some gains in graduation. Those experiments involved disadvantaged children entering particular participating schools.
Causal confidence: moderate.
Size: no demonstrated average advantage of private provision as such. Specific schools differ, and the features that matter (teachers, peers, expectations, safety, fit) have to be judged directly.

Home education.
What it is: teaching a child at home instead of sending them to school, ranging from highly structured tuition to almost nothing.
Evidence: much weaker than for any other schooling comparison. There are no randomised or lottery-based studies, and most widely quoted results come from volunteer samples recruited through home-education organisations, where families opt in to being measured. Those families are on average wealthier, better educated, and more likely to have a parent at home, which predicts good outcomes regardless of schooling. The few studies that match children on background find structured home education performing at or slightly above school on academic measures and unstructured home education performing below, but the samples are small. On social outcomes, no consistent average deficit has been found. On adult outcomes, results are mixed and mostly from religious homeschooling populations that may not generalise.
Causal confidence: low, on all of it. This is one of the least well-evidenced major decisions a parent can make.
Size: unknown on average. The important number is the spread, not the average: the range of what “home education” means is far wider than the gap between its average and school’s, and in many jurisdictions nobody checks which end a given family is at.
Implication: the framework in this document says judge the provision, not the label, and home education is the extreme case. A decent school and very good home education are probably not far apart on academic outcomes, and the evidence cannot resolve which is ahead. What differs is what has to be supplied deliberately rather than arriving by default. School supplies, without anyone arranging it, several things section 5 identifies as mattering: a peer group, the chance of a close friendship, a good teacher some years, and an adult outside the family who takes the child seriously. It also supplies detection. Teachers are the largest single source of referrals when something is wrong at home, and home education removes that. None of this argues against home education. It argues that a family choosing it is taking on the job of providing those things on purpose, and that the honest comparison is not “school versus home” but “this school versus what this family will actually do.”

Parental mental health and functioning.
Evidence: parental mental illness, substance misuse, and severe instability alter children’s environments substantially. Maternal depression is the best-studied case. Children of depressed mothers have roughly two to three times the rate of depression themselves. In follow-ups of maternal treatment trials, diagnoses among children fell by about 11% where the mother’s depression lifted and rose by about 8% where it did not, over a year. Child outcomes were not themselves assigned by lottery.
Causal confidence: moderate.
Implication: treating a parent’s condition is likely to help the child; the evidence for fathers is thinner but there is no reason to expect it to differ in direction.

Parental conflict and divorce.
Evidence: children of divorced parents do somewhat worse on average, by roughly 5–8 percentile points on academic and behavioural measures, and the gap has narrowed as divorce has become more common. Children in high-conflict intact households do worse than children of low-conflict separations. Conflict is the most consistently supported mechanism, alongside loss of income, moving, and changes in parenting, and it is also correlated with parental traits that could independently affect children.
Causal confidence: low to moderate.
Size: small on average, and highly variable between families.
Implication: the evidence gives little justification for preserving a high-conflict household solely because intact households do better on average.

Parental engagement with school.
Evidence: across studies of nearly half a million families, parental expectations for a child’s education show the strongest association with achievement of any involvement measure, at roughly 8–12 percentile points, while attending school events shows almost none. Direct homework help correlates with worse results, largely because struggling children get more help; that negative association should not be read causally.
Causal confidence: low, since all of this is observational and expectations track many other things.
Implication: engagement and expectations have more consistent support than routine direct involvement in completing homework. Whether a particular child would benefit from help is not answered by these studies.

5. Experiences outside the home

Section 1 said the differences between siblings raised together are usually larger than the ones they share, and that researchers have struggled to identify what those differences consist of. This section is what has been identified. It is short because the honest list is short.

The strongest design compares identical twins who differed in one experience. Because the pair share their genes and everything about the family they both grew up in, an association cannot be explained by genetic differences between them or by family-wide factors. It is not an experiment: a pre-existing difference between the twins could produce both the exposure and the outcome, and reverse causation and measurement problems remain. These are strong designs, not decisive ones. Items using them are marked as higher confidence.

Bullying.
What it is: repeated victimisation by peers, usually at school.
Evidence: roughly a third of children report being bullied at some point and about 10% frequently. The bullied twin has more anxiety, depression, and self-harm years later than their genetically identical sibling: in the largest such study, self-harm by age twelve occurred in about 8% of frequently bullied children against about 3% of others, and the difference persisted within twin pairs. Depression at eighteen is roughly doubled, from about 6% to about 12%, among those bullied frequently at eleven. Effects scale with duration and remain detectable into the twenties.
Causal confidence: moderate to high.
Size: moderate and lasting; among the larger identified non-shared effects.
Implication: what is supported is knowing what is happening, taking a first report seriously, and removing a child from a situation that is not being resolved. What is not established is that any particular parental response works better than another once it has started.

A close friendship.
What it is: at least one reciprocated, good-quality friendship.
Evidence: among bullied children, those with at least one good friendship show roughly half the later emotional harm. Friendship quality predicts better adjustment generally, at roughly 5–8 percentile points, and children without a reciprocated friendship at age ten are several times more likely to report loneliness and depressive symptoms in adolescence.
Causal confidence: moderate; almost entirely observational, and children who are already doing better make friends more easily.
Size: moderate as a buffer against bullying; small as a general effect.
Implication: this argues for protecting the conditions friendships form in, meaning stability, unstructured time, and not over-scheduling, rather than managing friendships directly, which has no evidence behind it.

One reliable adult who is not a parent.
What it is: a coach, teacher, relative, family friend, or formal mentor who sees the child regularly over a sustained period.
Evidence: in a randomised trial of formal mentoring, over 18 months mentored children were about 46% less likely to start using illegal drugs (roughly 11% of controls started, against 6% of the mentored), about 27% less likely to start drinking, about a third less likely to hit someone, skipped about half as many days of school, and reported better relationships with their own parents. Observational work finds similar patterns for informal mentors, though selection is a real problem there.
Causal confidence: moderate to high; unusually for this list, from a randomised trial.
Size: moderate. Probably the largest positive effect in this section.
Implication: make such relationships possible and do not resent them.

Peers.
What it is: classmates, friendship group, neighbourhood social circle.
Evidence: studies using randomly assigned classmates and university roommates find effects on grades of roughly 2–5 percentile points, and larger effects on behaviour: having a randomly assigned peer group that drinks raises a student’s own drinking substantially, and delinquency spreads more strongly than achievement does. Programmes that group high-risk adolescents together can make behaviour worse rather than better; several trials have found increases in problem behaviour of roughly 5–10 percentage points relative to controls.
Causal confidence: moderate to high, where assignment was random.
Size: small for achievement, moderate for behaviour.

Delaying substance use.
What it is: age at first regular drinking or cannabis use.
Evidence: about 15% of people who begin drinking before age 15 develop alcohol dependence, against about 5% of those who begin at 18 or later, roughly a threefold difference. Cannabis before about 17 is associated with lower rates of school completion. In twin comparisons, where the earlier-starting twin is compared with their identical sibling, the gap narrows by roughly half but does not disappear, so much of the raw association is family background and some of it is not.
Causal confidence: moderate.
Size: moderate.
Implication: delay is the variable with evidence behind it, more than any particular conversation about drugs.

Sport and physical activity.
What it is: regular participation, organised or otherwise.
Evidence: participating adolescents report depressive symptoms at roughly 10–25% lower rates than non-participants, with team sport slightly better than individual sport. Randomised physical-activity programmes move an average child to roughly the 55th percentile in maths and reading. Much of the observational association runs backwards, since happier and healthier children take part more.
Causal confidence: low to moderate.
Size: small.

Social media and night-time phone use.
What it is: specific patterns of use, as distinct from total screen time, which barely predicts anything.
Evidence: broad measures show little. All screen use combined explains about 0.4% of the variation in adolescent well-being, roughly the size of the association with eating potatoes. Narrower measures show more but still modestly: the heaviest social media users, particularly girls in early adolescence, report depressive symptoms at roughly 1.5 to 2 times the rate of light users against a baseline of roughly 10–15%, and studies of social media arriving in a population find falls in mental health of a few percentage points. Experimental evidence has since accumulated: pooling randomised studies that asked people to cut back on social media finds small average improvements, moving a typical participant to roughly the 57th percentile on well-being and the 60th on depressive symptoms. Most of those participants were young adults rather than adolescents. Phone use that displaces sleep remains the firmest link, because the sleep effects are independently established. Gaming with existing friends looks neutral.
Causal confidence: low to moderate, and this is the fastest-moving area in the document.
Size: small on average; larger for heavy use and for the sleep pathway.
Implication: total hours is a poor target. Reducing social media specifically has small average benefits in trials, and keeping phones out of the night has the clearest case of anything here.

Serious illness, head injury, and abuse. All show effects in twin-difference designs. Abuse is covered in section 2, head injury in the safety checklist there.

Two conclusions. First, the identified influences a parent can affect are mostly about exposure and access: which school, which peers, whether an outside adult is allowed to matter, whether early substance use is delayed, whether sleep is protected. Second, even the well-identified ones are individually small, and they do not add up to the large hidden factor that would explain why siblings differ. Some of what makes siblings differ is demonstrably arbitrary: children who are youngest in their school year are diagnosed with ADHD markedly more often and picked for selective sports less often, purely because of birth month. Decades of searching have not turned up a large, plannable factor, though the search is limited by how poorly environments are measured.

6. Things people often overestimate

Each entry is classified by what the evidence shows. “No effect found,” “the effect runs the other way,” “the claimed advantage is absent,” and “nobody has tested it” are four different findings, and treating them alike is how weak claims survive.

Evidence of roughly zero effect.

Birth order on personality. Large studies find no meaningful effect; there is a small IQ effect of about 1–3 points favouring first-borns.
Family size, on children’s outcomes. The raw association with worse outcomes largely disappears once you account for which families have more children.
Being an only child. No deficit in social skills or adjustment has been found.
Differences among ordinary functioning homes, on adult personality and life satisfaction. Section 1.

Popular claims that did not hold up.

Praising effort rather than ability to build a “growth mindset.” The original small studies did not replicate at scale; large trials find effects near zero on average, with possible small benefits for struggling students.
Learning styles. No evidence that matching teaching to a child’s supposed visual, auditory, or kinaesthetic style improves learning, despite repeated testing.
The “30 million word” gap. The original finding came from 42 families; larger studies find a real but far smaller difference in words heard. Talking with a child matters; the number was wrong.
Classical music raising intelligence. The original effect was a brief arousal effect on one spatial task, not an increase in intelligence, and it does not apply to infants.
Sugar causing hyperactivity. Repeatedly tested in blinded trials and not found. Parents who believe a child has had sugar rate them as more hyperactive.
Educational screen media for under-twos. Randomised trials find no language benefit from video, including from products marketed for it.

Claimed advantage not found.

Full-time parental care over competent childcare. Exclusive maternal care did not predict better outcomes in the main US study; care quality and hours had modest associations. Section 4.
Private schooling as such. Adjusted comparisons and voucher lotteries do not show an average advantage. Section 4.
Single parenthood as a cause in itself. Children of single parents do somewhat worse on average, but income, conflict, and selection account for most of it; the household count is not the active ingredient.

Evidence runs the other way. Not merely unproven; the better studies point to harm.

Grouping troubled adolescents together for treatment. Randomised trials find problem behaviour increases relative to controls. Section 5.
Bed-sharing with an infant, whatever its other merits. It raises the risk of sudden infant death, particularly alongside soft bedding, sofas, smoking, or alcohol. Section 2.
Switching to a private school on the assumption it must be better. In the one large lottery study, children lost roughly a third of a year of maths in the first year. That was a specific set of participating schools serving disadvantaged children, so it is a warning rather than a general result. Section 4.
Physical punishment. Consistently associated with worse behaviour, not better, and a worse relationship later, though almost none of that evidence is randomised. Section 8.
Expanding childcare quickly without ensuring quality. Quebec is the warning case: a rapid universal expansion, with about 5% of providers rated high quality at launch, was followed by lasting increases in anxiety, aggression, and crime. The study did not vary quality experimentally, so the lesson is about how expansion is done, not a general claim that more childcare harms children. Section 4.

No demonstrated advantage, and a real opportunity cost. These are not shown to cause harm directly; they consume something that does have evidence behind it.

Doing homework with a child rather than expecting them to do it. Correlates with worse results, but much of that is because struggling children get more help, so this is not evidence of harm. It is evidence that the benefit people assume is not there. Section 4.
Adding activities past the point where sleep or free time suffers. No study shows more activities help more, and the sleep and friendship time they displace do have evidence behind them. Sections 2, 3, and 5.
Paying substantially more for a school that is only slightly better on the features that matter. Section 4.

Not well tested for lasting effects.

Attachment parenting as a package, and baby-wearing.
Home education, in either direction. Widely claimed to be better and widely claimed to be worse; neither is established. Section 4.
Raising children in several countries.
Moving between two already-good neighbourhoods.
An elite private school rather than a good one.
Leaving a career solely for the child’s development.
Any particular parenting style producing a particular adult personality.
Letting children experience consequences to build judgement.
Exposing children to many different things so they find what suits them.
Birth spacing, and specific routines, chores, or reading schedules beyond the near-term effects in section 3.

Cost. The expensive items above deserve arithmetic rather than intuition. Thirteen years at a UK independent day school is roughly £300,000–£400,000 at current fees, and the same money invested would be worth more than that by the time the child leaves. Leaving work costs not only income but promotion, pension, and accumulated experience. Neither decision is wrong, and both are legitimate on preference alone. But given the evidence above, they should be made knowing that no measurable developmental return has been demonstrated in exchange.

7. What the evidence does not tell us, and how much room that leaves

Four things could in principle be true even though the studies have not detected them. Each is a real limit. But each is also constrained by evidence, and none leaves as much room as it first appears.

Rare practices. Something used by a few percent of families contributes almost nothing to population-wide variation even if it works well for those families, so it would not show up. What constrains it: if a rare practice were powerful, we would expect to see it when researchers deliberately teach average parents to do better. Parenting programmes do exactly that, and they produce modest effects that mostly do not persist into adulthood (section 3). Rare and powerful is possible; it has not been found, and it has been looked for.

Child-specific parenting. Getting two different children’s different needs right produces an effect counted as non-shared environment, which is exactly the part that is hardest to detect. What constrains it: comparing identical twins whose parents treated them differently, the twin treated more harshly or with less warmth does worse — so within-family differences in parenting do have detectable effects, and they are consistently small. Whatever child-specific parenting does, it is not large enough to show up as large in the one design built to find it.

Outcomes nobody measures. Judgement, values, taste, specific skills, the adult relationship, and whether a childhood was any good are rarely the outcome variable. What constrains it: nothing, really. This is the genuine gap, and it is why section 8 exists. But note the asymmetry: it is a reason to believe parents matter in ways not captured, not a reason to believe any particular expensive intervention works.

Which children respond. A small average can hide some children who benefit a great deal. What constrains it: to act on that, a parent would need to know in advance which children respond, and studies looking for those characteristics have generally not found reliable ones. The heterogeneity is real; the ability to exploit it is not established.

Taken together: the escape routes exist, but the evidence has been pointed at three of the four and has not found much hiding there. The honest position is not “the studies missed it,” but “several ways it could have been missed have been checked, and the effects found were small.”

8. Childhood, and the relationship, as outcomes in themselves

Childhood is an outcome. Everything above measures childhood by what it predicts later. But a childhood is also lived. Whether tonight’s dinner involves affection, conversation, and humour, or fear, humiliation, and violence, matters in itself, and parents have very large influence over it regardless of what twin studies say about adult traits. Estimates of the genetic share of adult personality do not make that irrelevant, empirically or morally.

The relationship is an outcome. Most parents would want a close relationship with their adult children even if it changed nothing else. It is a legitimate end, and one parents influence directly, from the start, and for life.

Estrangement. How common it is depends heavily on definition. In one US survey, about 27% of adults reported current estrangement from some family member and about 10% from a parent or child. A nationally representative study following people over time found about 6% had ever been estranged from their mother and 26% from their father, with many later reconciling. Estrangement from a parent is associated with worse mental health in the adult child.

What adult children report as reasons. This evidence comes from surveys and interviews in which adults look back. It identifies what they say drove closeness or estrangement; it cannot establish which parenting practices causally prevent estrangement. Consistently reported:

Perceived rejection or lack of warmth in childhood.
Harsh or physical punishment. Across 75 studies, almost none of them randomised, spanking is associated with somewhat worse behaviour and worse parent–child relationship quality.
Control through guilt, shame, or withdrawal of affection. Across studies covering over 126,000 children, this form of control shows a moderate correlation with worse well-being, and support for the child’s growing independence a moderate correlation with better well-being; limits on behaviour do not show the same pattern.
Favouritism between siblings, as perceived by the child, in studies of adult siblings and their parents.
Continued control into adulthood, and rejection of the adult child’s partner, career, or beliefs, in surveys of estranged adults.
Loss of contact after divorce, most often with fathers, in studies of adult children of divorce.

Causal confidence: low. Consistency across studies: high.

Implication. Treat childhood experience and the relationship as outcomes, not instruments. Weigh them against measured developmental advantages, not below them. The behaviours associated with a lasting relationship do not require money.

9. Open questions, and what would change these conclusions

Every conclusion here should come with a statement of what would move it. This section pairs the main claims with the evidence that would substantially change them, and notes where that evidence is being generated. None of these is a single decisive test: a synthesis resting on dozens of studies should shift on replicated findings, not on one paper.

Claim: parenting differences within ordinary families do not much affect broad adult traits.
What would substantially change it: several large studies using genetically informed designs — comparing siblings, or using measured genetics as a control — that finds substantial effects of measured parenting on adult personality or mental health after genetic confounding is removed. Or an intervention trial with genuine adult follow-up showing large effects.
Where to watch: population registries in Nordic countries, which link whole populations across generations and increasingly include genetic data; sibling-comparison analyses of large cohorts.

Claim: non-shared environment is mostly small, unpredictable influences.
What would substantially change it: better measurement finding a large, stable factor. Almost all attempts so far have used retrospective questionnaires. Continuous measurement — phones, wearables, repeated sampling through the day — measures environments far better than a survey does, and it has barely been applied to this question. A substantial part of the “non-shared” component may simply be measurement error, and better instruments would show that.
This is the single most likely place for the picture to change.

Claim: early-childhood and parenting programme effects fade.
What would substantially change it: long-term follow-ups of the current generation of trials. Most existing evidence stops within a few years. Several large trials from the 2000s and 2010s are now reaching adulthood.

Claim: cash alone does less than the correlations imply.
What would substantially change it: longer follow-up. The main randomised trial has so far reported at age four. Age-six data collection is complete and age-eight follow-up began in 2026. If effects appear once children reach school, the current null looks like a matter of timing rather than a result.

Claim: quality matters more than who provides childcare.
What would substantially change it: evaluations of newer large-scale childcare expansions, including current programmes in Canada and England, which operate in different regulatory and service environments from 1990s Quebec. These are the closest thing to a replication the question will get.

Claim: social media has small average effects.
What would substantially change it: randomised restriction trials in adolescents specifically, rather than the young adults who make up most existing trials, and work that separates use patterns rather than measuring total hours. This is the fastest-moving evidence in the document and the conclusion most likely to be out of date soonest.

Claim: the genetic share of cognitive ability varies with a family’s circumstances in the US but not Europe.
What would substantially change it: further replication attempts, which have so far been mixed. If the pattern repeatedly fails to appear in new US samples, one of the more interesting claims in section 1 goes away.

Claim: bullying, mentoring, and peer effects are real and causal.
What would substantially change it: for bullying and mentoring, the twin-difference and randomised evidence is reasonably firm. For peers, most estimates come from schools and universities with random assignment, which may not generalise to how children actually form friendships.

Questions nobody has answered, and probably should.

Does moving between two good neighbourhoods do anything? Every credible estimate concerns families leaving severe disadvantage.
Is home education better or worse than a decent school, for a comparable child? A fast-growing choice with almost no usable evidence, because the families who do it differ so much from those who do not and, in many places, nobody records outcomes at all.
Do parenting-programme effects on behaviour translate into anything in adulthood, or do they simply make childhood easier? Both would be worth knowing, and only one is usually studied.
Does anything a parent does after a child is bullied help? Enormous literature on prevalence, almost none on parental response.
What actually accounts for the shared-environment component in education? It could be schools, neighbourhoods, expectations, money, or all four; the twin model cannot say.
Which children respond to which interventions? Averages dominate the literature, and the characteristics that would let a parent apply them to their own child have not been identified.

A note on how this document should decay. The safety findings in section 2 are unlikely to change much. The genetics in section 1 will get more precise but probably not reverse. Sections 5 and 6 contain the claims most likely to look wrong in ten years, because they rest on the youngest and thinnest evidence. If you return to this, start there.

10. A decision rule

Prevent large harms: harmful exposures before birth, lead, abuse and neglect, unsafe sleep, unrestrained travel, non-vaccination.
Treat actual problems: a parent’s mental illness, a child’s diagnosed condition, emerging depression or anxiety in an adolescent, a high-conflict household.
Secure adequate conditions: a safe neighbourhood, competent childcare, a decent school, enough sleep.
Use modest interventions where they are cheap: reading, consistent praise and consequences, a few well-run activities, a parenting course if behaviour is a problem.
Protect the outside-the-home things with evidence: respond to bullying, leave room for close friendships and for an adult outside the family to matter, delay early substance use, keep phones out of the night.
Be sceptical of expensive claims resting on small or uncertain effects: elite schooling, leaving work for developmental benefit, enrichment beyond a few activities, relocation between good areas.
Beyond that, decide on the family’s own preferences and on the quality of childhood and the relationship. The evidence does not distinguish between the remaining options, and those two things are outcomes in their own right.