Why it always takes longer

Lesson 1 found that your picture of a week is unreliable. This lesson is the same failure at the scale of a single job, where it has been measured much more precisely and where the fix that everybody reaches for turns out not to work.

The finding has a name, the planning fallacy, which Buehler and colleagues credit to Kahneman and Tversky in 1979 and put in quotation marks in their own title.² Buehler, Griffin and Ross, 1994, in the Journal of Personality and Social Psychology.¹ It's the best-evidenced thing in this course, and it is worth knowing its scope before its numbers: every subject in the studies below was a Canadian university student, and the paper was published in 1994.

The thesis study

Thirty-seven psychology students in the final semester of the Honors Thesis course at the University of Waterloo were telephoned and asked to predict when they would submit. Not once, but three times over, in three different ways:

• their best estimate of the date
• the date if "everything went as well as it possibly could"
• the date if "everything went as poorly as it possibly could"

The course coordinator recorded when each thesis actually arrived. The means below are from 33 of them.¹

Read the last column on its own, because it is where the lesson is.

These students named a date on the assumption that everything that could go wrong would, and the average of those dates was 48.6 days. The work took 55.5. Fewer than half of them finished by their own worst case. The paper states it plainly: "Interestingly, fewer than half of the respondents (48.7%) finished by the time they had predicted assuming that 'everything went as poorly as it possibly could.'"¹

So the worst case a person can imagine isn't the worst case. It is a slightly slower version of the run that goes well.

What this does to the obvious fix

You are probably already thinking it, and it is the first thing anybody thinks: fine, so I'll pad my estimates.

The same study tested the nearest thing to it. Reading the pessimistic date as a padded estimate produced under instruction is this course's own reading of that condition rather than a claim the authors make,⁴ and it is worth saying so, because a worst-case scenario and an arbitrary multiplier are not obviously the same operation. Here is what the authors found: "Although the instructions to make a pessimistic prediction decreased the optimistic bias in prediction, it did not increase the accuracy of respondents' forecasts."¹

The absolute error was 23.2 days under the pessimistic instruction against 22.6 days under the instruction to be accurate. Six tenths of a day apart, on a task that ran to fifty-five.

That distinction is the one to carry out of this section, and it recurs twice more in this course. Bias is which side of the truth you land on. Accuracy is how far from it you are. Padding moves you across the truth without moving you closer to it, because it adds no information about this particular job. You are still guessing; you have just changed the direction you guess in.

It isn't only theses

Study 2 took 104 undergraduates and asked for predictions on two projects each, one academic and one not, that they intended to finish in the coming week.¹

And the figure that closes off the easy explanation: these people were not hedging or being modest. "[S]ubjects reported feeling 74.1% certain that they would meet their forecasts for academic projects and 69.9% certain for nonacademic tasks."¹

Three quarters certain, and right about a third of the time.

The ratio, which is the part you can use

Something worth noticing across those studies, and it's this course's own arithmetic on the published means rather than a finding the authors report.⁴

Divide the actual by the predicted:

• Thesis study, best estimate: 55.5 over 33.9, about 1.6
• Study 2, academic projects: 10.7 over 5.8, about 1.8
• Study 2, nonacademic projects: 9.2 over 5.0, about 1.8

Different tasks, different lengths, ratios in the same neighbourhood. Now do the same with the differences: 21.6 days, 4.9 days, 4.2 days. Those have nothing in common and nothing to say to each other.

This is why the exercise at the end of this lesson asks for a ratio. A shortfall of forty minutes on a forty-minute job tells you nothing about a job that will take a fortnight. "About double" does.

Worked: five tasks and one number

Rosa's week is constructed, and so is every figure in it.⁴ She is a district nurse on a fixed rota, and these are five things she did around it.

Line the five ratios up in order: 0.89, 1.25, 1.38, 1.75, 4.7. The middle one is 1.38, and that is her median. She would have got 1.99 by averaging, and the average is the wrong tool here, because one task is doing all the work in it.

That one task is the finding, not the noise. The bank call ran nearly five times over, and it is the only one of the five where the length was not hers to decide. She could not have known it would be forty-seven minutes. What she could have known, from the last three times she rang them, is that it is the kind of task where somebody else sets the clock.

So her sheet gives her two things rather than one. A working number of about 1.4 for tasks she controls, which is close enough to the studies above to be worth trusting on herself. And a named category, anything that involves waiting on an institution, where the ratio is not a ratio at all and the only honest plan is to give the whole morning to it.

Notice what a median did that an average would not: it kept the outlier visible as a separate fact instead of smearing it across the other four.

Now run the same division on the Study 4 table further down this lesson, which is a computer assignment rather than a thesis: 6.8 over 5.5 is about 1.2. So the neighbourhood is loose, and it is loosest on the shortest task.

That is the first of two warnings, and the three figures above would have hidden it. Three ratios from two studies of students in one country do not give you a constant, and your own number is the one that matters anyway. The second warning is that a ratio is only stable within a kind of work, which the exercise is designed to show you.

The finding that inverts the moral

Inside Study 2 there is a smaller analysis that is, practically speaking, the most useful thing in the paper.¹

Sixty-two of those students had an external deadline on their academic project. A date somebody else set.

• 80.6 percent of them finished in time to meet it.
• The projects were due in 12.9 days on average, and were reported finished in 11.0.
• They had predicted 5.9 days.
• Only 38.7 percent finished within their own prediction.

Look at what that describes. These were not people who couldn't finish things. Four out of five delivered against the date they'd been given. Their problem was entirely in the forecast, which sat "well in advance of both the deadline and the reported completion time".¹

The authors put the conclusion carefully: "subjects underestimated the importance of deadlines in determining when they would finish their projects."¹ The correlations make it concrete. Their predictions correlated only weakly with their deadlines, r = .23. Their actual completion times correlated strongly with them, r = .82.

In plain terms: the deadline was running their schedule and their forecast didn't know. They were predicting from the work, and the work was not what governed when they finished.

You should take two things from this and they pull in different directions, which is why the lesson gives you both rather than choosing.

One. If you want to know when something will be done, the date somebody else set is better evidence than your own estimate. That is uncomfortable and it is what the correlations say.

Two. Only a certain kind of task has one. A deadline is not available for the things nobody is waiting for, and those are exactly the things that slide for years. Lesson 4 is about what to do there, and lesson 6 is about deciding whether they should be on the list at all.

Why it happens, in one set of figures

Study 4, the one in the next section, also asked its 123 subjects to list what they were thinking about while they predicted, and the answer is the whole mechanism.¹

• 93.5 percent reported considering future plans and scenarios for the task, mostly about how they'd successfully complete it.
• 9.8 percent mentioned any potential impediment.
• 8.9 percent reported thinking about their own past experiences.

Read those three together and the picture is not one of people ignoring their history. It is of people for whom the history never came up.

Forecasting a task means imagining how it will go, step after step, and an imagined run is made of steps that work. That is what makes it imaginable. The specific thing that will actually delay you, the document you can't find, the reply that takes four days, the morning you're too tired, has no place in the story because you don't know which one it will be. You know there will be one. You cannot picture it.

So you do not produce a bad forecast through carelessness. You produce it by doing the thing forecasting is, and your history isn't in the room while you do it.

And remembering doesn't fix it

If the history is absent, the obvious repair is to fetch it. Study 4, the source of the three figures above, tested exactly that on 123 students doing a computer assignment.¹

One group described their past experience with similar assignments just before predicting, and was told to keep those experiences in mind.

Almost nothing. And the authors' description of why is worth reading in full, because it names something most people will recognise:

"The absence of an effect in the recall condition is rather remarkable. In this condition, subjects first described their past performance with projects similar to the computer assignment and acknowledged that they typically finish only 1 day before deadlines. Following a suggestion to 'keep in mind previous experiences with assignments,' they then predicted when they would finish the computer assignment. Despite this seemingly powerful manipulation, subjects continued to make overly optimistic forecasts. Apparently, subjects were able to acknowledge their past experiences but disassociate those episodes from their present predictions."¹

They said out loud that they always finish the day before it's due. Then they predicted that this time they'd be early.

Remembering is not connecting, and the gap between those two is where this whole subject lives. There was a third group in that study who were made to do the connecting, and what happened to them is lesson 4's subject. Their share finishing on time roughly doubled, though their forecasts were no more accurate than anybody else's, which is a distinction this lesson has already made you careful about. Lesson 3 stops here on purpose: you should feel the size of the problem before you meet what can be done about it.

Four things people believe about their own estimates

"I just need to pad them." Measured, in the study this lesson opens with. Padding cut the bias and left the error where it was.¹ The pessimistic estimate was still a week short of the truth.

"I should remember how long it took last time." Measured, and it moved the on-time share from 29.3 to 38.1 percent while the students in question had just recited their own history accurately.¹

"Other people manage this." The paper's first hypothesis was that "[p]eople underestimate their own but not others' completion times",¹ and the pattern it reports is that the bias belongs to predicting for yourself rather than to predicting in general. This course has read the abstract and Studies 1, 2 and 4, not the observer study, so treat this as what the paper reports rather than as something the course has checked line by line. Either way, the folk version has it backwards. You are not unusually bad at this, and on what the paper reports the asymmetry is not between you and other people but between predicting for yourself and predicting for somebody else. Which is a reason to ask somebody, and lesson 4 gives you a way of doing the same job on your own.

"Deadlines are the problem." In Study 2, 80.6 percent of the students with deadlines met them.¹ Deadlines were the thing that worked. What failed was the forecast.

Practice

Connections

Back. Lesson 1 established that your impression of your own time is unreliable and asked you to measure rather than to introspect; this lesson is that argument at the scale of one task, with much better evidence behind it. Lesson 2's habit of writing the prediction down before the record is what makes the exercise above work at all. From earlier Core courses: Logic and Argument lesson 6 gave you the habit of asking how much a piece of evidence should move you, which is what the correlations of .23 and .82 in the deadline analysis are doing. Using AI Effectively taught you to ask what a measurement was taken on before you carry it anywhere, which is why the scope of these studies is in the second paragraph rather than in a footnote.

Forward. Lesson 4 is the two corrections with measured effects behind them, the third condition of the study this lesson stopped short of and the one planning technique in this subject with a meta-analysis, and it is careful about what each does and does not improve. Lesson 5 is why the day itself doesn't hold still. Lesson 7 builds a week using your own ratio. Lesson 8 returns to what the whole evidence base is worth.

Go deeper

• Exploring the "Planning Fallacy" (Journal of Personality and Social Psychology, 1994). Hosted openly at MIT. This course has read the abstract, Studies 1 and 2 in full with their tables, and Study 4 with its table and discussion. If you read one thing, read the tables: they carry the argument on their own, which is rare. Reading Well lesson 8 gives you the questions to put to a figure, and they work unchanged on these tables.
• The American Time Use Survey appears here for a third time and a third reason, after lessons 1 and 2. This one: it is a case of a large organisation deciding that asking people how long things take is not good enough, and building an instrument instead. That is the institutional version of the move this lesson asks you to make on yourself. United States only, and this course has read the article that uses its data rather than the survey's own documentation.

Sources

1. Roger Buehler, Dale Griffin and Michael Ross, "Exploring the 'Planning Fallacy': Why People Underestimate Their Task Completion Times", Journal of Personality and Social Psychology 67(3), 1994, pages 366 to 381. Read in substantial part: the abstract, Studies 1 and 2 in full with their tables, and Study 4 with its table and discussion. Supports: Study 1's design, its three predicted means and its actual mean, and the 48.7 percent figure with its quoted sentence; the quoted finding that the pessimistic instruction decreased the bias without increasing accuracy, and the 23.2 against 22.6 absolute errors; Study 2's means and completion shares and the quoted confidence figures; the deadline sub-analysis including 80.6 percent, 12.9 and 11.0 days, 5.9 predicted, 38.7 percent, the quoted sentence about underestimating the importance of deadlines, and the correlations of .23 and .82; the thought-listing figures of 93.5, 9.8 and 8.9 percent; Study 4's control and recall columns and the quoted passage about the absence of an effect in the recall condition; and the abstract's first hypothesis about own against others' completion times. Scope: university students in Canada, published 1994. Study 1 is 37 Honours Thesis students at the University of Waterloo, means from 33; Study 2 is 104 undergraduates; Study 4 is 123 students.
2. The term is credited by Buehler and colleagues to Daniel Kahneman and Amos Tversky, "Intuitive Prediction: Biases and Corrective Procedures", 1979. This course has not opened that paper and the attribution above is the only claim it makes about it.
3. Brigitte J. C. Claessens, Wendelien van Eerde, Christel G. Rutte and Robert A. Roe, "A review of the time management literature", Personnel Review 36(2), 2007. Read in part: the abstract and pages 255 to 257. Supports the standing point, made in lesson 1 and relied on here, that the evidence base for this subject is mostly students and office workers.
4. Three things in this lesson are the course's own and are labelled as such in the body where the reader meets them. First, the ratios of actual to predicted time are this course's arithmetic on the published means in source 1: the authors report means and completion shares, they do not report ratios, and nothing in the paper establishes that a ratio is stable for an individual. Second, reading Study 1's pessimistic condition as a test of padding an estimate is this course's reading of that condition and not a claim the authors make. Third, Rosa and her five tasks are constructed, along with every figure in them; no source read for this course records one person's predictions against their own times.