Choice Architecture, Defaults and Auto-Escalation

Choice architecture is how a set of options is arranged in front of somebody choosing. Architecture cannot be avoided. Every arrangement is one arrangement among many, and no way of presenting a choice is neutral. So the question is never whether to have architecture. The question is who sets the arrangement and in whose interest, and asking it turns a technique into a question about conduct.

Most treatments of this subject open by asking whether arranging somebody's choices is acceptable. Acceptability is the wrong opening question, and getting it wrong is what makes the rest of the argument go in circles. The arranging has already happened by the time anybody chooses. Somebody decided what to list first, what to leave off, how many options to show, what happens if the form is never returned, and what the screen says before the button is pressed. Every one of those decisions was made by somebody, and there is no version of the form on which none of them was made. Thaler and Sunstein set this out in Nudge in 2008, and the reason the book has been argued about ever since is that people keep hearing it as a proposal when it is first of all an observation.

What is choice architecture, and can it be avoided?

The idea is not a financial one, and a version of it that only works on a portfolio has been written too narrowly, so the clearest starting point lies outside money entirely. A hospital form asks whether the person filling it in would like to register as an organ donor. One version says: tick here to join. Another says: the name is already on the register, and a tick is needed to come off. A third says nothing about it at all and waits to be asked. Nothing about the law has changed between the three. Nobody has been forbidden anything, and nobody has been made to do anything. All that has changed is what happens when the chooser does nothing, and what happens when the chooser does nothing is called the defaultWhat happens if the chooser does nothing at all. Every arrangement has one, including an arrangement nobody designed..

Now for a reasonable-sounding question that is not: could the form have been written with no arrangement? No form could. A form has to be printed in some order, with some number of boxes, with something happening when it is not returned. An unchosen arrangement is not an absent arrangement. An unchosen arrangement is one nobody thought about. That is worse rather than better, because an arrangement nobody thought about was still set by somebody, and that somebody was optimising for printing costs, or for the software they already had, or for nothing at all. The choice available to a designer is between a considered arrangement and an unconsidered one.

Why does a small amount of effort not explain how much this matters?

Here is where most readers stop believing the subject is interesting, and the objection they raise is a good one. Surely a default only matters because changing it is a nuisance. Tick a box, sign a form, wait on a phone line. People are busy and lazy, the reasoning goes, so of course the pre-ticked option wins, and there is nothing deep here at all. If that were the whole story, the fix would be obvious: make switching easier and the effect goes away.

Johnson and Goldstein, writing in Science in 2003, tested exactly that on organ donor registration across countries. Some countries used an opt-inAn arrangement that requires an action from anyone who wants to take part. Doing nothing leaves the person out. arrangement, where joining the register required an action. Others used an opt-outAn arrangement that requires an action from anyone who wants to stop taking part. Doing nothing leaves the person in. arrangement, where leaving it required an action. The paperwork involved on either side was a few minutes at most. The differences between the two kinds of country were far too large for a few minutes of form filling to account for, and that mismatch, not the direction of the effect, is the finding that matters. When the size of a cause and the size of its effect do not match, the cause has been misidentified.

The arithmetic of inconvenience is worth setting out. Suppose a switch takes ten minutes and the time is valued generously. The ten minutes are real, and they will stop a few people. Ten minutes will not stop most of a population from taking something they say they would rather have. Something else is doing the work, and the reason it matters which something is that the fixes are completely different. If effort were the mechanism, better forms would solve it. Since effort is not the mechanism, better forms will not.

What does explain it, if effort does not?

Three explanations do reach, and the useful thing about them is that they are separate. Knocking one down leaves the other two standing, and the effect survives every attempt to design it away.

The first is that a default reads as a recommendation. Whoever set it presumably knew something the chooser does not, so departing from it feels less like exercising a choice and more like overruling somebody who had the file open. Nobody says this out loud. The reading shows up as a slight hesitation and a thought along the lines of, they must have had a reason. Samuelson and Zeckhauser, in the Journal of Risk and Uncertainty in 1988, gave the general form of this stickiness the name status quo bias, and the recommendation reading is one of the reasons it is so hard to shift.

The second is that a default fixes the reference pointThe position a person measures gains and losses from. Change the position and the same outcome can be scored as either.. Once a person is in an arrangement, that arrangement is where the counting starts from, and anything that would have to be given up to move counts as a loss rather than as a smaller gain. Kahneman and Tversky set out this asymmetry in Econometrica in 1979, and it produces something with real bite: losses are weighted more heavily than gains of the same size. The Palash decision log, an invented record of 240 decisions taken by 60 investors over eight quarters, measured that weighting on its own 60 people. Asked what gain would make a fifty-fifty gamble against a Rs 10,000/- loss worth taking, the median answer was Rs 22,000/-, so the measured coefficient is 22,000 divided by 10,000, being 2.2. Putting a default in place does not just make one option easy. Moving now means giving up what is already held, so every other option starts to cost something.

The third explanation is the quietest and, for a lot of people, the strongest. A default removes the decision. Not the option, the decision. Nobody had to sit down and weigh anything, so there is no moment at which the chooser can later be shown to have chosen wrongly. An unmade decision cannot be got wrong, and to somebody who would rather not be responsible for an outcome, that is worth a great deal. This is not laziness and it is not stupidity. Declining to decide is a rational response to being held accountable for outcomes nobody can control. Most people occupy that position at work and a good many occupy it at home.

Set the effort of departing to nothing and see what happens. In the illustration below, 100 people are shown a default. Of those 100, 50 genuinely prefer the alternative on the merits, and that 50 is held still while the effort moves. If effort were the whole mechanism, then at zero effort all 50 of them would depart. In practice the share that departs runs at 20 of the 100 even when departing costs nothing, leaving a gap of 30 people. The 30 person gap is what the three explanations above are for, and the gap is still there when the effort has been reduced to nothing at all. The 100, the 50, the 20 and the 30 are a made-up illustration, chosen to make the shape of the argument visible, and are not measured figures.

What is auto-escalation, and what problem was it built for?

Asking somebody to put more away each month starting now is asking them to give something up today, at the moment they are being asked. The request is answered with today's feelings about today's money, and today's money always has somewhere to be. Agreeing costs nothing at the moment of agreeing, so asking the same person to put more away starting at a future date brings an answer from a completely different place. The gap between today's money and a future date is the whole of the idea, and Thaler and Benartzi built an arrangement on it in Save More Tomorrow, in the Journal of Political Economy in 2004.

Auto-escalationAn arrangement under which a contribution rises automatically at future dates unless the person stops it. works in two moves. The first move puts the cost in the future, where it is easy to accept. The second move is the one people miss: when the future date arrives, the increase happens unless the person stops it. So the arrangement made at the future date is itself a default. Auto-escalation is a default applied to a default, and a default applied to a default gets agreement where a direct request for the same increase does not. The third explanation is doing all the work here. Nobody has to decide anything at any point, so nobody can be wrong at any point.

Take the household in the invented decision log. Meera Sundaram, 41 and salaried, has Rs 25,000/- a month going into her investments by standing instructionAn arrangement with a bank or a provider that repeats on a schedule until somebody stops it., against monthly outgo of Rs 55,000/- and a reserve of Rs 1,10,000/-, or two months of outgo. Ask her today to raise the Rs 25,000/- to Rs 32,500/- and she will look at this month and say no. Offer her three future increases of Rs 2,500/- each, arriving at Rs 32,500/- in the end, and the answer at the moment of asking is about a month she has not lived through yet. The Rs 2,500/- steps and the Rs 32,500/- figure are a made-up illustration, and the decision log records no escalation.

Where does an arrangement stop being architecture and become a promotion?

Everything so far has been a technique, and a technique is neither good nor bad. The same two moves that get somebody to put more away each month will get them into a scheme that pays whoever arranged it. Nothing in the mechanism cares which. The indifference of the mechanism is the uncomfortable part of the subject and the part worth sitting with. Behavioural work cannot settle the question on its own.

Choice Architecture vs Financial Promotion

A financial promotionA communication designed to bring about a particular financial action by the person receiving it. is a communication meant to produce a particular financial action. Choice architecture is how options are arranged when they are presented. Set side by side, the two descriptions show that the second does not exclude the first. An arrangement can be built to produce a particular action, in which case it is doing both jobs at once. An arrangement set in the chooser's interest and one set in the setter's interest can be pixel for pixel identical, so what the arrangement looks like does not distinguish the two. They are distinguished by whose interest decided the shape, which is not visible on the screen at all.

Which four tests separate the two, and why must all four hold?

Four questions do the separating, and the discipline is that all four have to hold. Three out of four is not a pass. The one that fails is usually the one that mattered. The four are asked in this order, of any arrangement, including one a person is about to set for themselves.

One. Was the default set in the chooser's interest, judged by what the chooser needs rather than by what the setter would like them to want? Two. Is the arrangement disclosed, so the chooser can find out what happens if they do nothing without having to work it out from the outcome? Three. Can it be reversed at no cost, quickly, without a phone call that never gets answered? Four. Does whoever set the default gain from it being kept? The first three can all hold while a payment quietly turns on whether the box stays ticked, so the fourth catches most cases. That fourth question is also the only one of the four that the chooser can almost never answer for themselves, which is why it has to be asked out loud.

How would anyone know a default was working, and why is that hard?

The most useful turn in the subject is the one from technique to measurement problem. Suppose a default has been set and the question is whether it is holding. The record is where the answer is sought. The Palash decision log has 240 logged decisions from 60 investors over eight quarters, and its composition is fixed: 96 buys, 84 sells, 36 switches and 24 pauses of a standing instruction, summing to 240. The 24 pauses divided by 240 is 10.0 per cent.

Reaching for the complement gives 90.0 per cent, and the complement is the trap. The complement does not follow at all. The other 216 decisions are 96 buys, 84 sells and 36 switches, not the default holding. Buys, sells and switches are different decisions entirely and say nothing about the standing instruction one way or the other. And the months in which the instruction simply ran are not in the log at all, because continuing generates no decision and a log records decisions. Nobody logged a line saying nothing happened this month.

So the log counts departures from the default and never the default itself. Counting only departures is not a defect in this particular record. Counting only departures is the general case, and the same fact makes a default powerful in the first place. The path that requires nothing also produces no evidence of having been taken, so the more effective a default is, the less trace it leaves. The consequence, stated plainly: 24 pauses establishes that a pause was logged 24 times, and nothing whatever about how often the default held.

Who sets the default, and in whose interest?

Whose interest set the arrangement is the question the whole subject reduces to, and the question has an awkward property: the person best placed to answer it is the person with the least reason to. Somebody sets the arrangement. Whoever sets it has an interest, even if the interest is only wanting a quiet week. Where the interest and the chooser's interest point the same way, nothing much happens. Where the two interests point different ways, the arrangement will drift towards the setter. A hundred small design decisions each go the easy way, so the drift needs nobody deciding to be dishonest.

Devika Rao, the adviser at the invented Palash Advisory Services Private Limited, is in this position every working day. The scheme document in the same record runs to 46 printed sides with the risk statement on the 31st in eight point type; when 30 readers were tested, 7 of them could state the main risk afterwards, being 23.3 per cent, against 24 of 30, being 80.0 per cent, for a 90 word version placed at the top. Nobody hid anything. The risk statement was there both times. Where in a document a fact is placed is a design decision. A design decision made by somebody with nothing to gain from the fact being read will drift. The same record also shows that when 30 readers were shown 214 options only 11 chose, while 21 of 30 chose when shown 7, so the count of options presented is itself part of the arrangement.

How does somebody setting an arrangement actually use all this?

Three readers use this differently, and all three are worth walking through. A person deciding alone, with no adviser and no committee, uses it on themselves. Nobody can stop being subject to somebody else's arrangement, but a person can set their own before tiredness arrives: a standing instruction that runs on a date nobody has to think about, and a written note of what would make them stop it. The written note is what survives a bad month, and the note works exactly the same way for a career move or a large purchase as for a monthly investment.

An adviser or a plan designer uses it as a checklist before setting anything. The four tests get written down with the answers, and the fourth answer names who is paid and on what. Devika Rao would find the discipline mildly uncomfortable, and the discomfort is the point: an arrangement she cannot answer test four about cleanly is one she should not be setting. Setting for one person is where suitabilityWhether an arrangement fits the particular person it is being offered to, rather than being sound in general. meets architecture. A default that suits most people is still the wrong default for the person in front of the adviser, and the default is applied to the person in front of the adviser.

An analyst or a supervisor reading somebody else's arrangement uses the measurement point hardest. A default that works perfectly generates no rows at all, so when a number arrives showing that a default is working well, the first question is what event generated each row of the record. The right question to a provider is not what share of people stayed. Ask instead what the provider's system writes down when nothing happens, and if the answer is nothing, on what basis the count is made. A record that only fires on departure can establish that departures happened and cannot establish what the rest of the population did.

Which rule governs any of this, and where is it set?

No rule is stated above, and the silence is deliberate rather than an omission. Everything above is the behavioural case: what an arrangement does to behaviour, why the size of the effect rules out the obvious explanation, and which questions separate an arrangement built for the chooser from one built for the setter. None of that is a statement about what is permitted.

Meera Sundaram, Devika Rao, Palash Advisory Services Private Limited and the Palash decision log are invented.
Educational material. Not advice on any investment, tax, budget or market position.