The Bigger Picture.

Cue

The Diagram, someone who would settle the reversal by a cause. The two true readings at odds, the lurking variable behind them, the arrows there to be drawn and the assumption there to be staked.

Action

Adjudicating: The Modeller, holding the modeller's seat, draws the causal structure to pick the level to trust, against the pull of the authoritative total and the honest split.

Stage

The Modeller, on Draw the Cause, sets the arrows between the variables, judges which is confounder to adjust and which mediator to leave, and chooses the level the model licenses. The paradox can be resolved this way, but only by a causal claim brought from outside, one the table can neither prove nor refute.

Tension

Drawing the cause is the one move that can choose between the levels, and it stakes the answer on an assumption the data cannot check. There is nothing to fault: the choice of level is a causal question and the modeller answers it in the only currency that can. But the arrows are guessed, not measured, so the plot closes one of two ways: the structure is drawn correctly and the right level chosen, bought against a cause the numbers cannot confirm ... or the arrows are wrong, the wrong variable adjusted, and the reversal reversed once more with equal confidence.

Cue

The Breakdown, a third variable tied to both the grouping and the outcome. The world assigning cases to groups unevenly, the hard cases to one treatment, the many applicants to the selective departments, the subgroups filling unalike.

Action

Sorting: the lurking variable loads the subgroups differently, so the groups are no longer comparable and the setup for a reversal is laid.

Stage

The world sorts the cases the way it does, sicker patients to the newer treatment, more applicants to the harder programmes, so the subgroups differ in more than the thing being compared. No one arranges it: the uneven allocation is the ordinary structure of how cases fall, and it quietly makes the groups unalike beneath any comparison drawn across them.

Tension

The sorting is the world behaving as it does, and the world behaving as it does is what will make the total mislead. There is nothing to fault: the uneven allocation is real structure, not a bias introduced. But it loads the groups differently, so the plot closes with the subgroups made incomparable and the setup for a reversal laid, the confounder's uneven hand in place and no trace of it yet in any total.

Cue

The Breakdown, the same data read group by group. The total's one clean answer on the table, the subgroups there to be read apart, the suspicion that the parts may not behave like the whole.

Action

Splitting: The Splitter, holding the splitter's seat, separates the subgroups and finds the opposite trend in every one, equally honest.

Stage

The Splitter, on Split It Out, breaks the table into its groups and reads the trend within each, and every single subgroup points the other way from the total. Both readings stand, computed from the identical rows: the whole says one thing and each part says the opposite, and neither is a miscount.

Tension

Disaggregating is the rigorous move, and reading every part correctly settles nothing about which level to trust. There is nothing to fault: the strata are honestly computed and truly hold. But whether the grouping should be adjusted away or kept turns on a cause the data does not hold, so the plot closes with the parts and the whole openly at odds, the reversal laid bare and neither level able to prove it is the one to act on.

Cue

The Ledger, all the rows gathered and the whole there to be read. The subgroups summed into one table, the larger sample to hand, the instinct to step back to the fuller picture.

Action

Pooling: The Pooler, holding the pooler's seat, combines the data and reads the aggregate trend, one clear direction with the authority of the big number.

Stage

The Pooler, on Trust the Total, sums every subgroup into the total and reads the overall trend, honestly and completely, from all the data there is. The number is large and clean and points one way, and the completeness that built it makes it feel like the truth risen above the noise of the parts.

Tension

Reading the biggest honest number is the sound instinct, and the pooling that built it mixed groups that should not have been mixed. There is nothing to fault: the aggregate is computed correctly from every row. But the subgroups were loaded unevenly, so the plot closes with the total giving its confident answer, authoritative and clean, over a confounder it averaged away and cannot show.

Cue

The Ledger, a decision that will not wait and two readings that disagree. The total authoritative and clear, every subgroup against it, a treatment to choose or a policy to set, and no time to wait for a certainty that will not come.

Action

Pooling: The Decider, holding the decider's seat, acts on the aggregate and against every subgroup, trusting the larger sample and the clearer whole.

Stage

The Decider, on Act on the Number, reads the biggest, most authoritative summary and commits, taking the total as the truth to act on because rising above the slices feels like rising above the noise. The call is made on the whole and against the parts, confidently, and nothing on the page can say whether the level chosen was the right one.

Tension

Acting on the clearest, largest number is the sound instinct, and it commits against everything every subgroup shows. There is nothing to fault: a decision must be made and the aggregate wears the authority of size. But the data cannot license a level, so the plot closes with the verdict taken on the total, right or wrong with no way to tell, the accountability that forced the call discharged on a number that could not ground it.

Tenor

Authoritative, reasonable, above the noise. The plain good sense of stepping back to the whole, reading the big number built from all the data, and trusting the larger sample over the smaller slices. Nothing false in it: the aggregate is honestly computed and really is built from every row, and trusting the fuller picture is a sound instinct.

Undertow

Under the authority of the whole, a lurking variable that sorted the cases unevenly into the groups, so the total points the opposite way from every subgroup. Because the groups are not alike and pooling mixes them with unlike weights, the aggregate can reverse the trend that holds in each part, and both readings are correct, and which should guide the decision turns on a cause the numbers do not carry. The bigger picture that feels like rising to the truth is a mixture that can point the wrong way.

Nerve

The feeling of stepping back to the whole and trusting the big honest number, sure you are seeing past the noise, because the aggregate is larger and computed from all of it, while every subgroup in the same data says the opposite, and the table cannot tell you which level to believe.

Echo

After the run: a decision taken on the pooled total and against every subgroup, right or wrong with no way on the page to know, no one miscounting and no one lying, only an honest aggregate and honest strata that disagree, both true of the level they describe, and the one thing that would settle it, the causal structure that says which variable to adjust, brought in from outside or not at all, unprovable by the data it must interpret.

Projection

Decisive. Responsible. Needing an answer. Is the one who cannot stare at the table forever, who has a treatment to choose or a policy to set and must read the data and act. No rashness in it: only someone accountable for a decision, taking the clearest, largest, most authoritative number and moving on it.

Action

Acts on the total, choosing the level that carries the most data and the plainest signal, and setting the policy or the treatment accordingly. Takes the aggregate as the truth to act on, because it is built from everything and rising above the slices feels like rising above the noise. Does what a decision-maker must: reads the clearest summary and commits.

Shadow

Not a reckless chooser, an accountable one, and that is the whole difficulty: acting on the biggest honest number is the sound instinct, and it can act against everything every subgroup shows. Because the data cannot say which level to trust and the aggregate wears the authority of size, the Decider commits to the total and, when the grouping was the thing that caused, moves exactly the wrong way, confidently, or the right way, with no way to tell which from the table, so the accountability that demands a decision is discharged on a number that cannot license it. It acts on the clearest picture, and the clearest picture can be the reversed one.

Tell

Says "I have to choose, and I will go with the number built from all of it." Chooses in good faith, and cannot know, from the data alone, whether the whole or the parts should have carried the call.

Projection

Thoughtful. Causal. Looking behind the table. Is the one who asks not what the numbers say but what produced them, drawing the arrows between the variables to see which is confounder and which is cause. No mysticism in it: only the discipline of knowing that a table shows association and a decision needs causation.

Action

Draws the causal structure, deciding which variable to adjust away as a confounder and which to leave alone as a mediator, and so picks the level to trust. Brings to the table the one thing it lacks, a model of what caused what, and reads the reversal through it. Does what causal reasoning does: settles which level licenses the decision, by an assumption about the world the data cannot hold.

Shadow

Not an oracle reading truth from the figures, a modeller staking a claim, and that is the whole difficulty: the causal structure is what settles the paradox, and it cannot be proven by the data it interprets. Because the arrows are an assumption brought from outside and the table can neither supply nor confirm them, the Modeller who draws the wrong structure adjusts the wrong way and reverses the reversal with the same confidence as the right one, so the move that resolves the paradox also stakes the answer on a judgement the numbers will never settle, and being right requires a cause correctly guessed. It brings the cause the table lacks, and the cause it brings is a bet the table cannot check.

Tell

Says "the numbers cannot tell you: what caused the grouping decides which level is real." Says it, and must then commit to a causal story the data can neither prove nor refute, right only if the story is.

Projection

Sensible. Rigorous. Above the noise. Is the one who gathers all the rows into one table and reads the overall trend, trusting the larger sample and the fuller picture. No sloppiness in it: only a careful analyst computing the honest aggregate from every piece of the data there is.

Action

Pools the data and reads the total, combining the subgroups into one number and reporting the direction it points. Trusts the size of the sample and the completeness of the whole, and steps back from the slices to the picture they sum to. Does what a careful analyst does: builds the biggest honest number available and reads the trend it shows.

Shadow

Not a careless counter, a rigorous one, and that is the whole difficulty: the aggregate is honestly computed and truly built from all the data, and it can still point the opposite way from every part. Because a lurking variable has loaded the subgroups unevenly, pooling them mixes unlike things with unlike weights, so the Pooler's honest total can reverse the trend that holds in each group, and the very completeness that makes the number authoritative is what lets the confounder flip it. It reads the whole correctly, and reading the whole correctly can be reading the wrong thing.

Tell

Says "look at the big number: it is built from all the data, so it is the one to trust." Built it honestly, and cannot see that a whole built from unlike parts can say the opposite of every part.

Projection

Careful. Granular. Alert to the detail. Is the one who breaks the table into its groups and reads the trend within each, unwilling to let the slices be blurred into one. No pedantry in it: only a careful analyst who sees that the parts may not behave like the whole and looks at each on its own.

Action

Splits the data by group and reads each subgroup's trend, reporting that within this stratum and that one the direction is the same, and opposite to the total. Holds that the parts deserve their own reading, and refuses to let an average hide what every group plainly shows. Does what a careful analyst does: disaggregates, and lets each subgroup speak for itself.

Shadow

Not a nitpicker splitting hairs, a rigorous reader of the parts, and that is the whole difficulty: each subgroup trend is honestly computed and truly holds, and it can still be the wrong level to act on. Because whether the grouping should be adjusted away or kept depends on what caused it, the Splitter's honest strata are the right guide only if the grouping is on the causal path, and the data cannot say that it is, so disaggregating correctly can be disaggregating on a variable that should have been left pooled, splitting the truth into the wrong answer with equal rigour. It reads the parts correctly, and reading the parts correctly can be reading the wrong level.

Tell

Says "do not blur the groups: every single one of them says the opposite of your total." Reads them truly, and cannot show, from the numbers alone, that the groups are the level the decision should be made on.

Shown

The place the table is broken into its groups, each subgroup laid out on its own with its own trend, and the uneven counts across them there to be seen. The splitter reading each stratum, the lurking variable's uneven load showing in the cell sizes, every group pointing the same way and against the total.

Holds

The sorting and the splitting, and the parts the whole hid. Here the data meets its groups, so the Breakdown holds the point where the strata speak apart: each subgroup read on its own and the uneven allocation shown, so it is where the confounder's fingerprints appear and every part points against the pooled total, the local trend plain where the aggregate blurred it.

Offers

The true trend within each group, and the sight of how unevenly they were filled. Because the Breakdown lays the subgroups out apart, it offers the splitter each stratum's honest reading and shows the uneven counts that loaded them, so it hands out the parts the total hid and the fingerprints of the confounder, and hands them out without the one thing that would say whether this grouping is the level to trust.

Withheld

Any proof that the split is the right level to act on. The Breakdown can show the true trend in every group and the uneven sorting between them, but it cannot say whether that grouping should be adjusted away or kept, so it withholds the causal fact that would license its level, leaving the strata true and mute on their standing, and, all parts and no cause, it withholds any sign that reading each group correctly can still be reading the wrong level of all.

Shown

The place the variables are drawn with arrows between them, a graph of what is taken to cause what, standing off the table of counts. The modeller setting the structure, marking one variable a confounder to adjust and another a mediator to leave, and reading the reversal through the arrows.

Holds

The adjudicating, and the cause the table lacks. Here the counts meet a claim about the world, so the Diagram holds the point where the level is chosen: the causal structure drawn and the confounder named, so it is where the reversal can be resolved, the arrows deciding what the numbers alone never could.

Offers

A way to pick the level, and the cause the counting left out. Because the Diagram carries what the table cannot, a model of what produced the data, it offers the modeller the ground to adjudicate the reversal and choose which level licenses the decision, so it hands out the resolution the numbers withheld, and hands it out resting on arrows that are assumed, not measured.

Withheld

Any proof that the arrows are right. The Diagram can settle which level to trust once the causal structure is drawn, but it cannot draw that structure from the data or confirm it against the table, so it withholds the certainty that the model is correct, leaving the resolution as sound as the guess beneath it, and, all cause and no check, it withholds any guarantee that the wrong arrows will not reverse the reversal with equal confidence.

Shown

The place every row is gathered into one, a single table where the data is summed and the overall trend read off with the authority of size. The pooler combining the subgroups into the total, the decider reading the big number and committing to it, the groups blurred into one figure.

Holds

The pooling and the deciding, and the authority of the whole. Here the data meets the aggregate, so the Ledger holds the point where the parts become one number: every subgroup summed into the total, so it is where the fuller picture is read and acted on, the larger sample always more persuasive than the slices it hides.

Offers

A single authoritative number, and the clarity of the whole. Because the Ledger gathers all the data into one, it offers the pooler the biggest honest figure and the decider the plainest signal to act on, so it hands out the authority of size and the relief of one clear answer, and hands it out having blurred the unlike groups whose difference would have changed the reading.

Withheld

Any sight of the confounder the pooling hid. The Ledger cannot both sum the groups into one authoritative total and show how unevenly they were loaded, so it withholds the lurking variable it pooled over, letting the aggregate stand clean and unqualified, and, all whole and no parts, it withholds any sign that the biggest, clearest number in the house is the one that can point against every group it was built from.

Place

Runs across all three: pooled at the ledger, split at the breakdown, resolved or guessed at the diagram. It is the whole circuit, the honest data that reads both ways and settles only on a cause.

Load bearing

Without it the pieces are separate: a total, some strata, a confounder, a reversal, a decision, with nothing to bind them. With it they are one mechanism: the lurking variable sorts the groups unevenly, the total pools them and points one way, the strata split them and point the other, the reversal stands, and the verdict acts on a level the data cannot license, each part driving the next. The Bigger Picture is load bearing because it is the structure entire (Simpson; Blyth; Pearl): the aggregation reversal and its irresolvability by data alone, so the piece carries the trap whole, the reason two honest readings of one table can contradict and no counting can choose between them.

Apparent

Good analysis, the plain competence of reading the data and acting on the clearest number. Seen at the surface, the whole trap looks like sound quantitative work, evidence gathered and a decision grounded, and the truth that the same data supports opposite calls, that the level to trust turns on a cause the numbers do not carry, that the clearest picture can be the reversed one, stays hidden behind how much acting on the biggest honest number looks like exactly the right thing to do.

Yearbook

Assembled from every element: the lurking variable that sorts, the total that pools, the strata that split, the reversal between them, and the verdict taken on one. It is the trap named: the Bigger Picture binds sorting, pooling, and splitting into one reversal, resolved only where adjudicating draws the cause the numbers never held.

Place

Loads the groups unevenly beneath the breakdown, its fingerprints in the cell sizes. Named or missed only at the diagram, where its causal role is drawn.

Load bearing

Without it the groups would be alike, the aggregate and the strata would agree, and there would be no paradox. With it the reversal has its engine: a third variable, tied to both the grouping and the outcome, distributed unevenly across the groups, so pooling them weights unlike things and can flip the trend. The Lurking Variable is load bearing because it is the confounder the whole trap turns on (Simpson; Pearl): the hidden cause of the reversal, so the piece carries the third thing that makes the parts and the whole disagree, the variable whose uneven hand does the flipping and whose causal role, not the data, decides which level to trust.

Apparent

Nothing, an unseen third thing behind a clean two-way comparison, easy to leave out. Seen at the surface, the Lurking Variable does not appear: the eye is on the two variables compared, and the third stays out of the frame, and the truth that it loaded the groups unevenly, that it is the cause of the reversal, that whether it is a confounder to adjust or a mediator to leave decides the whole question, stays hidden behind how complete a two-column comparison looks without it.

Yearbook

Spread unevenly by sorting and read only through adjudicating, the confounder behind the reversal. It is the engine of the trap: the Lurking Variable loads the strata unevenly and lets the total reverse them, the hidden cause of the reversal, the third thing the bigger picture must name to escape and the numbers never show.

Place

Appears between the ledger and the breakdown, where the whole and the parts are read against each other. Resolved or missed at the diagram, where a cause would say which to believe.

Load bearing

Without it the whole and the parts agree, and there is nothing surprising and no trap. With it the paradox is sprung: the same honest data points one way pooled and the opposite way split, and both readings are correct, neither a miscount. The Reversal is load bearing because it is the paradox the trap is named for (Simpson; Blyth): the direction that flips between the aggregate and the strata, so the piece carries the contradiction between two true readings, the surprise that the parts and the whole of one honest table can disagree, and that the numbers cannot say which of them to act on.

Apparent

A mistake, the plain look of a figure that must have been miscomputed to contradict itself. Seen at the surface, the Reversal looks like an error to be found and fixed, and the truth that both readings are honestly correct, that the contradiction is a genuine feature of an unevenly confounded table, that no recount will remove it because it is not a mistake but a property, stays hidden behind how much a whole that contradicts its parts looks like something someone got wrong.

Yearbook

Laid bare when the total meets the strata, the flip driven by the lurking variable. It is the paradox of the trap: the Reversal sets the true whole against the true parts, sprung by the confounder and settled only by adjudicating, the contradiction the bigger picture is built on and the numbers cannot resolve.

Place

Laid out at the breakdown, where the data is read group by group. Against the whole read at the ledger, whose total they contradict.

Load bearing

Without them there is only the aggregate, and no parts to contradict it and no reversal to see. With them the trap has its other true pole: the trend within each subgroup, honestly computed and holding in every one, and opposite to the total. The Strata are load bearing because they are the parts the reversal sets against the whole (Simpson; Blyth): the within-group trends, equally correct and pointing the other way, so the piece carries the honest parts whose disagreement with the honest whole is the paradox itself, true of the groups they describe and unable to prove their level is the one to act on.

Apparent

The real story, the granular truth the crude aggregate blurred away. Seen at the surface, the Strata look like the finer and truer reading, the detail an average concealed, and the truth that reading each group correctly can still be the wrong level, that whether to trust the parts turns on a cause the numbers do not hold, that the granular reading is right only if the grouping is on the causal path, stays hidden behind how much a finer view feels like a truer one.

Yearbook

Read by the splitter through splitting, the within-group trends that hold in every part. They are the true parts of the trap: the Strata point against the total and lay the reversal bare, the honest slices the lurking variable loaded unevenly, one of the two readings the bigger picture cannot choose between.

Place

Read at the ledger, where every row is summed into one. Its opposite laid out at the breakdown, where the strata that contradict it appear.

Load bearing

Without it there is only a set of subgroups, and no whole to reverse them and no authority to persuade. With it the trap has its seductive pole: a single number built from every row, carrying the weight of the larger sample, that reads as the fuller and truer picture. The Total is load bearing because it is the aggregate the reversal turns on (Simpson): the honest pooled figure that can point the opposite way from every subgroup, so the piece carries the whole whose very completeness is what lets an uneven confounder flip it, the big number that feels the most right and can be the most wrong.

Apparent

The truth, the plain and authoritative answer built from all the data there is. Seen at the surface, the Total looks like the clearest view, a number too large and too complete to doubt, and the truth that it mixes unlike groups with unlike weights, that its completeness hides the sorting beneath it, that a whole built from incomparable parts can reverse them all, stays hidden behind how much a bigger sample feels like a better one.

Yearbook

Built by the pooler through pooling, the aggregate read with the authority of size. It is the seductive pole of the trap: the Total pools away the lurking variable and points against the strata, carrying the reversal into the verdict, the whole the bigger picture trusts.

Place

Handed down at the ledger, where the decision is made on the total. Its rightness settled only at the diagram, where a cause would say if the level was correct.

Load bearing

Without it the reversal is a curiosity, a table that contradicts itself and harms no one. With it the paradox bites: a decision is made on one level, a treatment chosen or a policy set, and it commits against the other level's reading. The Verdict is load bearing because it is where the trap costs (Bickel; Charig): the call taken on the aggregate against every subgroup, so the piece carries the decision that turns a contradiction on paper into an action in the world, the choice made on the level that persuaded and grounded on numbers that could not license it.

Apparent

A sound call, the plain result of acting on the clearest, largest number available. Seen at the surface, the Verdict looks like good decision-making, a choice made on the fullest evidence, and the truth that it committed against everything every subgroup showed, that whether it was right turns on a cause the data never held, that the same table would have justified the opposite call, stays hidden behind how responsible acting on the biggest number plainly looks.

Yearbook

Handed down by the decider carrying the reversal, the call made on the total. It is where the trap costs: the Verdict acts on the total against the strata, the decision that only adjudicating could have grounded, the action in the world the bigger picture takes on a level it cannot justify.

Has

A decision that will not wait and the clearest number to hand: the standing to choose a treatment or set a policy, and the pull of the largest, most authoritative summary. A clear case: a decision must be made, and the number built from all the data is the one to make it on. No way to tell, from the table, whether the whole or the parts should carry the call, and no time to wait for a certainty that will not come.

Orders

Act on the number. Read the clearest, largest summary, take the aggregate as the truth to act on, and commit, because rising above the slices feels like rising above the noise. Choose, and move.

Loses

The chance to be right for a reason it can defend, exactly by acting on the level that carries the most authority. Because the data cannot license a level and the aggregate wears the authority of size, the seat commits to the total and, where the grouping was the cause, moves exactly the wrong way, or the right way, with nothing on the page to tell which, so the accountability that forces a decision is discharged on a number that cannot ground it. It gains a decision made and a picture that looked clearest, and loses any way to know the decision was right.

Drives

Toward the call made on the total, bought against the subgroups it acts against and the cause it cannot see, so the seat that must decide commits to the level that persuades, not the level that is right.

Has

The variables and the question of what caused what: the standing to draw the arrows between them and read the reversal through a model of the world. A clear case: a table shows association and a decision needs causation, so the level to trust is chosen by the causal structure, not the counts. No way to read the arrows off the data, and no cover from the charge of guessing at a cause the numbers cannot check.

Orders

Draw the cause. Set the causal structure, decide which variable is a confounder to adjust away and which a mediator to leave alone, and pick the level the model licenses. Bring the table the one thing it lacks, a claim about what produced it.

Loses

The safety of a conclusion the data can back, exactly by settling the paradox with a cause. Because the causal structure is brought from outside and the table can neither prove nor refute it, the seat that draws the wrong arrows adjusts the wrong way and reverses the reversal with equal confidence, so the move that resolves the paradox stakes the answer on a judgement the numbers will never settle, and being right requires a cause correctly guessed. It buys a decision the model can license against a model the data cannot confirm, right only if the causal story it committed to is.

Drives

Toward the cause drawn and the level chosen, bought against an assumption the data cannot supply or check, so the one seat that can settle the paradox must stake it on a cause the numbers never carry.

Has

All the rows and the authority of the larger sample: the standing to gather the data into one table and read the trend it sums to. A clear case: more data is a fuller picture, and the number built from all of it deserves the most trust. No way to see the lurking variable in the total, and no sign, in the aggregate alone, that the groups it pooled were unalike.

Orders

Trust the total. Combine every subgroup into one table, read the overall trend, and step back from the slices to the whole they make. Take the biggest honest number as the truth, and rise above the noise of the parts.

Loses

Any sight of the confounder it pooled over, exactly by combining the groups into one. Because a lurking variable has loaded the subgroups unevenly, the seat's honest total mixes unlike things with unlike weights, so the completeness that makes the aggregate authoritative is what lets the confounder flip its direction, and reading the whole correctly can be reading the reverse of every part. It gains the authority of the larger number, and loses the difference between the groups that the larger number hides.

Drives

Toward the total read and trusted whole, bought against the confounder it pools away, so the seat that reads the fullest picture can read the reversed one.

Has

The groups laid out apart and the trend within each: the standing to disaggregate the table and read every subgroup on its own. A clear case: the parts may not behave like the whole, and an average that hides what every group shows is hiding something. No way to prove, from the numbers, that the grouping it split on is the level the decision belongs to.

Orders

Split it out. Break the table into its subgroups, read the trend in each, and refuse to let an average blur what every group plainly shows. Let each stratum speak for itself, and hold the parts against the whole.

Loses

The certainty that its level is the right one, exactly by reading the parts correctly. Because whether the grouping should be adjusted away or kept turns on what caused it, the seat's honest strata guide the decision only if the grouping is on the causal path, and the data cannot say it is, so splitting correctly can be splitting on a variable that should have stayed pooled, breaking the truth into the wrong answer with full rigour. It gains the true trend within each group, and loses any proof that the within-group trend is the one to act on.

Drives

Toward the data split and each part read true, bought against the proof that the parts are the right level, so the seat that reads every group correctly may still read the wrong level of all.

Initiated by

A causal question the counts cannot answer, and a modeller willing to draw it. At the diagram, the modeller setting the arrows between the variables, judging which is confounder and which is mediator, to decide which level licenses the decision.

Direction

Toward the paradox resolved, if the arrows are right. Because whether to adjust the grouping away or keep it turns on what caused it, adjudicating brings a causal structure to the table and reads the reversal through it, so it drives toward the level chosen and the decision grounded: but the structure is assumed, not measured, so the process resolves the paradox only by staking it on a cause the data can neither supply nor confirm.

Lever

The truth that only causation can choose between the levels. Because a table shows association and the choice of level is a causal one, drawing the cause is the one move that can settle the reversal, so adjudicating runs on the correct insight: yet the same move that resolves the paradox commits to arrows the numbers cannot check, so the process is right exactly when its guess at the cause is, and no more surely.

Echo

A level chosen and a decision licensed, or the reversal reversed once more. Either the causal structure is drawn correctly and the right level chosen, bought against an assumption the data cannot confirm, or the arrows are guessed wrong, the wrong variable adjusted, and the reversal reversed with the same confidence, the paradox settled only as well as the cause beneath it was known.

Initiated by

All the rows and the instinct to trust the larger sample. At the ledger, the pooler summing every subgroup into one table and reading the overall trend, the decider taking the big number to act on.

Direction

Toward one authoritative number that can point against every part. Because the subgroups were loaded unevenly and pooling weights them by their sizes, combining them mixes unlike things, so the aggregate trend can reverse the trend that holds in each group, and the total, built honestly from all the data, carries the authority of size while hiding the sorting that flipped it.

Lever

The soundness of trusting more data over less. Because a larger sample really is a fuller picture and the aggregate really is computed from every row, reading the total is the sensible instinct, so pooling runs on good statistical sense, the reversal assembled from no miscount, only from summing groups that should not have been summed as if they were alike.

Echo

A big honest number that points the wrong way, and feels the most right. The subgroups have been combined and the total read with the authority of the larger sample, so what reads as rising above the noise to the truth is, in truth, a mixture of incomparable groups, the clearest number in the house the one most able to reverse every part it was built from.

Initiated by

A third variable correlated with both the grouping and the outcome. At the breakdown, the world assigning cases to groups unevenly, the hard cases to one treatment, the many applicants to the selective departments, so the subgroups fill unalike.

Direction

Toward groups that are not comparable and a total that will mislead. Because the lurking variable is spread unevenly across the groups, sorting makes the subgroups differ in more than the thing being compared, so pooling them will average unlike things with unlike weights, and the setup is laid for the aggregate to point the opposite way from every part, without anyone arranging it.

Lever

The ordinary way the world assigns cases. Because sicker patients really are given the newer treatment and more applicants really do apply to the harder programmes, the uneven sorting is the world behaving as it does, not a bias introduced, so sorting runs on real structure, the confound assembled from no manipulation, only from cases falling into groups the way they actually fall.

Echo

Groups that cannot be fairly pooled, and no sign of it in the total. The cases have been sorted unevenly and the subgroups made unalike, so what reads, once summed, as one clean comparison is, in truth, a mixture of incomparable groups, the confounder's uneven hand hidden in a total that shows no trace of the sorting beneath it.

Initiated by

The suspicion that the parts may not behave like the whole. At the breakdown, the splitter breaking the table into its subgroups and reading the trend within each, unwilling to let an average blur what every group shows.

Direction

Toward the true local trend and the reversal laid bare. Because the subgroups were loaded unevenly, reading them apart shows the within-group trend the aggregate hid, so splitting reveals that every group points the opposite way from the total, and lays the reversal bare, each stratum honest and the whole and the parts now openly at odds.

Lever

The rigour of refusing to blur unlike groups. Because the parts genuinely may not behave like the whole and an average really can hide what every group shows, disaggregating is the careful, rigorous move, so splitting runs on good practice, the wrong-level risk assembled from no error, only from reading correctly a grouping the data cannot confirm is the one to act on.

Echo

Every part true and against the whole, and no proof of which to trust. The strata have been read apart and each shown to point against the total, so what reads as the granular truth the aggregate concealed is, in truth, one of two correct readings, honest within each group and unable, from the numbers alone, to prove that its level is the one the decision belongs to.

Direction

Read the clearest, largest summary and commit. Take the aggregate as the truth to act on, choose the treatment or set the policy, and move, because a decision cannot wait for a certainty that will not come and rising above the slices feels like rising above the noise.

Orders

(The Decider acts on the authoritative total and against every subgroup, in good faith, with nothing on the page to say which level was right.)

Implementation

(The Ledger; the pooling process; the call rests on the total and lands as the verdict, taken against the strata with no cause to license it.)

Targets

• [ ] the clearest, largest summary is read and committed to
• [ ] the aggregate is taken as the truth to act on
• [ ] the decision is made against every subgroup, in good faith
• [ ] the call is grounded on a number that cannot say whether the whole or the parts were right

Direction

Set the causal structure. Draw the arrows between the variables, decide which is a confounder to adjust away and which a mediator to leave alone, and pick the level the model licenses. Bring the table the one thing it lacks, a claim about what produced it, even though the claim cannot be read off the data.

Orders

(The Modeller draws the causal structure to pick the level to trust, at the cost of an assumption the data can neither supply nor confirm.)

Implementation

(The Diagram; the adjudicating process; the causal model reads the lurking variable as confounder or mediator to resolve the reversal, turning the bigger picture back ... if the arrows are guessed right.)

Targets

• [ ] the causal structure is drawn and the arrows between variables set
• [ ] each variable is judged confounder to adjust or mediator to leave
• [ ] the level to trust is chosen by the model, not the counts
• [ ] the answer is staked on a cause the data cannot confirm, right only if the structure is

Direction

Break the table into its subgroups and read the trend within each. Refuse to let an average blur what every group plainly shows, hold the parts against the whole, and let each stratum speak for itself, even against the authority of the total.

Orders

(The Splitter disaggregates the data and reads each subgroup, honestly and equally true, the reading that points the opposite way from the pooled total.)

Implementation

(The Breakdown; the splitting process; the reading lays out the strata and shows the reversal against the total, true within each group and mute on which level should rule.)

Targets

• [ ] the table is broken into its subgroups and each trend read
• [ ] every stratum is shown to point the opposite way from the total
• [ ] the parts are held against the whole and not blurred into an average
• [ ] the split is left unable to prove, from the data, that its level is the one to act on

Direction

Combine every subgroup into one table and read the overall trend. Trust the larger sample, step back from the slices to the whole they make, and take the biggest honest number as the truth. Rise above the noise of the parts to the picture they sum to.

Orders

(The Pooler pools the data and reads the aggregate, honestly and with the authority of size, one of the two correct readings that will disagree.)

Implementation

(The Ledger; the pooling process; the reading builds the total over the lurking variable, so the aggregate can carry the reversal against every part.)

Targets

• [ ] every subgroup is combined into one table and the total read
• [ ] the larger sample and the fuller picture are trusted over the slices
• [ ] the aggregate trend is reported with the authority of the big number
• [ ] the lurking variable is pooled away unseen, so the total may point against the parts