Health as a Return on Capital

Volume V — Labour, Value, Flourishing

THE PLATE

A woman standing at a wooden desk beside a window, looking out, afternoon light on her and the dried grasses. — Plate V.09The Standing Desk at Four O'Clock.The claim run is the only document in the building that cannot flatter anybody. It does not know who joined the programme, and that is exactly what makes it worth reading.

THE LETTER

Somebody has put a number in front of you. It says that for every dollar your organisation spends on employee health it gets three and a quarter back, and it is on a slide with a photograph of a woman on a yoga mat.

The number has a real source. It is not made up, it is not a vendor's invention, and it comes from a careful meta-analysis by serious economists published in a serious journal. You should still not sign against it, and the reason is the whole of this chapter: between that paper and today, the field did the thing that fields are supposed to do and mostly do not. It ran large randomised trials. It found something close to nothing on cost. And one of the authors of the original meta-analysis is a co-author of the trial that found it.

That is not an embarrassment. It is the single most creditable episode in the recent economics of health, and it is why this chapter can be written with confidence rather than with hope. A field that publishes its own reversal has given you something worth more than the original figure: a set of numbers you can actually act on.

So this chapter does four things, in order. It states exactly what the randomised evidence says and does not say. It takes the naive return on a real programme apart, arithmetically, and shows you where three to five times too much came from. It states the QALY and the DALY precisely, because they are the instruments you will need once you stop using the broken one. And then — this is the half that matters on Monday — it sets out what does have a defensible return, because several things do, and three of them are sitting unfunded in your building right now while the yoga mat is paid for.

You will finish able to read any health ROI put in front of you and say, within ten minutes, which part of it is a measurement and which part is a selection effect. That is a small skill and it is worth a great deal of money.

— The Editors

DISCOVERY

What is already working

Begin where this field is at its best, because its best is much better than its reputation, and almost nobody has been shown the good version.

A trial that paid people to stop smoking, and worked. Kevin Volpp and colleagues randomised 878 General Electric employees to information alone or to information plus a structured financial incentive of up to 750.00 USD. Cessation at nine to twelve months was 14.7 per cent in the incentive arm against 5.0 per cent in the control arm — an absolute difference of 9.7 points, a relative effect of 2.94 x, and a number needed to treat of 10.31. That is a large, clean, replicated behavioural effect obtained by a workplace, and it is the opposite of a null.

A safety programme nobody chose to be in. David Levine, Michael Toffel and Matthew Johnson used California's randomised allocation of workplace safety inspections — a lottery, in effect, run by the state — and compared inspected establishments with uninspected ones. Inspected firms had 9.4 per cent fewer injuries and 26.0 per cent lower workers' compensation costs over the following four years, an average saving of 355,000 USD per firm, or 88,750 USD a year. And the effect everyone expects and nobody finds: there was no detectable difference in employment, sales, credit rating or survival. The safety did not cost the jobs. That result is randomised, large, and about money.

A programme that told the truth about its own two halves. RAND's seven-year evaluation of PepsiCo's Healthy Living programme reported an overall return of 1.46 : 1 — and then, honourably, reported it split. The disease-management component, aimed at people who already had a diagnosed condition, returned 3.78 : 1 and saved 136.00 USD per member per month, or 1,632.00 USD a year. The lifestyle-management component, aimed at healthy people's habits, returned 0.48 : 1. The two differ by 7.88 x and they point in opposite directions. The average of them is the number that got quoted and it is the least useful figure in the study.

A workplace trial that moved what it said it would move. The Work, Family and Health Network's STAR intervention was a genuine cluster-randomised trial of a change to how work is organised — schedule control and trained supervisor support, not a wellness benefit. It improved schedule control, lowered burnout, perceived stress and psychological distress, and lowered voluntary turnover. It did not demonstrate medical-cost savings and never claimed to. It moved the thing it was built to move.

And a lottery that settled an argument. Oregon expanded Medicaid by drawing names from a list, which produced the cleanest natural experiment in the history of health insurance. The result is the shape you should carry into every conversation in this chapter: a 30.0 per cent relative reduction in depression, catastrophic out-of-pocket spending effectively eliminated — and no significant change in blood pressure, cholesterol or glycated haemoglobin at two years. Large effects on suffering and on financial ruin. No movement in the biomarkers.

Five cases. Notice what they have in common, because it is the pattern this whole chapter runs on. In every one, the outcome was measured by somebody who could not choose who was in the programme. A lottery, a randomisation, a state inspection schedule, an administrative claim run. Where that is true, the field produces real and usable numbers. Where it is not true, it produces the slide with the yoga mat.

THE ARITHMETIC

Where three to five times too much comes from

First, what the randomised record actually says.

The Illinois Workplace Wellness Study randomised 12,459 university employees: 4,834 to treatment and 7,625 to control, which is 38.8 per cent and 61.2 per cent of the sample. After one year it found no significant effect on medical spending, health behaviours, employee productivity or self-reported health. At two years, with clinical measurement added, no significant effect on markers, spending or absenteeism. The one durable effect was on beliefs, and on whether an employee had a primary care physician — which is a real and good result, and is not the one that was sold. The confidence intervals are the part to carry: they are tight enough to exclude 78.0 per cent of the previously published estimates of savings.

The BJ's Wholesale Club trial cluster-randomised 160 worksites — 20 treatment, 140 control — covering 32,974 employees, 4,037 at treatment sites and 28,937 at control sites, about 201.8 and 206.7 people per site. At eighteen months and again at three years it found significantly higher self-reported rates of regular exercise and active weight management, and no significant difference in clinical measures, spending, utilisation, absenteeism, tenure or job performance.

Second, what the published ratio requires. Take the 3.27 : 1 on medical cost and the 2.73 : 1 on absence — 6.00 USD of claimed return per dollar. Apply the medical half to a programme costing 150.00 USD per employee a year and it asserts a saving of 490.50 USD per employee. Set that against what the employer actually spends: on the 2023 benefits survey, a single-coverage premium of 8,435 USD less a worker contribution of 1,401 USD leaves the employer paying 7,034 USD. So the ratio asserts that a 150-dollar programme removes 6.97 per cent of the employer's entire medical cost. State it that way once and the randomised confidence intervals are not a surprise.

Third, the decomposition. Here is where the money goes.

Take a self-insured employer of 5,000 people spending 150.00 USD a head — 750,000 USD a year — with participation at 56.0 per cent, so 2,800 participants and 2,200 others. The evaluation the vendor performs compares the two groups afterwards:

  AFTER    participants        5,100 USD    non-participants   6,600 USD
  the gap the study reports    1,500 USD
  naive saving = 1,500 x 2,800             =  4,200,000 USD
  naive return = 4,200,000 / 750,000       =       5.60 : 1

Now ask the one question the vendor's method cannot ask: what was the gap before the programme existed?

  BEFORE   participants        4,900 USD    non-participants   6,100 USD
  the gap that already existed 1,200 USD
  causal effect = 1,500 - 1,200            =        300 USD
  saving = 300 x 2,800                     =    840,000 USD
  return                                   =       1.12 : 1

The naive figure is 5.00 x the difference-in-differences figure. Not because anyone lied. Because healthier people join wellness programmes, and 1,200 USD of the 1,500 USD gap was there before anyone did anything.

And the general form, so the bracket is arithmetic rather than opinion:

  naive / true  =  1 + (baseline gap / causal effect)

    baseline gap = 2 x the effect   ->  naive overstates by 3 x
    baseline gap = 3 x the effect   ->  naive overstates by 4 x
    baseline gap = 4 x the effect   ->  naive overstates by 5 x

Three to five times is simply what a pre-existing gap two to four times the size of the effect produces. Finally: at the randomised point estimate the effect is 0.00 USD, the return is 0.00 : 1, and the year's net is −750,000 USD.

Fourth, a second engine, and it runs even on people who did nothing. Suppose the programme targets the highest-spending decile. With a population mean of 6,000 USD, a standard deviation of 12,000 USD and a year-to-year correlation of 0.35, the expected value of a standard normal above the ninetieth percentile is 1.7550, so the targeted group's mean this year is 27,060 USD. Next year, with no intervention whatever, its expected mean is 13,371 USD. An apparent saving of 13,689 USD a head — 50.6 per cent — produced by arithmetic alone. The normal approximation understates it, because real spending has a heavier tail.

The honest negative: two of them, and the second one is ours

The first is the one above and it should be said plainly. Most published returns on workplace health programmes are selection artefacts. The randomised evidence is close to null on cost. An edition that printed the advocacy figure would be repeating a known error, and a manager who signs against it is buying a number that two large trials have excluded.

The second belongs to this book's own argument and it is harder. The living-systems case for health at work says: improve the conditions — control, security, recovery, belonging — and the body follows. The causal evidence for the conditions is genuinely strong. The evidence that the biomarkers follow, inside the horizon a business plans over, is not. Oregon moved depression by 30.0 per cent and moved blood pressure, cholesterol and glycated haemoglobin by nothing measurable at two years. STAR moved burnout and turnover and did not demonstrate medical savings. So a board paper promising that a regenerative workplace will lower the claim run within three years is promising something the best available trials did not find, and it should not be written.

And there is a measurement weakness underneath both, which is where much of the missing money was manufactured. Presenteeism — the productivity lost by people who are at work and unwell — is usually self-reported, then converted to dollars by multiplying the wage. Take a salary of 60,000 USD, a reported impairment of 5.0 per cent of working time, and a claimed relative reduction of 10.0 per cent:

  multiplier 0.5   impairment  1,500.00 USD   saving  150.00 USD   1.00 : 1
  multiplier 1.0   impairment  3,000.00 USD   saving  300.00 USD   2.00 : 1
  multiplier 1.5   impairment  4,500.00 USD   saving  450.00 USD   3.00 : 1

The same programme returns 1.00 : 1 or 3.00 : 1 — a spread of 3.00 x — depending on a multiplier nobody defends in the appendix. Mark Pauly and colleagues showed that the ratio of lost output to the wage is a property of the job: high where work is timed and team-dependent, near or below one where it is not. It is not 1.00 by nature. And the instruments that produce the 5.0 per cent in the first place agree with one another only moderately. Absence, by contrast, is read from payroll: a smaller number, a harder one, and the correct trade.

The cut

Here is the thing that reorders everything above, and it is one piece of arithmetic.

Health is not an asset the employer owns. It is an asset the employer rents, and you can compute the lease. Median employee tenure is 3.90 years. Model separations as exponential and the hazard is ln 2 / 3.90 = 0.17773 a year. A benefit stream worth B a year in perpetuity is worth B / r to society and only B / (r + λ) to the firm paying for it, so the firm's share is r / (r + λ):

  0.03 / (0.03 + 0.17773)  =  0.03 / 0.20773  =  14.44 %
  the employer's hurdle multiple  =  1 / 0.1444  =  6.92 x

An employer captures about a seventh of the long-run value of the health it creates, and therefore needs any long-duration health investment to be roughly seven times better than society needs it to be before its own arithmetic says yes. That single ratio explains the whole shape of the literature. The things employers do fund and do measure returns on — safety, disease management, smoking, turnover — all pay back inside the lease. The things they were sold instead were cheap, self-reported and long-dated, which is what a 14.44 per cent capture rate buys when nobody has noticed the capture rate.

Read that way, the null results stop being a disappointment about human health and become a finding about ownership. The wrong party was holding the asset. Change the owner and the arithmetic changes: at twelve years of tenure in a trade rather than a firm, the hazard is 0.05776, capture rises to 34.18 per cent, the hurdle falls to 2.93 x, and the same investment is 2.37 x more valuable to the party making it. That is not a sentiment. It is a design instruction, and it is the instrument at the end of this chapter.

DREAM

What becomes ordinary

In the organisation that has understood this, nobody argues about whether health pays. They argue about which endpoint is powered, which is a much better argument and one that evidence can settle.

Every health proposal arrives with three lines on its front page: the endpoint, the sample size required to detect the claimed effect, and the control group. The control group exists. It was made by randomising the order of a rollout that was going to be staggered anyway, at no cost, because the organisation worked out some years ago that a staggered rollout is a free experiment and an unstaggered one is a free anecdote.

The finance function knows the capture ratio and says it out loud. When a proposal's benefits run for twenty years, the paper states that the firm expects to hold 14.44 per cent of them and asks the obvious next question — who holds the rest, and would they pay us to do it? Sometimes the answer is the insurer. Sometimes the pension scheme. Sometimes the state. Occasionally it is the employee, who would happily pay a small amount for something portable. None of those conversations were possible while everyone was pretending the firm owned the whole asset.

Presenteeism appears in the pack with its multiplier printed beside it and a sensitivity band around it, or it does not appear at all. Absence is read from payroll. Claims are read from the claim run. The three are never added into one figure with a single confident total, because the organisation learned that the total's size is decided by the softest of the three.

And the things that work are funded first and generously: the safety programme, the cessation incentive that pays real money, the case management for the forty people with a serious diagnosis, the supervisor training that lets a parent move a shift without asking twice. They are funded because their arithmetic survived contact with a control group, and everyone in the room can recite it.

Nobody has stopped caring about the whole person. What has stopped is caring about the whole person through an instrument that could not detect them. The yoga is still there. It is paid for out of the same budget as the coffee, for the same reason, and nobody is asked to believe it returned 3.27 : 1.

DESIGN

The structure that gets you there

First: separate the four endpoints, and rank them by whether you can measure them.

Endpoint	Where it is read	Can a firm of 5,000 power it?
Medical spending	The claim run	No — and the arithmetic is below
Workers' compensation cost	The carrier's loss run	Sometimes, over several years
Voluntary turnover	Payroll	Yes, comfortably
Self-reported wellbeing	A survey	Yes — but it is not money

Now the arithmetic that decides the table. To detect a 300 USD change in annual spending with a standard deviation of 12,000 USD, at eighty per cent power and a five per cent two-sided test, you need 25,088 per arm — 50,176 employees, or 10.04 x more than the worked employer has. To detect a four-point fall in a twenty per cent voluntary turnover rate you need 1,568 per arm, 3,136 in total, which the same employer clears 1.59 x over.

That is the operational conclusion of this whole chapter. You cannot measure the medical-spending claim, so do not buy it. You can measure the turnover claim, so buy that one and hold the vendor to it.

Second: randomise the rollout, because it is free.

Almost every workplace programme is deployed in waves — by site, by division, by region — for reasons of capacity. Assign the wave order at random and you have converted a scheduling constraint into a randomised controlled trial at a marginal cost of zero. This is the single highest-return administrative decision in the chapter, and the entire literature that this chapter had to correct would not exist if it had been standard practice.

Third: move the endpoint to where the effect is, not where the story is.

Fund on the basis of what a trial has moved. Smoking cessation with real money: funded, on Volpp. Disease management for people with a diagnosis: funded, on the 3.78 : 1 half of the PepsiCo evaluation and not on the 1.46 : 1 average. Safety: funded, on a randomised inspection study showing 26.0 per cent lower compensation cost and 9.4 per cent fewer injuries with no job loss. Schedule control and supervisor training: funded, on burnout and turnover, and not on a promise about cholesterol.

Fourth: price the health effect in the right unit, and state the unit.

A quality-adjusted life year is time multiplied by a utility weight where 1.00 is full health and 0.00 is dead, discounted like any other stream. A disability-adjusted life year is years of life lost plus years lived with disability, where the disability weight runs the other way — 0.00 is full health. QALYs are gained; DALYs are averted; and one minus a utility is not a disability weight, because the two are elicited by different methods on different populations, however routinely they are swapped.

Worked: move 100 people from a utility of 0.78 to 0.85 — a gain of 0.07 — and hold it five years. The annuity factor at three per cent is 4.5797, so the programme gains 32.06 QALYs. At a cost of 600,000 USD that is 18,716 USD per QALY: against the published threshold range of 20,000 to 30,000 GBP, which is 25,400 to 38,100 USD, it is 1.36 x cheaper than the cheap end. Against the measured displacement figure of 12,936 GBP — 16,429 USD — it is marginal.

And now hold the cut beside it. That programme is a bargain in QALYs and the employer receives 14.44 per cent of its cash. Both sentences are true, they are in different currencies, and the reason health looks unfundable inside firms is that nobody says the second one.

DESTINY

How it holds when you stop pushing

Three things make this survive a change of Chief Financial Officer, and only three.

The control group is in the standing process, not in the project. If the randomised rollout is a rule about how rollouts are scheduled, it survives everyone. If it is something a clever analyst did once, it dies with them.

The capture ratio is written into the capital paper template. One line: expected duration of benefit, share retained by this entity, and who holds the rest. It takes a template change and it permanently stops the organisation from paying for other people's assets by accident.

The endpoints are owned by the functions that already report them. Turnover belongs to the people function, claims to finance, injuries to operations. A health metric that lives in a wellness team's own report dies when the team is reorganised. One that lives in payroll is read every month for other reasons.

Now the failure modes, named honestly, because they are specific.

It fails when the pilot is sized for enthusiasm and reports an effect inside the noise band — which, at 5,000 employees and a spending endpoint, is guaranteed. It fails when the baseline is built after the intervention, because then the 1,200 USD that was always there gets counted as the 1,500 USD that was earned. It fails when presenteeism is added to absence and claims to make one total, and the total is then defended as though the softest of the three were as solid as the hardest. It fails when a genuine wellbeing effect is oversold as a cost effect, and the cost effect does not arrive, and the wellbeing programme is cut for having failed at something it never claimed. That last one is the most expensive failure in this field, and it is caused entirely by the business case, never by the programme.

DELIGHT

What it feels like

There is a particular pleasure in being the person in the room who asks what the gap was before. It is not a clever question and it does not take courage. It takes about nine seconds, and the room goes quiet in a specific way, and what follows is usually a better conversation than the one that was scheduled.

Then a second pleasure, slower and better. Once you have stopped defending a number that cannot be defended, you find you can be far more generous about the things that are simply good. The supervisor training. The shift somebody needed to move. The extra week for the man whose wife is ill. None of it has to be justified by a fraudulent ratio any more, and it turns out that a thing you can do plainly, because it is right and it costs little, is easier to keep than a thing you have to keep proving.

And the claim run is genuinely interesting reading. It is the one document in the building that does not know who anybody is, does not know who joined the programme, and cannot be persuaded. Spending an afternoon with a document that cannot flatter you is a rest, not a chore.

OPERATIONALIZE THIS

At the level of finance

The cut said the wrong party is holding the asset. Here is the instrument that moves it, in a form a treasurer will recognise — because the legal vehicle already exists and has since 1947.

The structure: a multi-employer portable health-capital pool, funded by a shared-savings note, paid against a randomised control.

A group of employers in one trade or one region establish a jointly governed health trust — in the United States, a multiemployer welfare fund of the kind Taft-Hartley already provides for; in other jurisdictions, a sectoral fund or a mutual. Health investment is made by the pool, follows the worker across member employers, and is repaid out of verified claim reduction across the pool.

The mechanics, worked.

Scale. 14 employers, 41,000 covered lives, investing 210.00 USD a life a year: an annual programme of 8,610,000 USD.
Why the pool and not the firm. At 3.90 years of median firm tenure the capture is 14.44 per cent. At 12.00 years of tenure in the trade it is 34.18 per cent — 19.74 points recovered, a multiple of 2.37 x — and the hurdle falls from 6.92 x to 2.93 x. Nothing about the medicine changed. The owner changed.
Verification. Member sites are randomised into waves. Payment is computed against the sites not yet rolled out, never against non-participants. This is the clause the whole instrument exists to carry.
The return. At a verified reduction of 320.00 USD a life, gross saving is 13,120,000 USD on an 8,610,000 USD investment: 1.52 : 1, a rate of 52.4 per cent, which is 43.4 points of headroom over a 9.0 per cent cost of capital.
The waterfall. 75.0 per cent of verified saving — 9,840,000 USD — to the note until it is repaid; 3,280,000 USD retained by the pool from day one, so member employers see cash in year one rather than a promise.
Security. The verified saving stream and nothing else. If the programme does not reduce claims against its randomised control, there is nothing to repay, and the note must say so in those words.

The balance-sheet treatment. You may not capitalise a workforce: IAS 38 and its US equivalents forbid recognising an internally generated intangible of that kind, and nothing here asks anyone to misstate a statement. So the employer's contribution runs through operating expense, and the asset sits in the trust, where it belongs, as a funded position against a future claim stream. That is the accounting answer to the capture problem as well as the economic one. Talk to your auditors early; this is a conversation about a welfare trust, which they have every year.

The counterparty. Start with your own stop-loss carrier, who already holds the tail of your claim distribution and is therefore the one party with an aligned economic interest in your population's health over a horizon longer than 3.90 years. A stop-loss carrier will discuss a shared-savings rider before a bank will discuss a note.

The number that decides it.

        verified claim-dollars avoided per year
    -----------------------------------------------   >   WACC / capture
        pool investment + verification cost

Note the denominator on the right. The ordinary hurdle rate is the wrong test for a health asset, because it silently assumes the buyer keeps all of the benefit. Divide it by the capture ratio and you have the honest test: at a 5.00-year tenure the capture is 17.79 per cent, the hurdle multiple is 5.62 x, and a 9.0 per cent corporate hurdle becomes a 50.59 per cent social one. That is the number that tells you to move the asset rather than abandon it.

The first ninety days on a page.

Day	Action	Artifact
1–15	Pull three years of claims, loss runs and payroll turnover	Three baselines, dated
16–30	Compute your own capture ratio from your own tenure data	One line, one number
31–45	Randomise the wave order of whatever is next being rolled out	The allocation, sealed
46–60	Power every proposed endpoint; strike the ones you cannot detect	The powered endpoint list
61–75	Approach the stop-loss carrier with the pool structure	Term sheet, draft
76–90	First verified result against the randomised control	One page, one person

APPRECIATIVE QUESTIONS

Twelve, for a room

Discovery — what is already working

When did somebody here get better care, or a better schedule, or a safer station, and stay as a result? Who made that happen, and what did it cost?
Which of our health or safety spends already has a number attached that came from a claim run rather than a survey — and who built it?
Where have we changed how work is organised and watched something improve that we were not even measuring at the time?

Dream — what becomes possible

If every health proposal arrived with its endpoint, its required sample size and its control group on the front page, what would we stop arguing about?
Imagine we knew exactly what share of the health we create we actually keep. What would we do differently in the next budget round?
If our people carried their health investment with them when they left, what would we be willing to fund that we currently will not?

Design — what we build

What is being rolled out in waves in the next six months, and what would it cost us to assign the wave order at random?
Which endpoint can we genuinely detect at our size — and are we currently buying a claim about one we cannot?
What is our presenteeism multiplier, who chose it, and what happens to the business case at half and at one and a half times that figure?

Destiny — how it holds

What would have to be in the capital paper template for the capture question to be asked automatically, by everyone, forever?
Who outside this company holds the rest of the health value we create, and which of them would pay us something for it?
If a good wellbeing programme were cut next year for failing at a cost target it never claimed, what would we wish we had written down today?

WORKS CITED

Baicker, K., Cutler, D. and Song, Z. (2010). "Workplace Wellness Programs Can Generate Savings." Health Affairs, 29(2), 304–311.

Jones, D., Molitor, D. and Reif, J. (2019). "What Do Workplace Wellness Programs Do? Evidence from the Illinois Workplace Wellness Study." Quarterly Journal of Economics, 134(4), 1747–1791.

Reif, J., Chelius, C., Jones, D., Molitor, D., Rahman, T. and others (2020). "Effects of a Workplace Wellness Program on Employee Health, Health Beliefs, and Medical Use: A Randomized Clinical Trial." JAMA Internal Medicine, 180(7), 952–960.

Song, Z. and Baicker, K. (2019). "Effect of a Workplace Wellness Program on Employee Health and Economic Outcomes: A Randomized Clinical Trial." JAMA, 321(15), 1491–1501.

Song, Z. and Baicker, K. (2021). "Health and Economic Outcomes Up to Three Years After a Workplace Wellness Program: A Randomized Controlled Trial." Health Affairs, 40(6), 951–960.

Caloyeras, J. P., Liu, H., Exum, E., Broderick, M. and Mattke, S. (2014). "Managing Manifest Diseases, But Not Health Risks, Saved PepsiCo Money Over Seven Years." Health Affairs, 33(1), 124–131.

Mattke, S., Liu, H., Caloyeras, J. P. and others (2013). Workplace Wellness Programs Study: Final Report. RAND Corporation.

Volpp, K. G., Troxel, A. B., Pauly, M. V. and others (2009). "A Randomized, Controlled Trial of Financial Incentives for Smoking Cessation." New England Journal of Medicine, 360(7), 699–709.

Halpern, S. D., French, B., Small, D. S. and others (2015). "Randomized Trial of Four Financial-Incentive Programs for Smoking Cessation." New England Journal of Medicine, 372, 2108–2117.

Berman, M., Crane, R., Seiber, E. and Munur, M. (2014). "Estimating the Cost of a Smoking Employee." Tobacco Control, 23(5), 428–433.

Levine, D. I., Toffel, M. W. and Johnson, M. S. (2012). "Randomized Government Safety Inspections Reduce Worker Injuries with No Detectable Job Loss." Science, 336(6083), 907–911.

Kelly, E. L., Moen, P., Oakes, J. M. and others (2014). "Changing Work and Work-Family Conflict: Evidence from the Work, Family, and Health Network." American Sociological Review, 79(3), 485–516.

Moen, P., Kelly, E. L., Fan, W. and others (2016). "Does a Flexibility/Support Organizational Initiative Improve High-Tech Employees' Well-Being? Evidence from the Work, Family, and Health Network." American Sociological Review, 81(1), 134–164.

Horwitz, J. R., Kelly, B. D. and DiNardo, J. E. (2013). "Wellness Incentives in the Workplace: Cost Savings Through Cost Shifting to Unhealthy Workers." Health Affairs, 32(3), 468–476.

Pauly, M. V., Nicholson, S., Xu, J. and others (2002). "A General Model of the Impact of Absenteeism on Employers and Employees." Health Economics, 11(3), 221–231.

Pauly, M. V., Nicholson, S., Polsky, D. and others (2008). "Valuing Reductions in On-the-Job Illness: 'Presenteeism' from Managerial and Economic Perspectives." Health Economics, 17(4), 469–485.

Ospina, M. B., Dennett, L., Waye, A., Jacobs, P. and Thompson, A. H. (2015). "A Systematic Review of Measurement Properties of Instruments Assessing Presenteeism." American Journal of Managed Care, 21(2), e171–e185.

Kessler, R. C., Barber, C., Beck, A. and others (2003). "The World Health Organization Health and Work Performance Questionnaire (HPQ)." Journal of Occupational and Environmental Medicine, 45(2), 156–174.

Lerner, D., Amick, B. C., Rogers, W. H. and others (2001). "The Work Limitations Questionnaire." Medical Care, 39(1), 72–85.

Baicker, K., Taubman, S. L., Allen, H. L. and others (2013). "The Oregon Experiment — Effects of Medicaid on Clinical Outcomes." New England Journal of Medicine, 368, 1713–1722.

Finkelstein, A., Taubman, S., Wright, B. and others (2012). "The Oregon Health Insurance Experiment: Evidence from the First Year." Quarterly Journal of Economics, 127(3), 1057–1106.

Chetty, R., Hendren, N. and Katz, L. F. (2016). "The Effects of Exposure to Better Neighborhoods on Children: New Evidence from the Moving to Opportunity Experiment." American Economic Review, 106(4), 855–902.

Ludwig, J., Sanbonmatsu, L., Gennetian, L. and others (2011). "Neighborhoods, Obesity, and Diabetes — A Randomized Social Experiment." New England Journal of Medicine, 365, 1509–1519.

Ludwig, J., Duncan, G. J., Gennetian, L. A. and others (2012). "Neighborhood Effects on the Long-Term Well-Being of Low-Income Adults." Science, 337, 1505–1510.

Sullivan, D. and von Wachter, T. (2009). "Job Displacement and Mortality: An Analysis Using Administrative Data." Quarterly Journal of Economics, 124(3), 1265–1306.

Chetty, R., Stepner, M., Abraham, S. and others (2016). "The Association Between Income and Life Expectancy in the United States, 2001–2014." JAMA, 315(16), 1750–1766.

Marmot, M. G., Stansfeld, S., Patel, C. and others (1991). "Health Inequalities Among British Civil Servants: The Whitehall II Study." The Lancet, 337, 1387–1393.

Bosma, H., Marmot, M. G., Hemingway, H. and others (1997). "Low Job Control and Risk of Coronary Heart Disease in Whitehall II." BMJ, 314, 558–565.

Karasek, R. A. (1979). "Job Demands, Job Decision Latitude, and Mental Strain." Administrative Science Quarterly, 24(2), 285–308.

Salomon, J. A., Haagsma, J. A., Davis, A. and others (2015). "Disability Weights for the Global Burden of Disease 2013 Study." The Lancet Global Health, 3(11), e712–e723.

Neumann, P. J., Sanders, G. D., Russell, L. B., Siegel, J. E. and Ganiats, T. G. (eds) (2016). Cost-Effectiveness in Health and Medicine, 2nd edn. Oxford University Press.

Claxton, K., Martin, S., Soares, M. and others (2015). "Methods for the Estimation of the National Institute for Health and Care Excellence Cost-Effectiveness Threshold." Health Technology Assessment, 19(14).

Boushey, H. and Glynn, S. J. (2012). There Are Significant Business Costs to Replacing Employees. Center for American Progress.

Kaiser Family Foundation (2023). Employer Health Benefits Annual Survey.

Bureau of Labor Statistics (2024). Employee Tenure Summary, January 2024.

Note on figures. Every figure above is computed in lib/verify/V_09.py and printed with its inputs, its units and its source. Figures describing the worked employer, the worked pool and the worked QALY case are stated parameters and are labelled ILLUSTRATIVE in that module; every other figure carries a citation. The value of a statistical life, the NICE threshold range and the Claxton displacement figure are established in Chapter V.01 and are used here without being re-derived.