Health Economics Series Guide: Read the Evidence Before the Costs

อ่านฉบับภาษาไทย (Thai version)

Abstract

A cost-effectiveness result is only as sound as the clinical numbers beneath it, so this series reads the evidence before the costs. This guide follows one illustrative decision along the whole route. A district hospital committee must decide whether to pay for a new preventive treatment, C, in a district of 50,000 adults. Part I counts the patients with the right denominator, names the effect measure and asks how much C lowers their risk: in a trial, 8 of 66 patients on C and 13 of 60 on usual care had a serious complication, and the uncertainty is wide. Part II adds what C costs, how much health it buys, what each extra year in full health costs and what the district would pay each year. Part III checks where the inputs come from, the highest extra cost that still pays off and how to brief the committee. Every step links to the article that teaches it in full.

🔍 Click to enlarge Click anywhere to close

One decision, read from the top

A district hospital committee must decide whether to pay for a new preventive treatment, C, for adults newly diagnosed with a chronic condition. The district has 50,000 adults. In a trial, 8 of 66 patients given C and 13 of 60 given usual care, A, had a serious complication. C costs 150,000 THB more per patient than A.

The analysts' summary ends with one line. C buys the equivalent of 1.00 extra year in full health per patient, so each such year costs 150,000 THB. The committee's threshold, the most it will pay for one such year, is 160,000 THB. A quick reader stops there and says yes.

That line is the last link in a chain. It rests on how many patients there are, on how much C really lowers their risk and on how sure the trial is. A wrong patient count or a biased trial makes the money line wrong before any price is entered, and no cost table can catch it.

So this series reads in the other order: evidence first, then costs, then the decision. Its 15 topic articles sit in three parts, and this guide walks the committee's case through them once.

Part I: count the patients, then test the treatment

The committee's first question is how many patients C would reach. An incidence rate counts new cases over the time people were watched. Here it is 2 per 1,000 person-years: 1,000 adults followed for one year produce about 2 new patients. In 50,000 adults that is 100 new patients a year, and every later cost is multiplied by it.

A count of everyone now living with the condition would be larger, because it also holds people diagnosed in earlier years. Pricing C on that count would inflate the bill. The article on incidence and prevalence teaches the difference and how a rate becomes a risk, the input of a model that follows a group of patients through health states year by year (Incidence, Prevalence and the Right Denominator: The Epidemiology Beneath Cost-of-Illness and Markov Models).

Next, how much does C change the risk? In the trial, 8 of 66 patients on C had the complication, a risk of 12.12%, against 13 of 60 on A, a risk of 21.67%. The risk ratio divides the first risk by the second: 0.56, so a patient on C carries a little over half the risk of usual care.

The committee pays per patient, so the absolute gap matters more. The risk difference is 21.67% minus 12.12%: 9.55 fewer complications per 100 patients treated. The number needed to treat (NNT) turns that around: 100 divided by 9.55 is about 10.5, so 11 patients must receive C to prevent one complication over the trial's follow-up.

Then chance. The 95% confidence interval of the risk ratio runs from 0.25 to 1.26. It includes 1, the value for no effect, so this one trial cannot rule out that C does nothing.

Five effect measures, and the article that teaches each

The trial's answer came as a risk ratio, but a result can arrive as any of five effect measures. They compare different quantities: risks, odds, rates or hazards. Reading one as another changes the money line before any price is entered, so each measure has one article that teaches it.

Which of the five a study can give is set by three things together: its design, its statistical model and its estimand, the exact quantity it sets out to estimate. The name of the design alone does not decide it.

• Risk difference (RD): the risk on C minus the risk on A over the same follow-up, here 9.55 fewer complications per 100 treated. The NNT is its inverse, and the article on one trial table's four answers teaches both with their intervals (Risk Ratio, Odds Ratio, Risk Difference and NNT: One 2×2 Table, Four Answers).
• Risk ratio (RR): one risk divided by the other, here 0.56. The same article shows why a model applies it to a local baseline risk.
• Odds ratio (OR): it compares the odds of an event, the events divided by the non-events, between two groups. It drifts further from 1 than the risk ratio when events are common. With 8 of 66 on C and 13 of 60 on A having the complication, it is 0.50 while the risk ratio is 0.56. The same trial-table article teaches the odds ratio with its interval (Risk Ratio, Odds Ratio, Risk Difference and NNT: One 2×2 Table, Four Answers).
• Incidence rate ratio (IRR): one rate divided by another, where a rate counts events per person-time, like the 2 per 1,000 person-years above. This trial reports only risks, not person-time, so it has no IRR; with two rates, say 2 complications per 1,000 person-years on C and 4 on A, the IRR would be 2 divided by 4, or 0.50. The hazard article estimates it with a Poisson model of event counts over person-time, and also teaches the restricted mean survival time, the average event-free time within a fixed window (The Hazard Ratio Is Not a Risk Ratio: Proportional Hazards, Restricted Mean Survival Time and Model Inputs). The design article works one out from person-time and shows that a case-control odds ratio can estimate a risk ratio, a rate ratio or a risk odds ratio (the odds ratio of the risks over the whole follow-up), depending on how the controls were sampled (Design, Model and Estimand: What Each Study Design Can Actually Estimate).
• Hazard ratio (HR): it compares how fast events happen among people still event-free, moment by moment; it is not a ratio of risks over a period. The hazard article shows that it equals the IRR only when both hazards stay constant. It also shows why a model takes rates and survival curves, not one HR multiplied onto a probability.

Bias, certainty and other kinds of evidence

Then bias and certainty. Was treatment assigned at random, were complications counted the same way in both groups, and do the trial's patients look like the district's? One article separates chance, bias and confounding (Chance, Bias and Confounding: Why an Association Is Not Yet a Treatment Effect), and the next rates how much trust the effect deserves, its certainty, on a scale called GRADE (Beyond the Evidence Pyramid: GRADE Certainty and Choosing Effectiveness Inputs for Economic Models).

The last article of Part I covers treatments that start with a screening test (From Sensitivity to Cost per Case Detected: Predictive Values, Prevalence and the Economics of Screening).

Part II: cost, health, value and the yearly bill

Money starts with whose costs count and in which year's prices. The 150,000 THB must say whether it holds only hospital bills or also what families pay, and which price year it uses. The costing article sets these rules.

Next comes the health side, and the analysts' extra year in full health has a name. A QALY, a quality-adjusted life year, is a year of life weighted by its quality, from 1 for full health to 0 for death. As an illustration unrelated to C, over 10 years a state valued at 0.70 counts as 7.0 QALYs, so restoring full health would gain 3.0. Such a quality weight is called a utility, and the full explanation is in QALY, DALY and the Utilities Behind Them: What Is Actually Being Measured.

C's claim is 1.00 extra QALY per patient. The incremental cost-effectiveness ratio (ICER) is the extra cost divided by the extra health gained against the next best option. Option C costs 150,000 THB more than A for 1.00 extra QALY, an ICER of 150,000 THB per QALY against a threshold of 160,000; the full explanation is in the article on value thresholds.

The same comparison can be stated in money. Net monetary benefit (NMB) is health valued at the threshold minus cost. Applied to the extra cost and extra health of C over A, it is the incremental net monetary benefit (INMB), and a positive INMB means the extra cost is worth paying. At 160,000 THB per QALY, 1.00 extra QALY for 150,000 THB more gives an INMB of 10,000 THB, as the same article shows.

Where would the 1.00 QALY come from? A model turns the trial's 9.55 fewer complications per 100 into years of life and quality by following the patients over time. It also reruns the calculation many times with the uncertain inputs, so the interval from 0.25 to 1.26 reaches the money result instead of vanishing. The modelling article builds those models.

Value is not affordability. 100 new patients a year at 150,000 THB more each is 15,000,000 THB a year from the same budget, whatever the cost per QALY says. Who gains is a third question, and the article on budgets and fairness keeps the yearly bill and fairness apart from value (Cost-Effective Is Not Affordable, and Efficient Is Not Fair: Budget Impact and Equity).

Part III: sources, a price ceiling and the brief

The last three articles turn the numbers into a decision. The first checks whether each input, such as the 150,000 THB or the 1.00 QALY, comes from a source that fits the district (Parameter Sourcing, Reference Cost Lists and Utility Catalogues When Local Data Are Sparse).

The second asks the price question backwards (Early Health Technology Assessment, Headroom Pricing and Benefit-Package Decisions). At 160,000 THB per QALY, 1.00 extra QALY is worth 160,000 THB. The extra cost can therefore rise to 160,000 THB before C stops being worth it, and today it sits 10,000 THB under that ceiling.

The third writes the brief (Communicating Health Economic Evidence: From Model Outputs to Policy Briefs). For C it would put three facts side by side: the 10,000 THB margin per patient, a trial interval that still includes no effect, and a bill of 15,000,000 THB a year.

Trace the trial numbers

# Illustrative trial behind treatment C (base R). Writes: stdout only.
events_c <- 8; n_c <- 66; events_a <- 13; n_a <- 60
risk_c <- events_c / n_c; risk_a <- events_a / n_a; gap <- 100 * (risk_a - risk_c)
cat("Risk on C (%):", formatC(100 * risk_c, format = "f", digits = 2), "\n")
cat("Risk on A (%):", formatC(100 * risk_a, format = "f", digits = 2), "\n")
rr <- risk_c / risk_a; se_log_rr <- sqrt(1 / events_c - 1 / n_c + 1 / events_a - 1 / n_a)
cat("Risk ratio:", formatC(rr, format = "f", digits = 2), "\n")
ci <- formatC(exp(log(rr) + c(-1.96, 1.96) * se_log_rr), format = "f", digits = 2)
cat("95% CI, lower end:", ci[1], "\n")
cat("95% CI, upper end:", ci[2], "\n")
cat("Fewer complications per 100 treated:", formatC(gap, format = "f", digits = 2), "\n")
cat("100 / that gap:", formatC(100 / gap, format = "f", digits = 1), "\n")
cat("Number needed to treat:", ceiling(100 / gap), "\n")

Risk on C (%): 12.12
Risk on A (%): 21.67
Risk ratio: 0.56
95% CI, lower end: 0.25
95% CI, upper end: 1.26
Fewer complications per 100 treated: 9.55
100 / that gap: 10.5
Number needed to treat: 11

Trace the effect measures

# Two effect measures from the reading map (base R). Writes: stdout only.
events_c <- 8; n_c <- 66; events_a <- 13; n_a <- 60
odds_c <- events_c / (n_c - events_c); odds_a <- events_a / (n_a - events_a)
cat("Odds ratio in the trial:", formatC(odds_c / odds_a, format = "f", digits = 2), "\n")
rate_c <- 2; rate_a <- 4   # illustrative complications per 1,000 person-years
cat("IRR, illustrative rates:", formatC(rate_c / rate_a, format = "f", digits = 2), "\n")

Odds ratio in the trial: 0.50
IRR, illustrative rates: 0.50

Trace the value and the yearly bill

# Illustrative district, health, cost and value for treatment C (base R). Writes: stdout only.
adults <- 50000; new_per_1000 <- 2; patients <- adults * new_per_1000 / 1000
cat("New patients a year:", patients, "\n")
years <- 10; state_value <- 0.70; qalys <- years * state_value
cat("QALYs in 10 years at that value:", formatC(qalys, format = "f", digits = 1), "\n")
cat("QALYs gained by full health:", formatC(years - qalys, format = "f", digits = 1), "\n")
extra_cost <- 150000; extra_qaly <- 1; threshold <- 160000
thb <- function(x) formatC(x, format = "f", digits = 0, big.mark = ",")
cat("Extra QALY per patient:", formatC(extra_qaly, format = "f", digits = 2), "\n")
cat("ICER (THB per QALY):", thb(extra_cost / extra_qaly), "\n")
cat("Extra QALY valued at the threshold (THB):", thb(threshold * extra_qaly), "\n")
cat("INMB per patient (THB):", thb(threshold * extra_qaly - extra_cost), "\n")
cat("Extra spending a year (THB):", thb(patients * extra_cost), "\n")

New patients a year: 100
QALYs in 10 years at that value: 7.0
QALYs gained by full health: 3.0
Extra QALY per patient: 1.00
ICER (THB per QALY): 150,000
Extra QALY valued at the threshold (THB): 160,000
INMB per patient (THB): 10,000
Extra spending a year (THB): 15,000,000

Use the calculators to test assumptions

Three calculators let you rerun the committee's case with your own inputs. Open the cost and value calculator to compare costs, QALYs and price ceilings. Open the decision tree calculator when the complication is a single event with a clear end, and the state-transition calculator when it can recur year after year.

Keep the patient count, units, timing and comparator the same as in the evidence. Change one input at a time. Agreement with a calculator checks the arithmetic; it does not validate the evidence or approve coverage.

The series, in learning order

1. Health Economics Series Guide: Read the Evidence Before the Costs (this article)
2. Incidence, Prevalence and the Right Denominator: The Epidemiology Beneath Cost-of-Illness and Markov Models
3. Design, Model and Estimand: What Each Study Design Can Actually Estimate
4. Risk Ratio, Odds Ratio, Risk Difference and NNT: One 2×2 Table, Four Answers
5. The Hazard Ratio Is Not a Risk Ratio: Proportional Hazards, Restricted Mean Survival Time and Model Inputs
6. Chance, Bias and Confounding: Why an Association Is Not Yet a Treatment Effect
7. Beyond the Evidence Pyramid: GRADE Certainty and Choosing Effectiveness Inputs for Economic Models
8. From Sensitivity to Cost per Case Detected: Predictive Values, Prevalence and the Economics of Screening
9. Whose Cost, in Which Year? Perspective, Opportunity Cost, Inflation and Discounting
10. QALY, DALY and the Utilities Behind Them: What Is Actually Being Measured
11. ICER, Dominance and the Cost-Effectiveness Plane: Reading the Thai Threshold of 160,000 Baht per QALY
12. Decision Trees, Markov Cohorts and Probabilistic Sensitivity Analysis: How a Model Turns Evidence into a Decision
13. Cost-Effective Is Not Affordable, and Efficient Is Not Fair: Budget Impact and Equity
14. Parameter Sourcing, Reference Cost Lists and Utility Catalogues When Local Data Are Sparse
15. Early Health Technology Assessment, Headroom Pricing and Benefit-Package Decisions
16. Communicating Health Economic Evidence: From Model Outputs to Policy Briefs

Related in the wiki: Incidence, Prevalence and the Right Denominator: The Epidemiology Beneath Cost-of-Illness and Markov Models · Design, Model and Estimand: What Each Study Design Can Actually Estimate · Risk Ratio, Odds Ratio, Risk Difference and NNT: One 2×2 Table, Four Answers · The Hazard Ratio Is Not a Risk Ratio: Proportional Hazards, Restricted Mean Survival Time and Model Inputs