Two guesses from the same person can sometimes work better together than either guess alone. The method is called dialectical bootstrapping: make a first estimate, deliberately reason from a different perspective to produce a second, and combine them when both estimate the same uncertain quantity. The possible gain comes from partly cancelling errors, not from a mysterious doubling of intelligence.
It works best when the second reasoning route brings different evidence and the mistakes are not too closely correlated. It can fail when the two guesses share the same mistaken assumption, when one input is already a precise measurement, or when a simple average ignores different reliability. This guide shows the calculations, classroom exercises and decision checks needed to use the idea carefully.
Did you know? Researchers Herzog and Hertwig studied whether one person can create a limited version of the wisdom of crowds by reconsidering a quantitative estimate through another mental route. Their original 2009 research and later work concern estimation and judgement under uncertainty; they do not establish that every thought needs to be averaged or that a second perspective always improves truth.
Here you can choose a quick route for two-estimate arithmetic, a deeper route for variance and correlated errors, or a classroom and organisational route for better procedures. The worked examples are deliberately constructed and labelled; none claims to report actual eduKate student outcomes or a newly run psychological experiment.
Contents and three reading routes
For the mechanism start with what the method is, units and denominators, how errors cancel, covariance, and creating different evidence. For practical use continue through the chapters in order or follow the targeted routes below.
For practice go to five worked questions, reliability-based weights, student exercises, planning estimates, changed-context transfer and the six-stage workflow. For research interpretation use forecast scoring, error correlation, boundaries and the original sources.
One mind can sometimes produce two useful estimates
Dialectical bootstrapping is a way to improve a numerical judgement by deliberately generating a second estimate from a meaningfully different line of reasoning and then combining it with the first. The trick is not simply to guess twice. It is to make the second thought less dependent on the same assumptions that shaped the first. When their errors are not perfectly aligned, averaging can improve expected accuracy.
Stefan Herzog and Ralph Hertwig studied this idea in research published in Psychological Science in 2009. They called attention to a crowd within one mind: a person can revisit a quantitative question after attempting to consult different knowledge or challenge an initial assumption. Their finding concerns some estimation tasks under specified conditions. It does not mean that repeated thinking guarantees correct answers in every situation.
The method is useful to someone estimating a classroom project’s duration, predicting attendance, sizing a resource budget, or checking a factual quantity with uncertainty. It is less useful when an exact measurement or authoritative record already exists and can be consulted. If a document specifies a deadline, finding the actual date is better than averaging two recollections of it.
The immediate learner action is simple: write the first number and its reasoning; deliberately adopt a contrasting plausible perspective; write a second independently justified number; then use a transparent combination rule. Keep both estimates visible so that the final answer does not erase the evidence of how it was formed. The combined number is a forecast, not a new observation.
Start with quantities, units and one unambiguous question
An estimate is only meaningful if everyone is estimating the same thing. Compare the phrases How long will our activity take, How long until the first student finishes, and How long until all twenty students submit a correct solution. They use the same word time but represent different measurable quantities. Before invoking two estimates, define the start event, end event, population and unit of measurement.
Suppose a tutor asks how many minutes a group will spend checking a mathematics worksheet. One estimate might count writing time, while another includes explanation and corrections. Averaging the two without noticing that they refer to different tasks is a category error, not wisdom. Fix the definition first: minutes from the beginning of checking until each participant has reviewed and corrected every marked item.
Units matter when the second perspective changes. One student may estimate 1.5 hours and another internal voice may say 100 minutes. Convert 1.5 hours to 90 minutes before combining. The unweighted arithmetic mean is then (90+100)/2=95 minutes. A system that accidentally averages 1.5 and 100 and labels the result minutes has produced numerical decoration, not a valid forecast.
Write the denominator as well as the answer. If the job is average time per worksheet, the denominator is number of worksheets; if the job is total class time, the sum of concurrent student times may not be appropriate because three people work simultaneously. Dialectical bootstrapping improves the inputs to a judgement; it cannot repair a badly defined estimand.
Why two errors may cancel rather than reinforce
Let an unknown true quantity be T. Write the first estimate as X=T+e1 and the second as Y=T+e2, where e1 and e2 are estimation errors in the same units. The simple average is M=(X+Y)/2=T+(e1+e2)/2. The average error is therefore (e1+e2)/2. If the errors point in opposite directions, some of their magnitudes cancel.
Take a constructed example where T=146 people, the first estimate is 120 people and the second is 160 people. The first error is minus 26 people and the second is plus 14 people. The average is 140 people, with an error of minus 6 people. Here the combined estimate is more accurate than either individual estimate. The gain is an arithmetic consequence of these chosen inputs, not a claim about observed research participants.
Changing the signs shows why this is not magic. If the first estimate is 120 and the second 125 when truth remains 146, their average is 122.5, still 23.5 people low. Both estimates missed in the same direction because both may have been derived from the same false starting assumption. Averaging reduces their disagreement but does not correct the shared mistake.
Even when errors have opposing signs, an average need not beat the better contributor. If truth is 146, one estimate is 145 and another is 165, averaging gives 155, an error of 9, worse than the first estimate’s error of 1. Distinguish a general expected benefit over repeated trials from a guaranteed improvement on any single question.
The variance calculation explains the boundary
To analyse many comparable tasks, assume two estimates are unbiased and have error variances v1 and v2. Let their error covariance be c. The variance of their arithmetic-average error is (v1+v2+2c)/4. The plus sign on covariance is important: errors that rise and fall together make averaging less useful. Errors that move differently permit more cancellation.
If both estimates have error standard deviation 20 units, each variance is 400 squared units. With zero correlation the covariance is zero and the average has variance (400+400)/4=200 squared units, for a standard deviation of about 14.14 units. The reduction is real but smaller than halving the standard deviation. Halving independent variances does not halve their square roots.
If correlation is 0.8, covariance equals 0.8×20×20=320 squared units. The average variance becomes (400+400+640)/4=360 squared units and its standard deviation is about 18.97 units. With perfect positive correlation of 1, the variance is 400 and the average brings no variance reduction. Correlation is about how errors co-move, not whether the estimates look aesthetically different.
If the estimates share a systematic bias of plus 10 units, the average retains that bias of plus 10 units. The expected squared error includes both variance and squared bias. It would therefore be misleading to celebrate a narrower distribution if all estimates were clustered reliably around the wrong value. A sound review checks average directional error alongside absolute or squared prediction error.
Produce a second estimate from genuinely different knowledge
A second number can be created without producing genuinely new information. Re-reading the first estimate and nudging it slightly creates a highly related judgement, particularly when the initial number stays in view. To gain diversity, return to the problem through an alternative decomposition: bottom-up parts rather than a global impression, comparison with an unrelated precedent rather than memory of this event, or a deliberately imagined opposing possibility.
For a worksheet, the first estimate may come from a teacher’s intuitive impression that ten minutes is sufficient. A second estimate can instead add three minutes to distribute papers, seven to locate mistakes, six to explain corrections and five to rewrite solutions, yielding 21 minutes. Neither decomposition is automatically true. The point is that it exposes different assumptions to critique.
A more disciplined second pass asks: What would make my original answer wrong in the opposite direction? If the first estimate assumes everyone understood the instructions, imagine a plausible class in which one misunderstood keyword forces a shared explanation. If the first estimate imagines the slowest possible session, ask which parts could be completed in parallel. The counter-perspective should be reasonable, not an invented extreme chosen merely to move the mean.
Time separation can reduce the hold of the first mental route, but it does not guarantee independence. Recording reasoning and then deliberately switching method is often more useful than returning after a minute and repeating the same heuristic. The goal is partially nonredundant information, not artificial disagreement for its own sake.
Considering the opposite is a method, not a slogan
A teacher can ask a student to defend the initial estimate and then write a brief case for why a different numerical range might be justified. The alternate case must contain a checkable reason, such as a different comparison group, missing task stage or new denominator. Merely saying I may be wrong does not supply useful information to the second estimate.
If the first forecast for a field trip is 50 minutes, a contrary path could examine the queue at the entrance and the walking route separately. Suppose the walking portion takes 18 minutes, the museum orientation 12 and a reasonable queue allowance 25. That produces 55 minutes without copying the original intuition. If the original 50 minutes omitted the queue entirely, the disagreement identifies a concrete repair.
Contradiction should not be interpreted as always selecting the opposite of the first number. An initial 50-minute estimate could be too high or too low. The second perspective should respond to actual uncertainty. Asking What evidence would move this estimate upward? and What evidence would move it downward? often exposes the causal levers before the estimator commits to one direction.
In social questions, especially those about people, the technique requires care. Constructing an alternative explanation for a colleague’s behaviour is useful for reducing unwarranted certainty, but the resulting views should not be averaged as if personality traits had a common physical unit. The numerical method applies when a numerical target and defensible combining rule exist; perspective-taking on its own is a related but different skill.
Protect the second estimate from the first estimate’s anchor
An anchor is an initial number that affects later judgements even when it contains little relevant information. A person who has just written 200 may find it difficult to propose 120 without feeling that the second answer is too different. If the original anchor is untrustworthy, using it as the centre of the second estimate can preserve bias while creating an illusion of independent corroboration.
In a classroom version, write the first estimate privately and place the paper face down. Next use a different prompt that requests a bottom-up breakdown rather than another single guess. Once both numbers and reasons are recorded, reveal them for discussion. The procedure does not prove independence, but it creates a practical barrier against obvious copying.
A more advanced team can assign two analysts different starting information or methods while defining a common final target. One analyst uses actual historical task data, while another computes a resource-based estimate. They reconcile the question definition only before work begins, not after seeing which forecast looks better. Different inputs can help; however, deliberately withholding essential safety information is not appropriate.
Inspect repeated prediction errors to see whether apparent diversity is genuine. If estimates consistently move together as conditions change, their correlation may remain high despite different prose explanations. A different method is only valuable if it exposes relevant new variation or systematically corrects a known weakness.
Five worked examples with explicit denominators
Consider five entirely illustrative forecasting problems with true numerical answers 40, 100, 200, 300 and 500 units. First estimates are 34, 116, 208, 280 and 540. Their absolute errors are 6, 16, 8, 20 and 40 units, summing to 90. Dividing by five questions gives a mean absolute error of 18 units per question.
Second estimates for the same five targets are 52, 84, 182, 316 and 480 units. Their absolute errors are 12, 16, 18, 16 and 20 units, summing to 82. Dividing by the same five tasks gives a mean absolute error of 16.4 units. The second attempt is somewhat better on this constructed set, but neither list knows the answer ahead of time in a genuine evaluation.
The pairwise averages are 43, 100, 195, 298 and 510 units. Against the five true answers their absolute errors are 3, 0, 5, 2 and 10 units, totalling 20. The mean absolute error is 20/5=4 units per question. This particular choice of counterbalancing errors makes the combination look strikingly effective; it is a teaching example of arithmetic, not evidence of an expected improvement of that size in people.
Now try three different constructed questions with true answers 50, 80 and 110, first estimates 60, 90 and 120, and second estimates 63, 92 and 122. The first mean absolute error is 10. The combined estimates 61.5, 91 and 121 have absolute errors 11.5, 11 and 11, for an average of about 11.17. Averaging worsens performance here because both passes share the same high bias.
Why choosing the apparently better estimate is difficult
After seeing two estimates, people may be tempted to keep whichever one sounds more convincing. This can be reasonable when one method has verified evidence of better reliability. It is less reasonable when the choice rests on confidence or elegance alone. You do not know an estimate was accurate merely because its explanation came easily to mind.
Herzog and Hertwig later investigated decisions to combine or choose in dialectical bootstrapping, reported in a 2014 journal article indexed in PubMed. Their research examined when people combined estimates and how these choices related to disagreement. That work is a reason to distinguish the arithmetic rule from the behavioural choice to use it; the article does not make every form of averaging optimal.
An example clarifies the difference. A student guesses 120 attendees from memory, then derives 150 from a recorded seating plan. If the seating plan is accurate and complete, it may deserve more weight than memory. But the student must check whether the plan is a maximum capacity rather than expected attendance. The best input is not always the one that looks most official.
Write the selection rule before checking the realised answer. A team can decide to use a simple average when two comparable independent methods exist and switch to a better-calibrated weighting rule once enough held-out evaluation data have accumulated. Choosing weights only after knowing the result allows the analyst to manufacture apparent accuracy.
Unequal reliability calls for transparent weights
When two unbiased estimates have independent errors but different variances, the minimum-variance linear weighted average gives more weight to the more precise estimate. If their variances are 100 and 400 squared units, the first receives weight 400/(100+400)=0.8 and the second weight 100/(100+400)=0.2. The weights sum to one, so the combined estimate stays on the original numerical scale.
For illustrative estimates of 120 and 160 people, those weights produce 0.8×120+0.2×160=128 people. The unweighted mean would be 140 people. If the truth in this one case happens to be 146, the unweighted mean is better, but that does not refute the variance-based weighting rule. The rule concerns long-run expected squared error under specific assumptions, not hindsight on one selected example.
If error covariance is nonzero, the optimal unconstrained weights change. Estimated variances and covariances can be unstable when a team has only a few old questions, and a beautifully precise decimal weight can merely reflect noise. Equal weights are often a transparent starting point; later refinements should be justified with enough comparable data, sensible regularisation and evaluation on new cases.
Consider what happens when one forecaster is deliberately biased upward for safety, such as estimating the maximum time needed for evacuation. That estimate serves a different objective from predicting the average duration. An inverse-variance weight designed for unbiased forecasts is inappropriate until the team agrees whether the decision target is mean time, a high percentile or a conservative planning allowance.
Measure improvement on future questions, not a favourite anecdote
A method should be evaluated on questions selected before their answers are known. Store the numerical target, unit, first estimate, second estimate, fixed combining rule, actual answer, and any explanation of unusual conditions. Repeat across many tasks of the same type. If an organisation changes which questions it counts only after seeing the errors, the measured benefit will be biased.
Mean absolute error is sum of absolute deviations divided by the number of evaluated questions. Mean squared error is sum of squared deviations divided by that same count. Squaring gives large errors more influence, which can be appropriate when very large misses are especially costly. These metrics answer different questions; report the one linked to the actual decision and include a sample count.
Directional bias is mean signed estimate-minus-truth error. A small bias can hide large positive and negative misses that cancel, so it must not replace absolute error. Also distinguish forecasts that are regularly too high on particular subgroups or task lengths. A classroom method may look accurate overall while systematically underestimating tasks requiring careful explanation.
If a procedure appears better on a training set, evaluate it again on questions from a later period or a different but relevant group. Keep the transformation rule fixed. A good record should show how many tasks were attempted, how many had known answers, how missing outcomes were handled, and whether the two estimates were produced under genuinely different prompts.
Correlation and nonredundancy are more important than headcount
Two estimates from one mind are obviously not fully independent in the same way as measurements made by unrelated instruments. Both draw on personal memory, available examples and the question’s wording. Yet their errors need not be identical. Deliberate alternative reasoning can expose a different subset of knowledge, which is the opening dialectical bootstrapping tries to use.
Even a panel of ten people can share one bias if everyone reads the same incorrect summary before estimating. Ten voices are not ten independent sources merely because ten names appear on a spreadsheet. When a group needs genuine diversity, collect initial judgements privately, avoid revealing an early dominant answer, and document distinct evidence sources where practical.
One useful diagnostic is to chart each estimator’s signed error across comparable tasks. If one method consistently overestimates when the other underestimates, averaging may be useful. If both follow the same error pattern, adding them may provide little variance reduction. Do not infer causal independence from a correlation coefficient based on two or three observations.
The companion eduKate guide on collective blind spots examines whether independent observers actually cover missing evidence. Dialectical bootstrapping asks the related but narrower quantitative question: can one person generate partly nonredundant numerical estimates and combine them responsibly?
How discussion can destroy the benefit before it begins
Suppose four teachers are estimating time for a new practical lesson. The first teacher announces 35 minutes confidently. Others may adjust their own numbers toward that anchor without ever sharing their independent calculations. When the group later averages 35, 37, 36 and 38, the narrow spread looks reassuring but could reflect social influence rather than evidence.
A better process asks everyone to submit a first estimate privately with a reason. Then each contributor performs an independent second pass using a contrasting method. Only after both passes are locked does the group compare ranges, definitions and evidence. This protocol separates the generation of candidates from the negotiation of a final plan.
The difference between dialectical bootstrapping and conventional brainstorming is that this method is anchored to a measurable numerical target, with predictions that can later be scored. Brainstorming may generate creative possibilities without requiring a scalar answer. Both can be valuable, but substituting one label for the other hides what the team is trying to improve.
Information cascades are another reason to separate estimates before social discussion. The existing information cascades guide examines how following others can hide independent evidence. A classroom or meeting that wants diverse judgements should design its sequence to prevent the first answer from becoming a template for every later voice.
A student-friendly method for estimation in Mathematics and Science
Students can practise the method without needing to know covariance matrices. Give a measurable but not immediately known answer, such as the number of grains in a small measured scoop or the number of squares covering a large irregular surface. First estimate by overall visual impression. Then count a small sample and scale it up using a clearly defined denominator.
Suppose a tray has twelve equal compartments. First guess there are 240 beads total. In a second method, count 18 beads in a representative compartment and multiply 18×12=216. The simple average is 228 beads. This is a provisional estimate. If compartments are visibly unequal or bead counts vary widely, sampling more compartments may be a better improvement than averaging the first two guesses.
After an actual full count, compare all three numbers with truth. Ask which assumptions mattered. Did the sampled compartment represent the others? Was the first visual estimate systematically influenced by the tray’s size? Was the final combined estimate better, or was one method decisively more accurate? The objective is to teach judgement under uncertainty rather than to make averaging a ritual.
For school science, the second route should not overwrite an actual measurement. A reading of 23.4 millilitres from a suitable instrument has a different evidential status from an estimate made by eye. If repeat measurements are needed, combine them according to measurement uncertainty and proper procedure rather than treating a guess and an instrument reading as interchangeable.
Use two routes when planning work and resources
Consider an organisation estimating effort for a training workshop. The first method uses the total duration of last year’s comparable workshop. The second decomposes the new programme into preparation, delivery, documentation and follow-up. Those methods may disagree because they expose different costs, which is useful before committing money or staff hours.
Construct an example: an initial analogy suggests 12 staff-hours. A fresh task breakdown gives three hours to prepare materials, eight hours to teach, four hours to assess and three hours for follow-up, totalling 18 staff-hours. The equal average is 15 staff-hours. But if the schedule contract requires all four components, the breakdown reveals a hard minimum that averaging cannot lawfully or practically erase.
The method therefore helps decision makers inspect alternatives rather than blindly enforcing a mean. Use the average as one candidate when both inputs estimate the same uncertain total. If one estimate represents a mandatory lower bound and another is a probability forecast, they cannot simply be averaged. Clarify whether the target is expected labour, guaranteed capacity, budget ceiling or recommended reserve.
Planning fallacy is a neighbouring concept, not a synonym. A planning-fallacy diagnosis asks why people underestimate time or resources, often by focusing on inside details. Dialectical bootstrapping offers a specific aggregation tactic: deliberately form a second estimate using different knowledge and examine whether their errors offset. Both may contribute to a better workflow without replacing real project measurements.
Do not average probabilities without defining events
Probability estimates can also be combined, but the event must be precisely specified. A prediction that a school activity will be cancelled tomorrow is not interchangeable with a prediction that it will rain at any point tomorrow. Even if both are percentages, they refer to different events and require different evidence. First define the event, date, threshold and outcome source.
Suppose one forecast assigns a 30 percent chance to a defined event and a second independently reasoned forecast assigns 50 percent. A simple arithmetic pool gives 40 percent. This is a combination of beliefs, not a guarantee the event will happen in four of ten repetitions. To test calibration, collect many comparable forecasts and compare announced probability bins with event frequencies.
Overconfidence and shared missing evidence can defeat averaging. If both estimates ignore the same adverse condition, 40 percent may be too low or too high. A well-calibrated team can use proper scoring rules, such as the Brier score for binary events, to assess whether combined probabilities become more accurate over a substantial set of outcomes. The details differ from combining a physical count.
Do not force probability pooling onto safety-critical thresholds without an explicit decision rule. If the cost of a false negative greatly exceeds a false positive, a decision maker may need a more conservative action threshold even when the estimated probability is numerically well calibrated. The probability forecast and the action policy are separate components.
Institutional memory can make two estimates too similar
An organisation may believe it has two independent planning methods because they reside in different departments. Yet both may copy the same spreadsheet assumption, refer to last year’s mistaken target, or use identical historical samples. A first step in honest dialectical bootstrapping is to map common data dependencies, not merely insist that participants work in different rooms.
A team can write a short evidence provenance note for each estimate: which records, comparison periods, assumptions and units were used. If both forecasts depend on the same underlying vendor quote, their agreement is not independent confirmation. A genuinely different second estimate might use observed production rates or a structured bottom-up process rather than copying the quote under another label.
That does not mean an alternative method must avoid all shared facts. Everyone should use the same correct task definition and up-to-date safety requirements. Diversity concerns uncertain inference routes, not deliberate disagreement about known facts. In teaching, it is valuable to separate shared premises from the judgmental steps that remain open.
Preserve the first estimate even when the team selects another answer. Reviewing the difference after the real outcome arrives provides evidence about where each method works. Without that record, decision makers remember convincing successes and forget unremarkable failures, which turns a potentially testable forecasting technique into an anecdotal reputation.
A fair after-action review separates judgement from luck
Suppose a team predicts a visitor count of 1,200, then receives 1,700 because an unannounced nearby event redirects people. A review should check what information was available when the estimates were submitted, which assumptions could reasonably be questioned, and whether the combining rule was appropriate. It should not label one participant clever simply because their arbitrary guess happened to lie closer to the surprising truth.
Record errors as estimate minus actual outcome, in the same units, and compare both individuals’ and combined estimates. Consider whether the surprise was a genuine rare event, a predictable factor that someone neglected, or a change in the data definition. These explanations call for different repairs. A new event category might need an explicit control variable rather than another round of self-contradiction.
Re-evaluate the rule over repeated tasks instead of choosing whichever one looks best on the most recent case. If a simple average reliably reduces error, retain it. If one component becomes systematically better after new data arrive, update the weights using a documented validation procedure. If both remain biased by the same source, correct the source before pooling.
A useful review ends with a changed future action: revise the second-prompt wording, improve outcome recording, add an independent reference class, or stop combining heterogeneous quantities. An honest failure can teach more about judgement than an unexamined lucky success.
When averaging is the wrong operation
Some questions are factual lookups, not estimates. A train timetable, examination rule or published technical standard should be checked in the authoritative source rather than approximated by two recollections. Two wrong memories of a deadline do not become evidence when averaged. The method is meant for uncertain quantities where exact observation is absent or costly.
Other questions ask for categorical decisions. Choosing between two medical treatments, deciding whether data access is permitted, or determining whether a claim has sufficient ethical support involves constraints, evidence and values that cannot be reduced to the arithmetic mean of two labels. Multiple perspectives remain useful, but they require a different decision procedure.
Two estimates can also be generated from the same flawed model. Correlated errors, identical missing variables and common incentives may make their average no safer. A second estimate produced merely to justify the first is not dialectical. If a team cannot create a reasonably distinct inferential route, it should say so and seek new evidence rather than manufacture a second number.
Finally, the method has costs: more cognitive work, more recording and potential confusion if workers treat the combined output as a fact. Use it where the cost of error warrants deliberation and where outcomes can eventually be checked. For trivial familiar tasks, a direct measurement or one calibrated forecast may be faster and more accurate.
Changed-context problem: a new subgroup breaks the old advantage
Try this independent problem. In five earlier events, a team’s two estimates had often fallen on opposite sides of actual attendance and the simple mean worked well. A new event is an online workshop, and both methods now rely on the same obsolete registration-to-attendance ratio of 80 percent. Method one predicts 160 viewers from 200 registrations. Method two predicts 168 because it adds a small optimistic adjustment. What should the team do?
The arithmetic mean is (160+168)/2=164 viewers, but that result does not remove the shared 80-percent assumption. If a current sample of 50 comparable registrations yields 30 actual viewers, the observed ratio is 30/50=60 percent. Applying 60 percent to 200 registrations gives a revised estimate of 120 viewers, provided the sample is relevant and selection effects are examined. The denominator is registrations, not people who clicked an advertisement.
The correct lesson is not always choose 120. The new sample could be too small or atypical. A careful answer explains why the original pair lacks useful independence, checks the new 50-person sample’s comparability and uncertainty, then updates a forecasting model. Simply averaging two highly correlated errors is not the same as combining independent evidence.
Now change the event again: attendance is capped at 100 viewers by the platform. Even the revised expected demand of 120 cannot become actual attendance above a hard technical limit. The final operational forecast must distinguish demand from served attendance. The technique may help estimate demand, but domain constraints govern the quantity a decision maker can actually plan to deliver.
A practical six-stage protocol for one person or a team
First define the target, units, population, observation window and later source of truth. A crisp question such as How many minutes until all twelve participants finish the task is more testable than Is the workshop fast. Record any known upper or lower bounds before guessing. An estimate outside a genuinely hard bound is not improved by pooling.
Second write the first estimate and its reason while avoiding group influence. Third create an alternative route that uses a different decomposition, historical comparison, causal model or plausible opposing scenario. Fourth write the second estimate without consulting the first number when feasible. Different wording alone is not enough; the reasoning must introduce meaningful new considerations.
Fifth apply the predeclared combining rule, normally a simple mean when two estimates target the same quantity and no credible weighting evidence exists. Keep a note of why the alternative was different. Sixth compare with actual outcomes later and score the first, second and combined estimates consistently. Include all completed cases and explain exclusions.
After multiple rounds, decide whether the second route improves accuracy enough to justify its cost. Look especially for shared signed bias, subgroup variation and unexpected extreme errors. If the advantage vanishes under changed conditions, diagnose why rather than forcing another batch of averages. Intelligence in this workflow is the ability to test and revise the judgement process itself.
Frequently asked questions
Can I take the average of my first and second guess every time?
Only when they estimate the same numerical quantity and you have a reasonable basis for pooling. Averaging can worsen a good individual estimate or preserve a shared bias. Record both estimates and verify improvement over repeated cases rather than assuming every average is wise.
Is dialectical bootstrapping the same as asking a different person?
No. It is a structured attempt to obtain partly nonredundant estimates from one person’s knowledge. Getting estimates from independent people is conventional crowd aggregation. Both depend on shared biases and correlation, so neither guarantees improvement.
How do I make my second estimate genuinely different?
Change the inferential route: visual impression versus sampled count, analogous past case versus bottom-up decomposition, or an initial optimistic scenario versus a plausible counterargument. Hide the first number where practical and write down the new evidence or assumptions.
Why do correlated errors matter so much?
If both estimates drift together whenever the task changes, their arithmetic mean carries much of the same error. The average variance equals (v1+v2+2 covariance)/4. Less correlated errors provide more opportunity for cancellation, all else equal.
Should two estimates receive equal weight?
Equal weights are simple and auditable. If a sufficiently large comparable evaluation set establishes different reliabilities, a documented weighted combination may be better. Do not infer reliable weights from one example or choose them after seeing the correct answer.
Does the method remove planning fallacy or confirmation bias?
Not automatically. Deliberately constructing a different perspective can expose assumptions, but the same person may share biases across both passes. Additional outside evidence, measurements or independent scrutiny may be needed.
Can this technique help children learn Mathematics?
Yes, as a way to compare estimates and check units. Ask for one intuitive estimate and one reasoned estimation by counting a representative sample. Reveal the actual answer when possible and discuss which assumptions produced error, rather than rewarding the numerical average by default.
What accuracy metric should we use?
For numerical outcomes, mean absolute error is intuitive and reports average miss in the original units. Mean squared error emphasises large mistakes. Also check signed bias and relevant subgroups. Keep the denominator, sample count and outcome definition fixed.
Can a single better source outweigh both guesses?
Yes. If an authoritative measurement or properly recorded fact is available, use it. Dialectical bootstrapping is not a substitute for direct evidence. Two opinions cannot transform a guessed number into a measurement.
What if the second estimate is identical to the first?
The average is unchanged and there is no numerical gain from this pair. That may indicate strong evidence, or it may reflect anchoring and a shared mental route. Examine the reasons before deciding whether an additional method or new observation is needed.
Original research and how to check the evidence
Herzog, S. M., and Hertwig, R. (2009), The Wisdom of Many in One Mind: Improving Individual Judgments With Dialectical Bootstrapping. The original research introduces the technique and its conditional accuracy claim. The public abstract is available; this guide does not invent unpublished effect sizes.
Herzog and Hertwig’s follow-up work, Think twice and then: combining or choosing in dialectical bootstrapping?, published in the Journal of Experimental Psychology: Learning, Memory, and Cognition (2014), tests what makes people combine multiple judgements.
The original researchers’ Harnessing the wisdom of the inner crowd (Trends in Cognitive Sciences, 2014) reviews the broader idea of self-generated nonredundant estimates.
A published scholarly reply, The Crowd Within and the Benefits of Dialectical Bootstrapping (2013), illustrates that quantitative conclusions about pooling require careful specification of the benchmark and error measure.
All arithmetic in this guide is an explicitly illustrative reconstruction using stated values, units and denominators. It is not a claim of access to private research data or a measured improvement in a real classroom.
Related verified eduKate learning paths
Start with the How Intelligence Works hub when you want to connect individual judgement to human and collective intelligence.
Read Information Cascades for the way social influence can make apparently independent voices copy one another.
Use Collective Blind Spots when the difficulty is missing information across observers rather than how to average two numerical estimates.
See Signal Detection Theory when the task is detecting a weak signal or selecting a response threshold rather than forecasting a scalar quantity.