What People Think They Keep
The baseline part of Dur, Harms, and Non’s Dutch experiment asks a prior question before any information is supplied: how accurately do working-age people perceive the effective marginal tax rate that applies to them? This chapter reads those answers as descriptive evidence. The important task is to separate the size of an error from its direction, then examine how both vary with income and confidence. That makes clear why the same information could make work look more attractive to one group and less attractive to another, without claiming that income alone determines what any individual believes.
Accuracy has size and direction
The study compares two rates for each respondent. One is the effective marginal tax rate calculated from income and household information: the share of an additional increment of gross income that does not become additional disposable income. The other is the respondent’s perceived rate, elicited by asking how much they would gain on net if gross monthly income rose by €100. The difference between these rates is the person’s misperception.
There are two useful ways to describe that difference. The signed error preserves direction. If someone believes their effective marginal tax rate is 70% when the calculated rate is 50%, the signed error is +20 percentage points: they overestimate the amount lost to taxes and withdrawn benefits. If another person believes the rate is 30% when it is actually 50%, the signed error is −20 points: they underestimate it. The errors have the same absolute size, but they point in opposite behavioral directions.
The absolute error removes direction and measures distance from the calculated rate. Both examples have an absolute misperception of 20 percentage points. This measure answers “How far away was the belief?” rather than “Did the person see the marginal return as too low or too high?” The distinction is essential. A program might improve average accuracy while producing different labor-supply responses among people whose errors move in opposite directions.
In the final analytical sample, the average calculated effective marginal tax rate was about 43%, while the average perceived rate was about 47%. That four-point difference in averages does not mean that the typical respondent was only four points off. Individual errors can cancel when signed differences are averaged. The study reports an average absolute misperception of 21.4 percentage points, indicating much greater dispersion in individual understanding than the two averages alone would reveal.
The researchers classify respondents into three groups so that the later treatment analysis can respect this direction. About 36.4% were overestimators, 40.4% underestimators, and 23.2% were judged to be about right. “About right” does not mean that a respondent reproduced an exact individualized calculation. The information treatment communicated ranges for broad income groups, so the classification allowed a five-percentage-point margin around the relevant range. Beliefs above the upper bound plus that margin counted as overestimates; beliefs below the lower bound minus it counted as underestimates; responses within the relevant band were treated as about right.
That procedure is a practical compromise. A precise marginal rate depends on a person’s income and household circumstances, while the information presented in the survey was intentionally coarse. The classification therefore should not be read as a timeless label attached to a person. It identifies whether the person’s first-wave answer was materially above or below the range that the intervention was designed to communicate. Changing the margin could change the number assigned to each category, although the paper reports that its main results were not sensitive to the magnitude of the error margin.
The categories also do not divide the population into knowledgeable and ignorant people. Someone classified as an overestimator may understand many parts of the tax system and still misjudge the combined effect of several credits and allowances. Someone classified as about right may have arrived at a reasonable answer without knowing the underlying rules. The categories describe the relationship between one belief and one calculated marginal rate at one point in time. They are useful for interpreting the experiment, not for judging personal competence.
A simple example shows why the sign matters. Suppose an additional €100 actually increases disposable income by €80, implying a 20% effective marginal tax rate. A person who thinks they will keep only €50 sees the extra work as generating a much smaller net return than it does. If information moves that belief toward €80, the person may reconsider an additional hour. Another person who expects to keep €95 may discover that the return is less generous than assumed and may reconsider in the other direction. Accuracy is the shared objective, but the possible work response depends on where the mistaken belief started.
The diagram’s “possible” response is deliberate. Baseline misperception is not itself evidence that a person changed hours because of the error. It identifies the channel through which a correction could matter. A person may prefer more leisure even when the net wage is high, may be unable to change their schedule, or may not treat a €100 monthly increment as relevant to a real job decision. The baseline data establish the belief gap; they do not yet identify its behavioral effect.
The pattern follows income—but not perfectly
The clearest descriptive pattern is an income gradient in the direction of the error. Respondents with gross monthly income below €2,000 overestimated their effective marginal tax rate by about 28 percentage points on average. The middle-income group underestimated it by about 2.38 points, and the high-income group by about 4.40 points. The contrast is substantial: the lower-income group tended to see the marginal tax burden as much heavier than the calculated schedule implied, while higher-income respondents tended to see it as slightly lighter.
The pattern is economically consequential because an overestimate lowers the perceived net return to an extra unit of work. In the lower-income group, a person who believes that taxes and withdrawn benefits absorb far more of the next euro may decide that extra hours yield too little disposable income. If the person is wrong, the decision is not merely a response to a high official burden. It is a response to a lower return than the actual schedule provides. By contrast, a higher-income underestimator may perceive a somewhat more generous return than the schedule delivers.
But the income gradient is not a complete explanation. The paper reports over- and underestimation in every income category, and the distribution of errors has a mode at the point where beliefs are correct. There is variation among people with similar incomes because household circumstances and individual understanding differ. A gross monthly income band is not a household identity: credits and allowances can depend on children, housing, childcare, and other facts. Even with the same income, two respondents may face different calculated marginal rates and may interpret the system differently.
This matters for how the finding should be used. It is supported to say that low-income respondents were more likely, on average, to overestimate their marginal rate and that higher-income respondents tended, on average, to underestimate it slightly. It is not supported to say that low-income people uniformly misunderstand taxes, that high-income people generally understand them, or that income causes the direction of the error. The study measures a pattern of beliefs; it does not isolate whether the pattern comes from differences in tax exposure, household complexity, experience with benefits, financial knowledge, or some combination.
The absence of a single explanation is not a weakness to hide. It changes the appropriate policy response. If the same average error arose because a tax schedule was impossible to follow, broad communication might help. If it arose because people rarely face a decision at the relevant margin, information might be ignored. If it reflected different household-specific schedules, a national average explanation could be too vague. The experiment tests what happens when a general range-based explanation is offered, but the baseline results alone cannot tell us which design would correct every subgroup most effectively.
The confidence data add another layer. Respondents were asked how sure they were about their estimate on a five-point scale. In general, the study finds that larger absolute misperceptions were associated with lower confidence. That is a sensible pattern: uncertainty sometimes accompanies a difficult calculation. Yet the relationship was not a reliable self-correction mechanism. The roughly 4% who said they were completely certain were almost as far from the calculated rate as the roughly 10% who said they were completely unsure.
This exception is more revealing than a simple statement that confidence and accuracy are correlated. A respondent can be uncertain and wrong, but can also be highly confident and wrong. Confidence is information about a belief state, not proof that the belief is well-founded. For a policy that relies on people seeking clarification, that means the people least likely to ask for help may not be the people who need it least. The study does not show why the completely certain group was so inaccurate; it only reports the relationship between confidence and absolute error.
The result also cautions against treating a confident answer as a stable preference. A person can be certain that an extra €100 would mostly disappear and still revise that belief when given a credible range. Conversely, a person with low confidence may update but still face a constraint that prevents any change in hours. Belief accuracy, willingness to update, and capacity to act are related but distinct parts of the behavioral chain.
The baseline numbers come from the experiment’s final analytical sample, not from every person originally invited. The researchers began with a larger LISS-panel invitation pool, followed respondents across two waves, removed cases with missing income and extreme beliefs under their stated rule, and excluded the personalized-information arm from the main analysis. Those decisions are relevant when interpreting the composition of the sample. They do not erase the observed income gradient, but they do mean that the percentages describe the analyzed respondents under the study’s definitions, not a census of every Dutch household.
There is another boundary in the comparison between calculated and perceived rates. The calculated rate is produced from background information about income and household circumstances, while the perceived rate comes from a survey answer to a hypothetical increment. The two measures are designed to be comparable, but they are not identical kinds of observation. One is a model-based calculation of the schedule; the other is a reported belief about a scenario. The gap is substantively meaningful, yet it should not be treated as a direct measurement of a person’s complete tax knowledge or of every work decision they make.
Even so, the baseline evidence creates a sharp testable prediction. If information makes the marginal return clearer, overestimators should revise their perceived tax rate downward and may see additional work as more rewarding. Underestimators should revise upward and may see the return as less rewarding. People whose initial beliefs were already about right should have less reason to change in either direction. These predictions concern the sign of the correction, not a promise that hours will move. Whether the correction reaches behavior depends on the design of the information test and on whether respondents can actually adjust their work.
The central lesson is therefore narrower and stronger than “people do not understand taxes.” The Dutch respondents held heterogeneous beliefs about a specific marginal return. The errors were often large, their direction differed by income, and confidence did not reliably reveal accuracy. That is enough to make perception a plausible link between tax complexity and labor supply, but not enough to establish causation. The next stage is to connect these descriptive patterns to an intervention that changes information while preserving a credible comparison for what would otherwise have happened.