Complaint Behaviour: Who Complains, and Who Silently Leaves

Faced with dissatisfaction, a person can leave, speak up, or stay put and say nothing. Only speaking up writes anything down. A complaint count therefore captures one response out of three and stays completely silent about the other two, so it measures not how many people were unhappy but how many unhappy people decided to say so.

There is a particular kind of number that feels like evidence and is not. The number arrives with a date, a total and often a trend line, it can be put in a table and compared across periods, and everybody in the room treats it as a measurement of something real. The complaint countHow many people complained, not how many were dissatisfied. is the cleanest example in the whole of conduct. A complaint count is a measurement of complaining, and complaining is a behaviour rather than a condition. The condition worth measuring is dissatisfaction, and the two come apart so easily that the number can move a long way in either direction while the condition sits perfectly still.

Start away from finance. Everybody has been on both sides of this. A tea stall near a bus depot serves a hundred people in a morning and the tea is bad on a particular Tuesday. Some people put the cup down, walk off and never come back, and the stall keeper never learns why the queue is shorter next month. A couple of people say something at the counter. Most of the rest drink it anyway, mutter to a friend, and turn up again the next day because the stall is on the way and the alternative is a ten minute walk. If the stall keeper counted complaints on that Tuesday he would have a number, and the number would be two, and two is not the answer to any question he cares about.

What can somebody actually do when they are dissatisfied?

Albert Hirschman set the structure out in Exit, Voice, and Loyalty in 1970, and it has survived because there are genuinely only three moves available. A person who is unhappy with an organisation can leave it. Hirschman called leaving exitLeaving, without necessarily saying why.. Staying and saying something he called voiceComplaining, the only one of the three that leaves a record.. Or they can stay and say nothing. Hirschman called that third move loyaltyStaying quietly despite dissatisfaction.. The word does more work than it looks like it should. Staying quietly can come from attachment, from inertia, from having nowhere else to go, or from a belief that speaking would achieve nothing.

Hirschman's interest was in when a person picks each one, and the answer he gave is that exit and voice compete. Where leaving is easy, people leave instead of speaking, and the organisation loses the information that speaking would have given it. Where leaving is hard, people speak. Speaking is the only move left. The finding is genuinely uncomfortable for anybody who assumed complaints go up when service gets worse. Complaints go up when speaking is easier than leaving. The count is therefore a fact about the exit and about the channelThe route a complaint takes. Its ease changes the count. and only partly a fact about the service.

Notice what has already happened to the arithmetic. Three responses, and only one of them produces a document. Exit produces an absence. Loyalty produces nothing at all, not even an absence. The person is still there, behaving exactly as a satisfied person behaves. So of the three ways a dissatisfaction can express itself, the record captures one, and the other two are not badly captured or partially captured but wholly missing.

Two of those four are studied things rather than guesses, and worth naming properly. Samuelson and Zeckhauser described in 1988 how strongly people are pulled towards leaving an arrangement as it is. Staying put wins by default rather than by decision. And the last of the four, the belief that speaking changes nothing, is the one that turns a data gap into a self-sealing one: the less anybody responds, the fewer people speak, and the fewer people speak the less there is to respond to.

Why does only one of the three leave anything behind?

The gap is not an accident of how complaint systems happen to be built, and no amount of building them better fixes it. A record is written by an event, and the event is somebody doing something the system was designed to capture. A complaint record captures the act of complaining. Capturing that act is the record's entire remit. The record was never pointed at dissatisfaction, only at the narrow behaviour of registering a grievance, and the gap between the two is where the whole difficulty sits.

The same shape appears in what a decision log does not contain. A log of decisions taken records the buying and the selling and the switching, and holds nothing at all about the mornings when somebody looked at a position, thought about acting, and did not. Inaction leaves no trace because no event fired. Any population defined by not doing something is invisible in a record generated by doing it. Improving the record cannot reach them. The record was never in the place where they were. A complaint record and a decision log fail in exactly the same way for exactly the same reason.

So the phrase to be careful with is the one that sounds most reasonable: the complaints data should be improved. Cleaning up categories, adding fields, training people to write better summaries and reconciling the totals all make the recorded complaints more useful. Not one of those steps touches a person who never complained. The quality of a record and the coverage of a record are separate properties, and it is entirely possible to hold a beautifully kept record of a fraction of the thing worth knowing about.

Who actually complains, and who does not?

If complaining were free, the twenty would be a random fifth of the hundred and the record would at least be a fair sample of a small size. Complaining is not free. Complaining costs three separate things, and none of the three is spread evenly across any real population.

The first cost is time. Somebody has to find the right address, assemble what happened in the right order, retrieve the dates and the amounts, write it out, send it, and then hold the thread open through whatever comes back. Anyone who has tried to get a wrong electricity bill corrected knows that the hard part is not the argument, it is the four separate afternoons the argument eats. A person working two shifts does not have four afternoons.

The second cost is confidence, meaning the willingness to say plainly that something was done badly and to keep saying it to somebody who disagrees. Confidence is partly personality and largely practice. Somebody who has successfully had a bill corrected before begins the second one already believing it can be done. Somebody who has never once been listened to begins from the other end.

The third cost is the belief that it will work at all, and it is the one that quietly does the most damage to the record. Nobody spends time and confidence on a process they expect to lose. The people most certain that complaining changes nothing are the people who complain least. The record is emptiest exactly where the sense of futility is strongest. That is not a random gap in the data. The gap has a direction.

The three costs in series give the shape of the filter. A person reaches the record only by paying all three, and each one removes somebody. The population that comes out the far end is not a smaller version of the population that went in; it is systematically the part of it that had the time, the practice and the expectation of being heard. Filtering of that kind is what selectionWhich part of a population reaches a record, here the part that spoke. means here, and it is the reason a bigger record does not fix the problem: a larger sample of a filtered population is still filtered.

What does a complaint count actually measure?

The arithmetic works only with a population visible in full, and nobody in real life ever has one. The worked case therefore starts from a hundred people who are dissatisfied.

Of the hundred, suppose twenty complain, fifty leave without saying anything, and thirty stay put and say nothing. The complaint count is twenty. Twenty is the only number anybody downstream will ever see, and it says very little on its own. A number without a denominator is not yet a measurement.

Give it a denominator and it turns into two completely different statements. As a share of everybody who was dissatisfied it is twenty divided by a hundred, or 20.0 per cent. As a share of everybody who did something about it, meaning the twenty who spoke plus the fifty who left, it is twenty divided by seventy, and since twenty divided by seventy is 0.28571 that comes to 28.6 per cent. Same twenty people. Two answers, and they are not close to each other.

The record contains the twenty and nothing else, so neither denominator is available to whoever holds it. Leaving writes nothing, so there is no line for the fifty who left. The record has no line for the thirty who stayed. Staying quietly is indistinguishable from being content. The hundred is not a number anybody could look up. Finding it is not hard or expensive. No system holds it in any form.

Why can a rising complaint count be good news?

The consequence that follows catches almost everybody. The hundred stay exactly as they are. Nobody becomes more dissatisfied, nobody becomes less dissatisfied, and nothing about the service changes at all. The only thing that changes is the channel. Complaining becomes easier: a shorter form, an address on the first sheet of the document instead of the thirty first, somebody answering the phone.

Ten of the fifty who would have left without saying anything now say something instead. The count goes from twenty to thirty. Measured the way a count is usually reported, that is a rise of ten on a base of twenty, or 50.0 per cent. A complaint count that rose by half is entirely consistent with a dissatisfied population that did not move by one person. It was a hundred before and it is a hundred after.

The reverse does real harm. The channel is made harder instead: the address moves, a reference number nobody has is added, the form comes off the opening sheet. Ten of the twenty who were going to complain give up and simply leave. The count falls from twenty to ten, a halving, and every dashboard in the building shows a line going in the direction everybody wanted. Nothing improved. Ten more people walked out of the door than before, the opposite of an improvement, and the number that was supposed to detect it is the number that fell.

The picture repays a moment longer than feels necessary. It holds the whole relationship between the count and the condition in one frame. The dashed line never moves. The bars move by a factor of three. Everything a reader of the record can see lives in the bars, and everything they wanted to know lives in the line, and there is no operation on the bars that recovers the line.

The second thing the picture makes plain is that the two directions are not symmetric in how much harm they do. A count that rises because the channel improved is merely misleading; somebody may worry unnecessarily, and the worry is at least pointed at a real population. A count that falls because the channel got worse is actively dangerous. It produces relief at precisely the moment when more people are walking out without a word. The falling number and the improving service look identical in the record, and so do the falling number and the obstructed channel.

The honest reading of a complaint count therefore always requires a second question that the count itself cannot answer: what happened to the channel over the same period? If the address moved, if a form was shortened, if a phone line opened or closed, if a reminder started going out, then the count and the previous count are not measuring the same thing and comparing them is arithmetic performed on two different instruments.

A complaint channel deserves a blunt description. A channel is not a neutral pipe that either exists or does not. A channel is a designed route with a length, a location and a tone, and every one of those choices changes how many people travel it. Thaler and Sunstein made the general point in Nudge in 2008: there is no neutral way to lay out a choice. Some layout has to be picked, and every layout leads somewhere. A complaint channel is that argument applied to a route rather than to a menu.

Bertrand and Morse showed the same thing empirically in 2011, in their study of disclosure and payday borrowing in the Journal of Finance: moving information and changing how it was expressed changed what borrowers went on to do, without a single new fact being added. If where a warning sits changes behaviour, then where a complaint address sits changes behaviour too, and the count is downstream of that placement rather than independent of it. Somebody who designs the channel is, whether they intend it or not, setting the value of the number that will later be used to judge them.

What goes wrong once the count becomes the target?

Everything above is a measurement problem. Measurement is not the end of it. The next problem is behavioural and worse. Nobody has to behave badly for it to happen. Suppose the count is adopted as the thing to be improved. The count appears on a monthly sheet, somebody is asked about it, and a lower number is understood to be better. Ask now what the cheapest available way to lower it is.

The cheapest way is not to serve people better. Serving people better is slow, expensive and uncertain. The cheapest way is to make complaining harder. Making complaining harder is fast, nearly free and completely certain in its effect on the number. Lengthen the form. Require a reference that people do not have to hand. Move the address deeper into the document. Route the phone through four menu levels. Every one of those moves the metric in the desired direction, and not one of them touches the condition the metric was chosen to represent.

Why is the record thinnest where the duty is heaviest?

Go back to the three costs. Complaining takes time, it takes the confidence to say plainly that something was done badly, and it takes some belief that saying so will change the outcome. Now ask who is short of those three, and the answer is not a random slice of anybody. The people short of all three are the person working two jobs, the person who has never once been listened to by an institution, the person who is unwell, the person doing this in a second language, the person who is recently bereaved and dealing with paperwork they did not choose.

The heaviest care is owed to those same people. The overlap is not a coincidence and it is not a data quality issue. The complaint record is systematically emptiest for exactly the population whose position most needs to be known. A measurement bias has become a matter of conduct. Somebody reading a thin file about a group and concluding that all is well there has read the file precisely backwards.

Agarwal, Driscoll, Gabaix and Laibson set out in The Age of Reason in 2009 that the capability to handle financial decisions does not stay flat across a life, and declines in that capability sit alongside a reduced willingness to take on an institution in an argument. Wherever capability is lower, the cost of the three tolls is higher, and the record thins in the same place. Ageing is one worked example of a general shape rather than the whole of it.

What does silent departure leave behind, and where would it be found?

A silent departureLeaving without complaint, and therefore without a trace. writes nothing in the complaint file by definition. A silent departure still writes something elsewhere. Departure writes absences, and an absence is visible to anybody who decides in advance which presence was expected.

Think about a standing instruction. Somebody paying Rs 25,000/- a month into a scheme by standing instruction is generating a monthly event, and the month it stops generating one is a month with a shape. Nobody sent a message. Nobody filled in a form. But the sequence of twelve payments a year, worth Rs 3,00,000/- across the year, has a gap in it, and the gap is as legible as a letter if somebody was watching the sequence rather than the inbox.

The same logic applies to anything that used to happen and stopped. A quarterly review that stopped being booked. A message that stopped being opened. A renewal that lapsed without a conversation. A departure completed without anybody asking why. Exit is invisible in the complaint record and quite visible in the records of ordinary activity, so the practical response is not a better complaints process but a habit of watching the things that stop.

What does the Palash decision log record about any of this?

Almost nothing, and the shape of the nothing is worth being precise about. The Palash decision log covers 60 investors across eight quarters. Over that period 9 investors complained in writing and 14 left without saying anything. The log carries those two figures and no others, and they already show that the visible response was the smaller one: for every 9 who spoke, roughly 14 went quietly, a ratio of about 1.6 to 1.

Now turn to the absences. The log does not carry how many of the 60 were dissatisfied. Dissatisfaction is not an event, and nothing in the system fires when somebody feels it. The log does not carry why any of the 14 left. Nor does it carry whether any of the remaining 37 were unhappy and stayed. The log holds 9 rows of voice, 14 absences it can date but not explain, and no measurement whatever of the condition all of this was supposed to be about.

The 9 can be expressed against the two denominators used above. Against all 60 investors it is 9 divided by 60, or 15.0 per cent. Against the 23 who either spoke or left, it is 9 divided by 23, or 0.39130, rounding to 39.1 per cent. Both are computable only because the log happens to record departures as well as complaints. Most complaint records do not, and even with both the reason column is empty for all 14.

The shape is easier to feel on a single row than on sixty, so bring it down to one person. Meera Sundaram appears in that log repeatedly. Her holding opened on 4 January at Rs 12,00,000/-. She added Rs 1,00,000/- to Suvarna Chemicals on 19 February after a television segment named it. On 12 October she sold Suvarna Chemicals whole at Rs 4,60,000/- and kept Kesari Logistics, saying she would sell it when it returned to Rs 3,00,000/-. Every one of those is an event, so every one of them has a row.

Now ask the log a different question: was Meera satisfied with any of it? There is nothing to read. She did not complain, so there is no complaint row; she did not leave, so there is no departure to date. The log knows precisely what she did and holds not one character about how she felt about any of it. The same absence has arrived on a single named person. And note where a complaint would have started if there had been one: usually as a message that landed differently from how it was meant, which is set out under communication conduct.

The last row of that table is the one to carry away. The answer is not zero. Zero would be a measurement. The answer is that there is no field, so the quantity is not small, not uncertain and not estimated; it is simply not a thing the record has any way of holding.

What duty does the invisibility of exit create?

If the record cannot see two of the three responses, then anybody relying on the record is relying on a partial view, and knowing that a view is partial changes what may honestly be concluded from it. The whole of the duty follows from that, and it is a duty about inference rather than about process.

The duty runs in three directions. First, silence may not be read as satisfaction. Silence is what two of the three responses look like from the inside of a record. Second, a falling count may not be reported as an improvement without saying what happened to the channel over the same period. Better service and a worse channel produce identical numbers. Third, a group that appears rarely in the record is a group about whom less is known, not a group about whom good news has arrived.

How does somebody reading a complaint record actually use all this?

The same fact lands differently on each of three readers, so take them in turn. A person deciding alone, with no adviser and no committee, meets this when they read a published figure about complaints and try to work out what it says. The answer is that it says how many people complained, and that whether that is many or few depends on a denominator nobody published, so the figure supports a much narrower conclusion than its prominence suggests.

Somebody deciding on behalf of others meets it as a reading discipline. When a complaint total falls, the first question is what changed about the channel, and the second is what the ordinary activity records show over the same window. A total that fell while contributions were also stopping is a different fact from a total that fell while nothing else moved, and the complaint total on its own cannot distinguish them.

Somebody comparing two practices meets it as a warning against ranking. Two practices with complaint counts of 9 and 30 may differ in service, or may differ only in how visible their complaint address is, and the two explanations produce the same ordering. A complaint count is comparable across periods and across practices only to the extent that the channel was identical, and it almost never is. That is a limit on inference, not a suggestion about what anybody should measure instead.

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