Statistics & decisions

Bayesian Updating Explorer

Bayes' rule on 2 to 6 discrete hypotheses with a likelihood table of 2 to 6 evidence types and up to 20 observations, updated one observation at a time: the beliefs after every step, the most probable hypothesis with its tie band, the target's probability, log10 odds and step Bayes factors, the entropy in bits, the belief shift, the first observation as natural frequencies of 1,000 people by the largest-remainder rule, the value of the next observation (beliefs if next, belief shift, entropy after, Bayes factor, expected information gain), the fewest repeats of one outcome to reach a belief level in closed form, and five scored challenges.

Last updated

Calculator overview

Inputs and outputs

This summary comes from the calculator's published input and output contract.

Inputs

Prior weight of H3 (required)
About this input

The prior weight of H3: 0, or a number from 0.1 to 100 (used when at least 3 hypotheses are in use). Only the proportions matter. Dragged as a bar on the canvas.

Default 10 Range 0 to 100
Prior weight of H2 (required)
About this input

The prior weight of H2: 0, or a number from 0.1 to 100. Only the proportions matter. Dragged as a bar on the canvas.

Default 98 Range 0 to 100
Prior weight of H1 (required)
About this input

The prior weight of H1: 0, or a number from 0.1 to 100. Only the proportions matter. Dragged as a bar on the canvas.

Default 2 Range 0 to 100
Prior weight of H4 (required)
About this input

The prior weight of H4: 0, or a number from 0.1 to 100 (used when at least 4 hypotheses are in use). Only the proportions matter. Dragged as a bar on the canvas.

Default 10 Range 0 to 100
Target hypothesis (required)
About this input

The target hypothesis whose probability, log10 odds, Bayes factors and repeats the readouts follow, a whole number from 1 to the number of hypotheses (the slider stops there).

Default 1 Range 1 to 2 (conditional)
Prior weight of H6 (required)
About this input

The prior weight of H6: 0, or a number from 0.1 to 100 (used when at least 6 hypotheses are in use). Only the proportions matter. Dragged as a bar on the canvas.

Default 10 Range 0 to 100
Prior weight of H5 (required)
About this input

The prior weight of H5: 0, or a number from 0.1 to 100 (used when at least 5 hypotheses are in use). Only the proportions matter. Dragged as a bar on the canvas.

Default 10 Range 0 to 100
Number of evidence types (required)
About this input

The number of evidence types (possible outcomes of one observation) in use, a whole number from 2 to 6. Set by the chips on the canvas.

Default 2 Range 2 to 6
Challenge number (required)
About this input

Selects a challenge: 0 for none, 1 to 5 for a predefined challenge, which fixes the prior, the table and the settings except what it leaves to you.

Default 0 Range 0 to 5
Belief level (required)
About this input

The belief level the repeats readout aims for, from 0.5 to 0.999 (Advanced).

Default 0.95 Range 0.5 to 0.999
Number of hypotheses (required)
About this input

The number of hypotheses in use, a whole number from 2 to 6. Set by the chips on the canvas; hypotheses above it are hidden and ignored.

Default 2 Range 2 to 6
Your guess (required)
About this input

Challenge 1: your guess of the chance that someone who tests positive has the condition, as one of six bands; Not yet predicted keeps the results hidden. Not used outside challenge 1.

Default Not yet predicted Allowed Not yet predicted, Below 10%, 10% to 30%, 30% to 50%, 50% to 70%, 70% to 90%, 90% or more
Observations in order (an evidence number per row; blank rows are skipped)
When omitted Blank
About this input

The observed outcomes in order: each row holds an evidence number from 1 to the number of evidence types, or is blank. Blank rows are skipped. Up to 20 observations, built on the canvas token strip. Optional: with no observation the beliefs stay at the prior.

Default 1 row
ColumnRange or allowed values
Observation 1 to 6
Likelihood of each evidence type under each hypothesis (row h is hypothesis Hh) (required)
About this input

The likelihood table: row h, column Ee is the percentage chance of outcome Ee when hypothesis Hh is true, 0 or from 0.1 to 100, or blank outside the hypotheses and evidence types in use. Every cell in use needs a value. A row that does not add to 100% is rescaled to add to 100%.

Default 2 rows
ColumnRange or allowed values
E1 (%) 0 to 100
E2 (%) 0 to 100
E3 (%) 0 to 100
E4 (%) 0 to 100
E5 (%) 0 to 100
E6 (%) 0 to 100

Outputs

Final belief in H1
About this output

The final belief in H1 after the observations applied. Blank when H1 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Observations used
About this output

The number of observations applied: all of them, or those before an impossible one. Blank when the inputs are not valid.

No unit declared
Final belief in H2
About this output

The final belief in H2 after the observations applied. Blank when H2 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Final belief in H4
About this output

The final belief in H4 after the observations applied. Blank when H4 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Final belief in H3
About this output

The final belief in H3 after the observations applied. Blank when H3 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Observations entered
About this output

The number of observations entered (rows in use). Blank when the inputs are not valid.

No unit declared
Formula module versions
About this output

The formula modules the model is built from, with their versions.

No unit declared
Model version
About this output

The model version.

No unit declared
People with the observed first result
About this output

Of 1,000 people split by the prior and the likelihoods, the people whose result is the first observation (whole people by the largest-remainder rule). Blank with no observation applied.

Unit people
Share of them with the target
About this output

The target people over the people with the observed result: close to, but not exactly, the probability after the first observation. Blank with no observation applied.

No unit declared
Of them, people with the target hypothesis
About this output

Of those people, the people with the target hypothesis. Blank with no observation applied.

Unit people
Final belief in H5
About this output

The final belief in H5 after the observations applied. Blank when H5 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Target certain or ruled out
About this output

Certain when every rival of the target is ruled out; Ruled out when the target has probability 0; blank otherwise (the odds are then numbers).

No unit declared
Target final log10 odds
About this output

The base-10 logarithm of the target's final odds against all its rivals together. Blank when the target is certain or ruled out, the inputs are not valid or the result is withheld.

No unit declared
Target final probability
About this output

The final probability of the target hypothesis after the observations applied. Blank when the inputs are not valid or the result is withheld.

No unit declared
Target prior probability
About this output

The prior probability of the target hypothesis (its weight over the sum of the weights in use). Blank when the inputs are not valid.

No unit declared
Target prior log10 odds
About this output

The base-10 logarithm of the target's prior odds against all its rivals together. Blank when the target is certain or ruled out, or the inputs are not valid.

No unit declared
Target hypothesis used
About this output

The target hypothesis the calculation used. Blank when the inputs are not valid.

No unit declared
The beliefs if the next observation is each evidence type
About this output

The beliefs if the next observation is each evidence type (one row per evidence type in use). Drawn as a heat map at Advanced.

No unit declared
Final belief in H6
About this output

The final belief in H6 after the observations applied. Blank when H6 is not in use, the inputs are not valid or the result is withheld.

No unit declared
Prior entropy
About this output

The entropy of the prior beliefs, in bits: 0 when one hypothesis is certain, LOG(H, 2) when the beliefs are equal. Blank when the inputs are not valid.

Unit bits
Strongest next observation for the target
About this output

The evidence type whose next observation would raise the target's odds the most (the largest Bayes factor; one that rules out every rival wins). Blank when the target is certain, the inputs are not valid or the result is withheld.

No unit declared
The target belief after n more repeats of each evidence type
About this output

The target probability after n = 0 to 50 more repeats of each evidence type, one column per evidence type in use (closed form, in log space). Drawn with the belief level as a line.

No unit declared
Challenge result
About this output

No challenge selected, MET or NOT MET.

No unit declared
Challenge score
About this output

100 when the selected challenge is met, else 0.

No unit declared
The value of each possible next observation
About this output

For each possible next observation: its chance, its Bayes factor for the target, the belief shift and the entropy after it, the repeats to the belief level and the target probability after it.

No unit declared
Evidence types used
About this output

The number of evidence types the calculation used. Blank when the inputs are not valid.

No unit declared
Final entropy
About this output

The entropy of the final beliefs, in bits. Blank when the inputs are not valid or the result is withheld.

Unit bits
Challenge: the value scored
About this output

The value the selected challenge scores: the target probability in percent (challenge 1), the target probability (challenges 2, 4 and 5) or the belief shift (challenge 3). Blank with no challenge, while challenge 1 awaits a guess, or when the inputs are not valid.

No unit declared
Belief shift from the prior
About this output

How far the final beliefs moved from the prior: half the sum of the absolute differences (the total variation distance), from 0 to 1. Blank when the inputs are not valid or the result is withheld.

No unit declared
Belief level used
About this output

The belief level the calculation used. Blank when the inputs are not valid.

No unit declared
The beliefs after every step (row k is step k; the prior is step 0)
About this output

The beliefs after every step: row k is step k (step 0 the prior), with the observation, the belief in each hypothesis in use, the most probable label, the target probability, its log10 odds, the step Bayes factor and the entropy. Rows after the observations applied are blank.

No unit declared
Challenge goal
About this output

The goal of the selected challenge; blank with no challenge.

No unit declared
Challenge feedback
About this output

The prompt while challenge 1 awaits a guess, the met text when met, otherwise a hint; blank with no challenge.

No unit declared
Expected information gain of one more observation
About this output

The expected information gain of one more observation, in bits: the entropy now less the expected entropy after it (the mutual information between the hypothesis and the next outcome). Blank when the inputs are not valid or the result is withheld.

Unit bits
Log10 Bayes factor of all the observations
About this output

The final log10 odds less the prior log10 odds: the base-10 logarithm of the Bayes factor of all the observations applied, for the target against its rivals. Blank when either odds is undefined.

No unit declared
The likelihoods used (each row rescaled to add to 1)
About this output

The likelihoods used: each row in use rescaled to add to 1. Drawn as the likelihood heat map.

No unit declared
Most probable hypothesis
About this output

The most probable hypothesis after the observations applied (H1 to H6), or Tie: with the tied hypotheses when two or more are equal within 1E-12 relative. Blank when the inputs are not valid or the result is withheld.

No unit declared
Model Status
About this output

NOT VALID with the reason when an input cannot be used; CHECK when a likelihood row was rescaled to add to 100%, or when an observation impossible under every hypothesis still possible stopped the update; otherwise OK. A ruled-out hypothesis, a certain target, a tie and an uninformative observation read OK.

No unit declared
Probability of the most probable
About this output

The probability of the most probable hypothesis. Blank when the inputs are not valid or the result is withheld.

No unit declared
Least informative next observation
About this output

The evidence type whose next observation would move the beliefs the least (the smallest belief shift). Blank when the inputs are not valid or the result is withheld.

No unit declared
Observation with the fewest repeats
About this output

The evidence type with the fewest repeats to the belief level (the lowest number on a tie). Blank as fewestRepeatsBest.

No unit declared
Fewest repeats to the belief level
About this output

The fewest repeats of one evidence type that take the target to the belief level (0 when it is already there). Blank when no evidence type reaches it within 50 repeats, the inputs are not valid or the result is withheld.

No unit declared
1,000 people by hypothesis and first result (natural frequencies)
About this output

The first observation as natural frequencies: 1,000 people split by hypothesis and result, in whole people adding to exactly 1,000 (largest-remainder rule).

No unit declared
The 1,000-person icon array: the people in each block and the last dot of the block
About this output

The 1,000-person icon array as blocks (hypothesis, evidence, people, last dot): dot i takes the colour of the first block whose last dot is at least i.

No unit declared
Hypotheses used
About this output

The number of hypotheses the calculation used (the challenge value when a challenge is selected). Blank when the inputs are not valid.

No unit declared
This page is provided by LogicCommons for informational purposes only. Results are analysis outputs computed from the inputs you supply and are not professional advice. Verify all inputs and results independently.

LogicCommons is in beta. If a result, label, or reference looks wrong, tell us here; we read every message.

Methodology

Purpose and model boundary

This lab updates beliefs about 2 to 6 competing hypotheses, H1 to H6, one of which is true, as evidence arrives. Each observation has 2 to 6 possible outcomes, the evidence types E1 to E6. You set a prior weight for each hypothesis, a likelihood table giving the chance of each outcome under each hypothesis, and a sequence of up to 20 observed outcomes. The model applies Bayes' rule one observation at a time and returns the beliefs after every step, the most probable hypothesis, a chosen target hypothesis's probability and log odds, each step's Bayes factor, the entropy of the beliefs and how far they moved. It also shows the first update as natural frequencies in 1,000 people, and, at Advanced, what each possible next observation would do. Five scored challenges use the same model.

It is a lab for learning Bayesian updating: why a rare condition stays unlikely after one positive test, why the likelihood ratio and not the likelihood decides how much an observation counts, why some observations teach nothing, and how the order of observations does not matter. It is not a diagnostic or decision tool. It uses a finite set of hypotheses, assumes repeated observations are independent given the hypothesis, and draws no random numbers: you supply every observation.

The model is the calculation authority. The page sends the values you set to the model and draws only what it returns: the belief bars, the stacked history of beliefs, the token strip of observations, the log-odds path, the 1,000-person icon array and the heat maps. The page computes no probability, count or odds itself.

Inputs and units

Probabilities are numbers from 0 to 1. Prior weights are numbers from 0 to 100, and likelihoods are percentages.

Input Meaning and unit
Number of hypotheses A whole number from 2 to 6 (default 2): H1 to H<H> are in use.
Number of evidence types A whole number from 2 to 6 (default 2): E1 to E<E> are the possible outcomes.
Prior weight of H1 to H6 Each 0, or a number from 0.1 to 100 (defaults 2, 98, 10, 10, 10, 10). Only the proportions matter: the model rescales the weights in use to add to 1, so 1 and 99 are a 1% prior. A weight of 0 rules a hypothesis out from the start. At least one hypothesis in use needs a weight above 0.
Likelihood table Row h, column e: the chance of outcome Ee when Hh is true, 0 or from 0.1% to 100%. Every entry in use needs a value; entries outside the hypotheses and evidence types in use may be blank. Each row in use needs a sum above 0 and is rescaled to add to 100%.
Observations Up to 20 observed outcomes in order, each an evidence type in use. Blank rows are skipped. With none, the beliefs stay at the prior.
Target hypothesis A whole number from 1 to 6, and at most the number of hypotheses (default 1): the hypothesis the target readouts follow.
Belief level From 0.5 to 0.999 (default 0.95): the level the repeats readout aims for.
Challenge number 0 for free exploration, or 1 to 5 to load a challenge.
Your guess Not yet predicted, Below 10%, 10% to 30%, 30% to 50%, 50% to 70%, 70% to 90% or 90% or more: your guess in challenge 1.

Where each input is set:

  • On the canvas only: the numbers of hypotheses and evidence types (chips), the prior weights, the observations and your guess in challenge 1. Select an outcome chip to add it at the end of the token strip; select a token to remove it.
  • Draggable on the canvas: the prior weights, as the heights of the prior bars, in steps of 0.1.
  • In the input panel: the likelihood table, the target hypothesis, the challenge number, and the belief level under Advanced.

Governing relationships

Normalisation. With weights w_h for the hypotheses in use, the prior is π_h = w_h / (w_1 + ... + w_H). Each likelihood row is rescaled to add to 1: L(h, e) = a(h, e) / (sum of row h over the evidence types in use), where a(h, e) is the entered percentage.

Bayes' rule, one step at a time. Starting from p_0 = π, observation k with outcome e_k gives

u_k(h) = p_(k-1)(h) x L(h, e_k), Z_k = u_k(1) + ... + u_k(6), p_k(h) = u_k(h) / Z_k.

The posterior is proportional to the prior times the likelihood, and each posterior becomes the prior of the next step. Because the steps multiply, the order of the observations does not change the final beliefs. If Z_k = 0 the observation is impossible under every hypothesis still possible; the update stops before it.

Odds and Bayes factors. For the target hypothesis t with probability p(t) and rival mass R (the sum of the other hypotheses' probabilities, summed directly), the log odds are log10(p(t)) - log10(R). The Bayes factor of step k is the factor by which it multiplies the target's odds against the rest:

BF_k = L(t, e_k) x R_(k-1) / (sum over rivals h of u_k(h)).

With two hypotheses this is the likelihood ratio L(t, e) / L(other, e). The overall log10 Bayes factor is the final log odds less the prior log odds. Each step's Bayes factor is the rise of the log-odds path at that step.

Entropy. The uncertainty of the beliefs, in bits, is H = -Σ p(h) log2 p(h), over hypotheses with p > 0. It is 0 when one hypothesis is certain and log2 of the number of hypotheses when all are equally likely.

Belief shift. The total variation distance between the prior and the final beliefs, (|p(1) - π_1| + ... + |p(6) - π_6|) / 2, from 0 (no change) to 1.

Natural frequencies. For the first update, 1,000 people are split by hypothesis and outcome: the raw count of block (h, e) is 1000 x π_h x L(h, e). Each block first gets the whole-number part of its raw count, and the people left over go one at a time to the blocks with the largest fractional remainders (the largest-remainder method; equal remainders go to the earlier block). The counts are whole numbers that add to exactly 1,000. Of the people whose outcome is the one observed, the share with the target hypothesis is close to, but not exactly, the posterior after one step, because of the rounding to whole people.

The next observation (Advanced). From the final beliefs p:

  • the chance of outcome e next is P_e = Σ p(h) L(h, e);
  • the beliefs if e is seen next are p(h) L(h, e) / P_e, with the entropy and belief shift they would give, and the Bayes factor of e for the target;
  • the expected information gain of one more observation is the current entropy less the expected entropy after it, H - Σ P_e x H_after(e), in bits; it is 0 when every likelihood row is the same;
  • the strongest outcome is the one with the largest Bayes factor for the target (an outcome that would make the target certain comes first), and the least informative outcome the one with the smallest belief shift.

Repeats (Advanced). After n further repeats of outcome e,

p_n(t) = p(t) L(t, e)ⁿ / Σ p(h) L(h, e)ⁿ,

computed in log space so large n cannot overflow. The fewest repeats is the smallest n from 0 to 50 at which p_n(t) reaches the belief level. The curve need not rise steadily: a rival that explains the outcome even better eventually overtakes the target.

Calculation sequence

  1. Check the inputs in the order of the status list under Errors and warnings. In a challenge, take the challenge's fixed values, except the inputs it leaves to you.
  2. Rescale the prior weights and the likelihood rows.
  3. Read the observations in row order, skipping blank rows.
  4. Apply Bayes' rule step by step, stopping before an impossible observation.
  5. For each step, find the most probable hypothesis, the target's probability, log odds and Bayes factor, and the entropy; then the belief shift from the prior to the final beliefs.
  6. Split 1,000 people by hypothesis and outcome by the largest-remainder rule.
  7. From the final beliefs, find what each possible next observation would do, the expected information gain, and the repeats needed to reach the belief level.
  8. Set the status and, in a challenge, score it.

Outputs and interpretation

  • Observations entered and observations used (the steps applied before any impossible observation).
  • Final belief in H1 to H6, the most probable hypothesis (or a tie, such as Tie: H1, H3, within 1×10⁻¹² relative) and its probability.
  • Target prior share and target posterior, the target's prior and final log10 odds against the rest, and the log10 Bayes factor of all the observations together. When the target is certain or ruled out the odds are infinite or zero, the log odds are blank, and a note says Certain: every other hypothesis is ruled out, or Ruled out: this hypothesis has probability 0.
  • Prior entropy and final entropy (bits), and the belief shift (0 to 1).
  • People with the observed outcome and of them with the target (of 1,000), and their ratio, the natural-frequency share.
  • At Advanced: the expected information gain of one more observation (bits), the strongest and least informative next outcome, and the fewest repeats to reach the belief level and the outcome that does it.
  • Settings used: the counts, the target and the belief level in effect.
  • Tables: the beliefs after every step (with the outcome, most probable hypothesis, target, log odds, step Bayes factor and entropy); the rescaled likelihoods; the natural-frequency counts and the icon blocks; for each possible next outcome its chance, Bayes factor, belief shift, entropy after, repeats to the level and target after; the beliefs if next; and the repeat curve for n = 0 to 50.

Reading the canvas: each observation multiplies every bar by how likely that outcome is under its hypothesis, and the model rescales the bars to add to 1. The history bar stacks the beliefs after each step. On the log-odds path zero is even odds, and the rise of each segment is the log of that step's Bayes factor; a step with no point is one where the target is certain or ruled out. The icon array colours 1,000 dots by the model's whole-number counts. In the likelihood heat map darker squares are more likely outcomes; a row that differs from the others is evidence that can tell them apart.

Challenges

Each challenge fixes its own hypotheses, evidence types, prior and likelihood table, except the inputs it leaves to you. The model scores each one MET (100) or NOT MET (0) and gives a one-line hint while it is not met.

  1. Guess the base rate. Before you look, guess the chance that someone who tests positive has the condition. 1% of people have it; the test finds 80% of them and wrongly flags 9.6% of the rest. It is met when your guess names the band that holds the posterior (an exact band edge belongs to the upper band).
  2. Overturn the prior. The prior favours H1. Using at most four observations, make H3 at least 80% likely.
  3. An observation that teaches nothing. Add exactly one observation that leaves the beliefs where they are: a belief shift of at most 0.01.
  4. Fewest observations. Make H1 at least 95% likely using at most three observations.
  5. Move the prior. Keep the test and its one positive result. Change the prior until a positive test means at least a 50% chance of the condition.

Challenge 1 holds back the answer until you guess: while your guess reads Not yet predicted, the final beliefs, the most probable hypothesis, the target posterior, the final log odds and Bayes factor, the final entropy, the belief shift, the natural frequencies and every next-observation result are blank, as are the belief steps after the prior; the challenge reads NOT MET with a prompt to guess. The prior and the likelihoods stay shown.

What to take from the lab: a rare hypothesis stays fairly unlikely after one piece of good evidence, because the prior matters; the ratio of likelihoods, not the likelihood itself, decides how far an observation moves belief; an outcome equally likely under every hypothesis teaches nothing; and the same observations in any order give the same beliefs.

Validation and status logic

The status reads the first match in the list under Errors and warnings. Items 1 to 7 check each entered value against its own bounds; items 8 to 12 check the structure on the values in use, so in a challenge its fixed counts and table govern them. The two CHECK items follow, and every result is shown under them. A ruled-out hypothesis, a certain target, a tie, an observation that changes nothing and a level not reached within 50 repeats are results, not warnings.

Assumptions and limitations

  • A finite set of 2 to 6 hypotheses, exactly one of which is true, and 2 to 6 outcomes per observation.
  • Repeated observations are independent given the hypothesis, and every observation is one of the outcomes in the likelihood table. No observation is random: you supply each one.
  • Likelihood rows that do not add to 100% are rescaled; a row off by more than half a percentage point is flagged.
  • A posterior is exactly 0 only when a prior weight or a likelihood is exactly 0.
  • The natural frequencies round to whole people, so their share differs slightly from the exact posterior.
  • The repeats readout looks only at repeats of one outcome, up to 50.
  • Each test has a stated band: ties for most probable within 1×10⁻¹² relative, belief levels met within 1×10⁻⁹.

Restrictions and non-computing states

Under NOT VALID every result and every table is blank, apart from the model version, the challenge status (No challenge selected, or NOT MET when a challenge is selected), a score of 0, and the goal and hint of a selected challenge.

With no observation, the beliefs stay at the prior and the natural-frequency results are blank. Hypotheses and evidence types not in use are blank in every table. A next outcome with a chance of 0 shows only its label and chance.

In challenge 1, until you guess, the results are held back as described under Challenges.

Errors and warnings

The status messages, first match wins (h, e, i, x and t are the lowest failing indexes, H and E the counts in use):

  • NOT VALID: the number of hypotheses and the number of evidence types must each be a whole number from 2 to 6
  • NOT VALID: prior weight h must be 0 or a number from 0.1 to 100
  • NOT VALID: the likelihood of Ee under Hh must be 0 or a percentage from 0.1 to 100, or leave it blank
  • NOT VALID: observation i must be a whole number from 1 to 6 (an evidence type), or leave it blank
  • NOT VALID: the target hypothesis must be a whole number from 1 to 6
  • NOT VALID: the belief level must be from 0.5 to 0.999
  • NOT VALID: the challenge number must be a whole number from 0 to 5, and the guess one of the listed choices
  • NOT VALID: give at least one hypothesis in use (H1 to HH) a prior weight above 0
  • NOT VALID: enter the likelihood of Ee under Hh; every hypothesis in use needs a value for every evidence type in use
  • NOT VALID: the likelihoods of Hh are all 0 for the evidence types in use; give at least one of them a value above 0
  • NOT VALID: observation i is Ex, but only E1 to EE are in use; remove it or add evidence types
  • NOT VALID: the target hypothesis is Ht, but only H1 to HH are in use
  • CHECK: observation i (Ex) is impossible under every hypothesis still possible, so the update stops before it: the observations before it are applied, and later ones are not.
  • CHECK: the likelihoods of Hh do not add to 100%, so they were rescaled to add to 100%: the row's sum differs from 100 by more than 0.5.
  • OK

References

This laboratory is an educational model of Bayesian updating over a finite set of hypotheses. Each observation multiplies every belief by the likelihood of the observed outcome under that hypothesis, and the beliefs are rescaled to add to one, so the posterior after one observation is the prior of the next. The lab reports the target's odds against the rest and each step's Bayes factor, the Shannon entropy of the beliefs in bits and the expected information gain of one more observation, and it restates the first update as natural frequencies in 1,000 people, rounded to whole people by the largest-remainder method.

Sources and further reading

  • Bayes' theorem, its odds form and the ratio of likelihoods, and learning by conditioning on evidence: J. Joyce, Bayes' Theorem, Stanford Encyclopedia of Philosophy, Metaphysics Research Lab, Stanford University, sections 1 and 2. Stanford Encyclopedia of Philosophy
  • Natural frequencies as a way to reason about a screening test: G. Gigerenzer and U. Hoffrage, "How to improve Bayesian reasoning without instruction: frequency formats", Psychological Review 102(4), 1995, pages 684-704.
  • Entropy in bits, H = -Σ p log2 p, introduced by C. E. Shannon, "A mathematical theory of communication", Bell System Technical Journal 27, 1948: Wikipedia, Entropy (information theory). Wikipedia
  • The expected information an experiment provides: D. V. Lindley, "On a measure of the information provided by an experiment", Annals of Mathematical Statistics 27(4), 1956, pages 986-1005.
  • The largest-remainder rule, which gives each group the whole-number part of its share and the units left over to the largest fractional remainders: United States Census Bureau, Congressional Apportionment: Historical Perspective, the Vinton or Hamilton method. U.S. Census Bureau

No table or text from any of these publications is reproduced, and none of their authors or publishers has reviewed or endorses this laboratory.

This laboratory is for learning. Its results hold only within its stated ranges; its hypotheses, likelihoods and observations are your own, so it is not a basis for a medical or any other real decision.

Links were current at build time. If one moves, search the document's title rather than reporting a broken link as a defect.

Results are informational and not professional advice. See our Terms of Use. Powered by SpreadsheetWeb · About · Privacy · Terms