Ask a fifth grader to choose between a guaranteed 50 tokens and a 50-50 shot at either 20 or 200, and you’ll get an answer. Ask them again with slightly different numbers, and you may get an inconsistent one. Economists have long treated such inconsistencies as noise, but a growing body of work suggests they contain real information about how children think, guess, and process choices.
A new NBER working paper takes that question seriously, building a statistical model that tries to separate a child’s true preferences from the mental static surrounding them. The authors find that once you account for that static, many familiar claims about children and risk look different, including the long-standing observation that girls are more risk averse than boys.
Untangling preferences from noise
The paper is authored by Cesar Chavez Padilla of the University of Chicago’s Harris School of Public Policy, along with Shuaizhang Feng of Jinan University, James J. Heckman of the University of Chicago, and Zhe Yang of Liaoning University. Their goal was to figure out how a child’s IQ and personality shape their economic choices, and whether those traits act directly on preferences or work through some other channel.
The standard way to measure risk preferences is a multiple price list: a series of paired choices in which each option gets progressively better or worse. The point at which a person switches sides reveals how much variance they’re willing to tolerate. But this method assumes people understand each option clearly and choose accordingly. In reality, some choices are made after real deliberation, others are essentially guesses, and both can look identical in the data.
The authors built a structural model that treats observed choices as coming from three sources: genuine preferences, deliberation noise (imprecision in decision-making), and outright guessing. Each of these can depend on a child’s IQ, personality, and background. The model also allows shocks across a child’s choices to be correlated with one another, rather than assuming, as most prior work does, that each choice is independent.
The Mianzhu data
The empirical work draws on the Longitudinal Study of Children’s Development in Mianzhu, which follows roughly 7,200 students across 18 schools in Sichuan Province, China, from Grade 4 into high school. The analysis focuses on 2,204 fifth graders (1,115 boys and 1,089 girls) surveyed in 2019, who completed six incentivized risk-preference tasks on tablet computers, with real prizes at stake.
Cognitive ability was measured using Raven’s Standard Progressive Matrices, a 60-item nonverbal reasoning test. Personality was assessed by homeroom teachers using a 20-item version of the Big Five Inventory-2, covering conscientiousness, agreeableness, extraversion, openness, and emotional stability. The researchers used teacher reports rather than student self-reports because prior work in the same dataset found teacher assessments produced more predictive and reliable measures of children’s traits.
Family background information included urban versus rural residence, parental education, sibling status, and a detailed record of whether each parent had migrated for work, semester by semester, since Grade 1. This last measure matters in China, where large numbers of children are raised by grandparents while parents work in distant cities.
What the model revealed
Once the authors fit their model, several patterns emerged that reshape how these choices should be interpreted.
First, IQ and personality traits have relatively weak direct effects on preferences themselves. Instead, they operate mostly through deliberation noise and guessing. Higher IQ was linked to less guessing for both genders, and less noisy decisions for boys. Conscientiousness also sharpened decision-making for boys. In other words, traits don’t seem to make children fundamentally more or less risk-loving so much as they change how consistently and attentively kids engage with the choices in front of them.
Second, family background variables appear to work almost entirely through personality and cognition. Once IQ and Big Five traits are included in the model, socioeconomic factors have no additional direct effect on preferences or noise. But those same background variables do predict traits: parental education and household stability were linked to higher IQ scores and more favorable personality profiles, particularly for openness and conscientiousness.
Third, the assumption that choice errors are independent across items appears to be wrong. The authors found strong correlation in shocks within a child’s set of choices, consistent with persistent individual traits shaping decisions across the whole battery. Roughly 40 percent of the variance in deliberation noise reflected stable individual differences, not random per-item errors.
Fourth, guessing turns out to be a substantial and distinct source of choice mistakes. The estimated probability of guessing on any given item was about 14 percent for boys and 19 percent for girls. When the researchers decomposed the total utility lost to choice mistakes, guessing accounted for the largest single share, exceeding both random and persistent deliberation noise.
The gender gap, revisited
One of the paper’s more attention-grabbing findings concerns gender. Using a conventional model without correlated errors or guessing, the data reproduce the standard result that girls are more risk averse than boys. But that gap disappears once the authors add correlated errors, guessing, and psychological traits to the same functional form.
When they move to a more flexible utility function called Expo-Power, which allows risk aversion to change with the size of the stakes, a different pattern emerges. For small gambles, girls appeared less risk averse than boys. For larger gambles, the ordering reversed and girls became more risk averse. In the authors’ words, “the gender gap in risk aversion is a function of the stakes at risk rather than a single number.”
The researchers interpret this as evidence that the conventional story about female risk aversion partly reflects modeling assumptions rather than a stable behavioral fact, at least in this age group.
Boys and girls, different channels
The model also suggests that traits operate through different pathways for boys and girls. For boys, personality traits affected both preferences and the noise structure. Conscientiousness raised risk aversion, while openness flattened the way risk aversion changed with stake size. For girls, traits acted almost entirely on the noise structure and guessing rather than on preferences themselves.
The correlation between persistent deliberation noise and persistent guessing tendency also flipped signs across genders: negative for boys and positive for girls. The authors take this as a hint that boys and girls may employ different cognitive strategies when facing difficult choices, though they don’t push the interpretation further.
Caveats and context
Several limitations are worth noting. The sample is drawn from one region of China, and the children are all around ten or eleven years old, so extending the findings to other populations or ages requires caution. The model treats IQ and personality as exogenous inputs, which the authors argue is defensible for causal interpretation but which some readers may want to probe further. And while the Expo-Power function fits the data better than the standard constant-relative-risk-aversion form, the notion of “reference wealth” that appears in both models is not fully pinned down for children, an issue the authors flag and plan to address in follow-up work.
The methodological point that generalizes beyond childhood is that observed inconsistency in choices contains structure. Treating all of it as random error, the authors argue, conflates true preference variation with differences in the precision and engagement with which people make decisions. That distinction matters not only for children but for any population, such as older adults facing cognitive decline or survey respondents under time pressure, where deliberation itself varies systematically.
