The updated meta database covers more than 50 years of VSL studies and more than 4 000 VSL estimates. This chapter provides an overview of how studies were selected for the meta-analysis and how VSL estimates from these studies were adjusted for comparability. The chapter reviews stated preference (SP) and revealed preference (RP) studies published between 1973 and 2009 and included in the previous OECD’s 2012 study on mortality risk valuation from 2012. It then describes the new studies included in the meta-analysis of this report, i.e. studies published between 2009 and 2023, as well as additional RP studies published before 2009. The chapter provides descriptive statistics of these studies, such as study count and their methodological and geographical distribution.
Mortality Risk Valuation in Policy Assessment
Abstract
This chapter describes the meta-data used for the analysis, which is derived from three sources (Figure 3.1). The analysis makes use of data from the 2012 meta-analysis of value of statistical life (VSL) estimates (OECD, 2012[1]), as well as new meta-data that significantly expands the empirical basis for this report.
Note: * This RP dataset is the underlying data used in Viscusi and Masterman (2018[2]) and was made available to the authors by Kip Viscusi and Clayton Masterman. Although the dataset covers publications through 2016, the RP dataset used here only includes studies from this dataset published before 2009 since the 2009-2023 data collected covers both RP and SP using the systematic approach summarized in Figure 3.2.
3.1.1. OECD stated preference database (1970 –2008)
The meta-data source referred to in Figure 3.1 as the “OECD 2008 database” is the original database compiled for the OECD (2012[1]) report, which has also been analysed in many publications, including Lindhjem et al. (2011[3]), Robinson et al. (2019[4]) and Masterman and Viscusi (2018[2]). The construction of the database is described in OECD (2012[1]) and was intended to provide a comprehensive global coverage of all SP studies published in English, including literature not published in peer-reviewed outlets.
The OECD 2008 database contains about 850 VSL estimates from SP studies published between 1970 and the end of 2008 across 38 countries (Lindhjem et al., 2011[3]). A large range of study-related variables were extracted and coded from these SP studies. VSL estimates were either collected directly from studies or (for a smaller share1) calculated using reported willingness to pay (WTP) estimates for corresponding changes in mortality risks. To ensure the estimates were comparable over time and across countries, estimates expressed in local currencies were adjusted to 2005 price levels using the consumer price index and converted to 2005 USD using purchasing power-adjusted exchange rates (PPPs)2. This dataset is referred to as “Old SP” throughout the current report.
3.1.2. Revealed preference data (pre-2009)
The data referred to as “Old RP” in Figure 3.1 consists of 953 VSL estimates reported across 68 primary valuation studies (Masterman and Viscusi, 2018[2]).3 The full list of papers comprising this dataset is provided in the supplementary materials to Masterman and Viscusi (2018[2]). Of this sample, 518 VSL estimates are from the United States. As with the “Old SP” database, the search for studies included in the “Old RP” dataset did not follow systematic review procedures. The construction of the database involved identifying all labour market VSL estimates from previous meta-analyses and supplementing these by searching the economics database EconLit for papers mentioning “VSL” or “Value of a Statistical Life”, as well as searching the references in these studies for additional papers.4 The underlying dataset was made available for the current analysis by the authors (Viscusi and Masterman), and only studies from before 2009 are used for this analysis.5 This dataset is referred to as “Old RP” throughout the current report.
3.1.3. New meta-data using a systematic review procedure (2009 – 2024)
The new meta-data collected for the current analysis was compiled through a search for VSL estimates published between January 2009 and January 20246. Primary valuation studies were identified using systematic review procedures (Moher et al., 2009[5]; Page et al., 2021[6]) and from works cited by other recent literature on the subject (Jacobsen, Boyers and Avenell, 2020[7]; Mazzei et al., 2020[8]). This collection approach, based on the PRISMA 2020 guidelines, constitutes the best practice in current research synthesis and meta-analysis methodologies (Page et al., 2021[6]). The transparency of the search methodology is designed to enable reproducibility and to avoid any potential influence (direct or indirect) that researchers may have on the search procedure and results of the data collection effort.7
The aim of the latest review was to identify global SP and RP studies published in English,8 that report VSL estimates based on primary data. VSL estimates of interest are for the adult population, implying that VSL is a measure of how adults value mortality risks for themselves rather than for others, including children and family members. The search methodology employed was pre-registered at the Center for Open Science (see Annex A for details). The search process underwent trial periods to assess the extent to which the search terms and included databases capture the studies that have been collected by other sources (e.g. recent reviews such as (Keller et al., 2021[9])) and to evaluate the functionality of the internal screening procedure, in terms of the relevance of the returned studies and reported VSL estimate(s). Screening of the articles gathered by the search was performed manually, e.g. by reading abstracts to assess their relevance.
Following the screening process, the collected literature was reviewed by the project’s advisory group, which revealed several missing studies that should have been identified (and/or not screened out) in the search. Consequentially, search terms were updated, additional database sources were added, internal screening procedures were refined, and the search methodology was finalised in January 2024. The full list of studies included in the collected meta-data is provided in Annex H. The RP and SP studies published after 2008 that were identified through this search will subsequently be referred to as “New RP” and “New SP”, respectively, or collectively as the “New” data.
Figure 3.2 depicts the various steps comprising the PRISMA2020 methodology to construct the meta-data database, including identification, screening, eligibility and final study selection. The first step of the process (Identification) consisted of a search of scientific databases for VSL studies, which returned a total of 1 561 records. The next step (Screening) eliminated duplicate records, yielding 1 026 studies. In a subsequent step (Eligibility), the abstracts and full text of all records were manually checked against several eligibility criteria (cf. box in Figure 3.2 related to Full-text articles excluded), which resulted in the removal of 870 records. This search process identified a final number of 156 eligible records, of which 100 SP studies and 56 were RP studies.
Source: Produced by the authors.
The next step in the review and meta-analysis process was to extract information from the included primary valuation studies and code these as quantitative variables (e.g. dummies, categorical or continuous variables) in order to be able to provide descriptive information about the characteristics of the body of studies collected (Section 3.2), analyse factors that can explain variation in VSL estimates (Chapter 4) and better understand the relationship between VSL estimates and various factors that could potentially be used to adjust base VSL estimates for use in different policy contexts (Chapter 5). The variable of interest in this analysis is the estimate of VSL, i.e. the so-called “effect size” in the meta-analysis. To ensure the coherence of the analysis, the coding procedure for New SP was based in part on the way in which variables were coded in the Old SP database. The coding procedure for New RP data was inspired by and cross-checked against the Old RP database provided by Masterman and Viscusi (2018[2]). The coded variables are defined and discussed in Section 4.3 of Chapter 4, and descriptive statistics of all coded variables are provided in Annex C.
While a number of variables related to the quality of the primary valuation studies are coded at this stage, the overall quality of the studies included in the meta-data is not assessed until a later stage (cf. Section 4.3.3 of Chapter 4). Note that it is appropriate to determine whether to exclude or assign lower weights to studies considered to be of lower quality before undertaking an analysis of the data.
3.1.4. Conversion of VSL for comparability between data sources
The current analysis extracts VSL estimates reported anywhere in the full texts of the identified studies.9 There is considerable variation in how VSL estimates are reported. For example, studies generally report estimates in different currencies and for different years and only a small minority uses consumer price index (CPI) adjustments. To render VSL estimates comparable within and between data sources, all VSL estimates are converted to 2022 USD by using domestic CPIs10 and purchasing power parity (PPP) rates from 2022 (the latest year where data was widely available at the time of analysis). The conversion formula for the currency/country in year is:
Equation 3.1
The information used for the conversion was sourced from official OECD statistics.11 This conversion generally followed the same procedure as in OECD (2012[1]) which remains relevant in the current literature. US EPA (2016[10]) used estimates of income elasticity of VSL and GDP per capita to adjust VSL estimates over time (as increasing income can be assumed to affect VSL estimates) when compiling meta-data for analysis. However, this procedure was later advised against by the US EPA Scientific Advisory Board (SAB) (SAB/USEPA, 2017[11]).12,13 The details of the conversion procedure (including exceptional cases where it was not straightforward) is provided in Annex B. Unless otherwise noted, the notation USD is used throughout the report to denote constant 2022 US Dollars, calculated using the formula in Equation 3.1.
This chapter provides descriptive information about the meta-data sources. The purpose is to provide an overview of the data used in the statistical analysis carried out in Chapter 4, which will serve as the basis of the recommendations for use of VSL estimates in Chapter 6.
3.2.1. Number of studies and estimates
Figure 3.3 shows the total number of studies identified by the search procedure in each of the four databases. More SP studies (99) were identified in the period from 2009 until 2024 as compared to the previous database covering the earliest SP applications until the end of 2008 (72). For RP, only slightly more studies (4) were identified in the period after 2009 compared with older studies. On average, more estimates are provided per study for RP studies (27) than for SP studies (10).
Note: Old refers to primary valuation studies published before 2009, and New refers to primary valuation studies published since 2009.
Figure 3.4 and Figure 3.5 further divide primary valuation studies across databases into different types of SP and RP valuation methods. Comparing the valuation methods used by studies in the old and new SP databases, Figure 3.4shows that the use of choice experiments (CEs) has gained popularity over time relative to contingent valuation (CV). In the new SP data, the method mix is approximately equally split between the two methods. For the RP data, the Old RP dataset includes only studies using hedonic wage (HW) methods, whereas the New RP dataset also includes consumer market (CM) studies. In the New RP dataset, 9 out of 55 RP studies use CM methods.
Note: Old refers to studies published before 2009, and New refers to studies published since 2009.
Note: Old refers to studies published before 2009, and New refers to studies published from 2009.
3.2.2. Growth in mortality risk valuation studies over time
Figure 3.6 shows the total number of SP and RP valuation studies reporting VSL estimates by year, indicating substantial growth in the number of studies over time. While RP studies tended to outnumber SP studies in the years before 2000, there has been significant growth in SP studies since then, leading to SP studies outnumbering RP studies by the end of 2023. Note that a lag of several years can exist between the underlying RP and SP data and the publication of VSL estimates. This is due in part to the use of HW data, which reflect long time series and also due to the time needed for analysis before VSL estimates are published.
3.2.3. Geographical distribution of SP and RP studies
Figure 3.7 and Figure 3.8 show the geographical distribution of primary valuation studies using SP and RP methods, respectively. While both methods are used primarily in countries with relatively higher income levels, more recent studies tend to include a broader spectrum of countries and income levels, including in Latin America, Africa and Asia. Traditionally, RP methods have been used more frequently in North America, especially in the United States, compared to Europe and the rest of the world. Further, RP methods have a narrower geographical distribution then SP methods. SP methods are generally considered easier to implement given that conducting population surveys is relatively straightforward, while RP methods, especially hedonic wage studies, require large datasets on wage and risk factors for different sectors, which are typically not easily available or available at all in lower-income countries.
3.2.4. Characteristics of risks valued
Figure 3.9 provides an overview of the sources of risks underlying the VSL estimates reported in the New SP and New RP data. Non-contagious health-related risks are the largest single type of data source for the basis of SP estimates (501 estimates). These risks include fatality causes such as cancer, heart disease and stroke. The second largest source of risks include transport-related risks (163), followed by natural disasters (147) which include fatalities from events such as avalanches, floods and earthquakes. Climate-related risks (87), which include extreme weather, sea level change and the depletion of fisheries, and heat stroke also constitute a significant portion of risk estimates, as well as environmental risks (62) such as air and water pollution and contagious disease-related health risks (“virus” in Figure 3.1) (51)14. Several risk categories have very few estimates, including those related to e.g. crime, man-made (i.e. not natural) disasters and suicide. For the New SP, the mean (median) annualised risk change levels valued in the primary studies is 0.0015 (0.00015) for CV and 0.001 (0.0001) for CE.15 Regarding the sources of risk that serve as the basis for RP VSL estimates, many are job-related risks, the specific causes of which can be multiple and difficult to isolate (899 estimates). Health-related risks (94), transportation-related risks (77), and environment-related risks (57) are the next most common types of risk treated in studies in the New RP dataset. Several VSL estimates pertain to risk changes regarding contagious diseases (19), climate change (23) and natural disasters (12). For HW especially, it is not straightforward to estimate the risk levels associated with the reported VSL estimates.
New studies, stated preference and revealed preference
Note: The “health” category refers to non-viral diseases (e.g. heart disease). “Environment” refers to environmental hazards other than climate events and natural disasters (e.g. air pollution). “Disaster” category refers to man-made disasters (e.g. nuclear accident). In the New SP dataset, a number of missing values that have been assigned to the “Health” category based on the coding procedure for risk types used in the Old SP dataset.
The New SP and New RP datasets are also coded to distinguish between acute (traumatic) and chronic (illness-related) causes of mortality risks. 38% of VSL estimates are related to acute risks, 14% to chronic risks, and 48% relate to mixed risks, which are considered to be risks that are a mix of chronic and acute risks (as can be the case in many RP studies) or risks that are not clearly specified or easy to determine. Acute risks could for example include traffic accidents, while chronic risks are typically related to illnesses or conditions that last for a period of time before death, such as cancer.
This chapter reports summary statistics of the VSL estimates included in the meta-analysis. Depending on how the VSL estimates from each study is treated the results of the meta-analysis could potentially be different. The mean of VSL at the “estimate level” is calculated as the mean of all individual VSL estimates reported in all studies, with each estimate given equal weight. In contrast the mean of VSL at the “study level” reflects the mean VSL estimate calculated using a single estimate from each study. Each study-level estimate is calculated as the mean of all of the estimates reported in the study. Mean VSL at the “estimate level” implicitly gives more weight to the results of studies that report many estimates.
3.3.1. Estimate level VSLs in the dataset
Table 3.1 reports the basic descriptive statistics of VSL estimates for the four different datasets, including mean, standard deviation (SD) and number of studies (N). Following common statistical procedures, extreme outliers (or “far-outs”) are excluded from the values reported in the table.16 Outliers are defined based on the interquartile range and are further discussed in the context of the meta-analysis of the VSL data in Chapter 4.
For all data combined regardless of the valuation method, i.e. a total of 3 986 individual VSL estimates from 280 studies, the overall mean VSL estimate at the estimate level is USD 7.8 million (Table 3.1) .For the respective datasets New SP, New RP, Old SP and Old RP, the mean of VSL estimates at the estimate level are USD 7.3, 6.5, 6.8 and 12.9 million. Hence, the differences among the first three datasets are relatively small. Standard deviations of the mean of VSL estimates at the estimate level across the three meta-data sources are also similar.
Considering SP and RP data separately, and merging the Old SP and New SP datasets together, as well as the Old RP and New RP datasets together, yields mean VSL estimates of USD 7.1 and 8.4 million at the estimate level for SP studies and RP studies, respectively. Hence, the difference between the two methods, considering all publication dates, is not very large. For the new data only, the mean VSL estimate is USD 6.8 million at the estimate level. Note that these values reflect overall descriptive statistics and that a more robust comparison would account for other differences in the data that can explain variation in VSL estimates, such as income levels. The impact of such differences on VSL estimates will be examined in the meta-analysis conducted in Chapter 4, which controls for a number of such factors.
Excluding far-outs appears to have a substantial influence on the estimate-level mean VSLs. Including these values would result in an increase in the estimates to USD 60, 8.6, 12.9 and 15 .3 million for the New SP, New RP, Old SP and Old RP meta-data, respectively. Descriptive statistics of the various datasets, with and without outlier values, are provided in Annex C.
Note: SP refers to “Stated Preference”, and RP refers to “Revealed preference”. Old refers to studies published before 2009, and New refers to studies published in 2009 and later. This table presents summary statistics of all individuals VSL estimates collected across studies. The results exclude outliers (“far-outs”), which are defined as VSL . Estimates are inflation-adjusted to 2022 from the year a VSL estimate is reported according to the procedure described in Section 3.1.4.
Figure 3.10 reports the quartiles, median, minimum and maximum values of estimate-level VSLs. The left-hand side of the figure displays estimates for the four datasets, while the right-hand side of the figure displays estimates for various merged groups of meta-data. Median and mean estimate-level VSL estimates are relatively similar across the New SP, Old SP and New RP datasets. The width of the distributions (reflecting the interval between the 25% and 75% quartiles) is reflected in the width of the boxes. The horizontal lines (“whiskers”) reflect minimum and maximum values. Note that although some VSL estimates, especially for HW studies, may be reported as negative in the statistical analyses in primary valuation studies, they are nevertheless included here in order to reduce potential publication bias17. Mean estimate-level VSL estimates for New SP and RP together (“All New”) are very similar to the full dataset (“All merged”).
Note: Outliers (“far-outs”) are excluded, defined as VSL . Blue dots reflect mean VSL estimates, the midlines of boxes reflect median VSL estimates, the upper boundary of the boxes reflect the 75% quartile of VSL estimates, the lower boundary of the boxes reflect the 25% quartile of VSL estimates, and the horizontal lines (“whiskers”) reflect the minimum and maximum VSL estimates, respectively.
Annex C provides additional detailed information, including a display of the cumulative distribution functions of VSL estimates in the four datasets and for those obtained using RP and SP methods, respectively.
3.3.2. Study level VSLs in the dataset
Mean study-level VSL estimates are calculated as USD 5.3, 8.1, 6.4 and 11.3 million for the New SP, New RP, Old SP and Old RP datasets, respectively (Table 3.2). Mean study-level VSL estimates are lower than mean estimate-level VSL estimates, particularly in the New SP dataset. Smaller differences are observed between the mean estimates of VSL at the study and estimate levels for RP, SP and All new data combined. The overall mean VSL estimate at the study level combining all meta-data sources is relatively unchanged, at USD 7.2 million (compared to USD 7.8 million at the estimate level in Table 3.1.
One reason for the difference between study-level and estimate-level results is that high VSL estimates are likely to be generated as part of the robustness checks and sensitivity analyses that may be carried out in primary valuation studies. Moreover, such high estimates are more likely to be reported by studies that report many estimates, including from robustness checks and sensitivity analyses, whereas studies that produce fewer estimates are less likely to include very high estimates. Therefore “close-to-the-average” estimates from studies reporting fewer estimates may be underweighted when viewing VSL estimates at the estimate-level. This underweighting is reduced in the calculation of mean VSL estimates at the study level.
Excluding extreme values (far-outs) also has a substantial influence on the mean VSLs at the study level. For example, for the new SP data, the mean study-level VSL estimate increases from USD 5.3 million to USD 25.9 million when the extreme values are included. Descriptive statistics of the various datasets, with and without outlier values, are provided in Annex C.
Note: SP refers to “Stated Preference”, and RP refers to “Revealed preference”. Old refers to studies published before 2009, and New refers to studies published from 2009. Study Level = Weighted by the inverse of the number of estimates from each study. Outliers (“far-outs”) are excluded, defined as VSL Far-outs has a slightly different interpretation at the study level compared to Table 3.1. Here, e.g. 3.3 % are removed due to far-out estimates, implying that all estimates from those studies are removed. of VSL estimates are inflation-adjusted to 2022 from the year a VSL estimate is reported according to the procedure described in Section 3.1.4 of Chapter 3.
Figure 3.11 reports the quartiles, mean, median and minimum and maximum estimates of study-level VSL estimates.
Note: Outliers (“far-outs”) are excluded, defined as VSL . Blue dots reflect mean VSL estimates, the midlines of boxes reflect median VSL estimates, the upper boundary of the boxes reflect the 75% quartile of VSL estimates, the lower boundary of the boxes reflect the 25% quartile of VSL estimates, and the horizontal lines (“whiskers”) reflect the minimum and maximum VSL estimates, respectively.
References
[21] Banzhaf, H. (2022), “The Value of Statistical Life: A Meta-Analysis of Meta-Analyses”, Journal of Benefit-Cost Analysis, Vol. 13/2, pp. 182-197, https://doi.org/10.1017/BCA.2022.9.
[12] Bellavance, F., G. Dionne and M. Lebeau (2009), “The value of a statistical life: A meta-analysis with a mixed effects regression model”, Journal of Health Economics, Vol. 28/2, pp. 444-464, https://doi.org/10.1016/J.JHEALECO.2008.10.013.
[17] Fabiano, N. et al. (2024), “How to optimize the systematic review process using AI tools”, JCPP Advances, Vol. 4/2, https://doi.org/10.1002/JCV2.12234.
[16] Gentry, E. and W. Viscusi (2016), “The fatality and morbidity components of the value of statistical life”, Journal of Health Economics, Vol. 46, pp. 90-99, https://doi.org/10.1016/J.JHEALECO.2016.01.011.
[22] Ginbo, T., W. Adamowicz and P. Lloyd-Smith (2023), “Valuing Mortality Risk Reductions in Canada: An Updated Meta-Analysis and Policy Guidance”, Canadian Public Policy, Vol. 49/3, pp. 233-251, https://doi.org/10.3138/CPP.2022-052.
[7] Jacobsen, E., D. Boyers and A. Avenell (2020), “Challenges of Systematic Reviews of Economic Evaluations: A Review of Recent Reviews and an Obesity Case Study”, PharmacoEconomics, Vol. 38/3, pp. 259-267, https://doi.org/10.1007/S40273-019-00878-2.
[9] Keller, E. et al. (2021), “How Much Is a Human Life Worth? A Systematic Review”, Value in Health, Vol. 24/10, pp. 1531-1541, https://doi.org/10.1016/J.JVAL.2021.04.003.
[3] Lindhjem, H. et al. (2011), “Valuing Mortality Risk Reductions from Environmental, Transport, and Health Policies: A Global Meta-Analysis of Stated Preference Studies”, Risk Analysis, Vol. 31/9, pp. 1381-1407, https://doi.org/10.1111/J.1539-6924.2011.01694.X.
[2] Masterman, C. and W. Viscusi (2018), “The Income Elasticity of Global Values of a Statistical Life: Stated Preference Evidence”, Journal of Benefit-Cost Analysis, Vol. 9/3, pp. 407-434, https://doi.org/10.1017/BCA.2018.20.
[8] Mazzei, D. et al. (2020), “Systematic review of economic evaluations investigating education, exercise, and dietary weight management to manage hip and knee osteoarthritis: Protocol”, Systematic Reviews, Vol. 9/1, https://doi.org/10.1186/S13643-020-01492-6.
[5] Moher, D. et al. (2009), “Preferred reporting items for systematic reviews and meta-analyses: The PRISMA statement”, PLoS Medicine, Vol. 6/7, https://doi.org/10.1371/JOURNAL.PMED.1000097.
[19] OECD (2025), Annual Purchasing Power Parities and exchange rates, https://data-explorer.oecd.org/vis?lc=en&fs[0]=Topic%2C1%7CEconomy%23ECO%23%7CNational%20accounts%23ECO_NAD%23&fs[1]=Topic%2C2%7CEconomy%23ECO%23%7CNational%20accounts%23ECO_NAD%23%7CGDP%20and%20non-financial%20accounts%23ECO_NAD_GNF%23&pg=0&fc=Topic&snb=53&df[ds]=dsDisseminateFinalDMZ&df[id]=DSD_NAMAIN10%40DF_TABLE4&df[ag]=OECD.SDD.NAD&df[vs]=1.0&pd=%2C&dq=A.AUS%2BAUT%2BBEL%2BCAN%2BCHL%2BCOL%2BCRI%2BCZE%2BDNK%2BEST%2BFIN%2BFRA%2BDEU%2BGRC%2BHUN%2BISL%2BIRL%2BISR%2BITA%2BJPN%2BKOR%2BLVA%2BLTU%2BLUX%2BMEX%2BNLD%2BNZL%2BNOR%2BPOL%2BPRT%2BSVK%2BSVN%2BESP%2BSWE%2BCHE%2BTUR%2BGBR%2BUSA...PPP_B1GQ.......&to[TIME_PERIOD]=false&vw=ov (accessed on 13 September 2024).
[18] OECD (2025), Inflation (CPI) (indicator), https://doi.org/10.1787/eee82e6e-en (accessed on 5 February 2025).
[1] OECD (2012), Mortality Risk Valuation in Environment, Health and Transport Policies, OECD Publishing, Paris, https://doi.org/10.1787/9789264130807-en.
[6] Page, M. et al. (2021), “PRISMA 2020 explanation and elaboration: Updated guidance and exemplars for reporting systematic reviews”, The BMJ, Vol. 372, https://doi.org/10.1136/BMJ.N160.
[4] Robinson, L., J. Hammitt and L. O’Keeffe (2019), “Valuing Mortality Risk Reductions in Global Benefit-Cost Analysis”, Journal of Benefit-Cost Analysis, Vol. 10, pp. 15-50, https://doi.org/10.1017/BCA.2018.26.
[11] SAB/USEPA (2017), Review of EPA’s Proposed Methodology for Updating Mortality Risk Valuation Estimates for Policy Analysis.
[23] Tukey, J. (1977), Exploratory Data Analysis, Springer, New York, NY, https://doi.org/10.1007/978-0-387-32833-1_136.
[20] Tunçel, T. and J. Hammitt (2014), “A new meta-analysis on the WTP/WTA disparity”, Journal of Environmental Economics and Management, Vol. 68/1, pp. 175-187, https://doi.org/10.1016/J.JEEM.2014.06.001.
[10] USEPA (2016), Valuing mortality risk reductions for policy: a meta-analytic approach. White Paper, https://www.epa.gov/system/files/documents/2025-04/vsl-white-paper_final_020516-1.pdf.
[13] Viscusi, W. (2015), “Reference-Dependence Effects in Benefit Assessment: Beyond the WTA–WTP Dichotomy and WTA–WTP Ratios”, Journal of Benefit-Cost Analysis, Vol. 6/1, pp. 187-206, https://doi.org/10.1017/BCA.2015.3.
[14] Viscusi, W. and J. Aldy (2003), “The Value of a Statistical Life: A Critical Review of Market Estimates Throughout the World”, Journal of Risk and Uncertainty, Vol. 27/1, pp. 5-76, https://doi.org/10.1023/A:1025598106257.
[15] Viscusi, W. and C. Masterman (2017), “Anchoring biases in international estimates of the value of a statistical life”, Journal of Risk and Uncertainty, Vol. 54/2, pp. 103-128, https://doi.org/10.1007/S11166-017-9255-1.
[24] World Bank (2024), International Comparison Program, World Bank | World Development Indicators database, World Bank | Eurostat-OECD PPP Programme, https://data.worldbank.org/indicator/NY.GDP.MKTP.PP.CD?name_desc=false&year=2022.
← 1. This share was ca. 7.7 per cent (78 estimates).
← 2. An international dollar would buy a comparable amount of goods and services in the cited country that a U.S. dollar would buy in the United States. This term is often used in conjunction with Purchasing Power Parity (PPP) data (World Bank, 2024[24]). In what follows, the abbreviation “USD” is used to refer to this measure of international US dollars.
← 3. The database contains estimates from each study included in Bellavance et al. (2009[12]), Viscusi (2015[13]), Viscusi and Aldy (2003[14]) and Viscusi and Masterman (2017[15]). The full list of papers from which they draw VSL estimates can be found in the Supplementary materials to Masterman and Viscusi (2018[2]).
← 4. Based on personal communication with Masterman and Viscusi. One additional paper was identified and included due to co-authorship (Gentry and Viscusi, 2016[16]).
← 5. Data from 2009 to 2016 in the dataset from Masterman and Viscusi (2018[2]) is used to crosscheck the quality of RP data in corresponding years collected for the current analysis (e.g. with respect to variable coding procedures).
← 6. This included academic journals and other research outlets, the majority of which were subject to some type of peer review. 71% of the VSL estimates in the coded dataset were published in journals (cf. Annex C). Of the remaining 29%, 20% were classified as “unpublished”, 3% as conference proceedings, 6% as theses and less than 1% as book sections.
← 7. The search methodology for the current analysis has not made use of artificial intelligence tools. Although artificial intelligence may have potential for contributing to search methodologies (Fabiano et al., 2024[17]), the use of AI tools is still relatively immature in this field as of 2024.
← 8. This is the conventional approach in the systematic reviews and meta-analysis papers known to the authors, including OECD (2012[1]). To investigate potential limitations or valuation biases introduced by only using English-language studies, prominent researchers based in non-English speaking countries were consulted. For several countries (e.g. Poland, China, Germany, Spain, France, Nordic countries) it was confirmed that most reasonable-quality studies are published in English language journals. For some countries such as Japan and France (and some Latin American countries), where publication in the national language has historically been more common, some additional studies were identified. However, among the contacted researchers, none expressed concerns that ignoring a relatively small number of non-English VSL studies would lead to a systematic bias in the VSL estimates presented in this study.
← 9. In contrast to the database used for OECD (2012[1]), the current analysis does not include VSL estimates derived from other available measures (e.g. WTP for a specified change in mortality risk). This choice was made to avoid any calculation errors and/or misunderstandings related to the intentions of original study authors and to make the data collection procedure as transparent and replicable as possible. As noted above, the proportion of such estimates was relatively small in the Old SP database.
← 10. Note that this practice assumes that the way people value mortality risks over time follows the same pattern as their WTP for the basket of consumer goods the CPI is based on. This constitutes standard practice in the meta-analysis literature. Since there is no “CPI for mortality risk reductions” and making other assumptions is difficult to justify.
← 11. CPI and PPP values are retrieved from OECD (2025[18]) and OECD (2025[19]), respectively, and are available from the OECD Data Explorer: https://www.oecd.org/en/data/datasets/oecd-DE.html.
← 12. The US EPA SAB states that it is not appropriate to conduct this transformation as part of constructing the input data for meta-analysis due to the fact that it makes strong assumptions about income elasticity of VSL over time. However, the SAB considers that it may be appropriate when making recommendations for the use of VSL estimates in specific applications (i.e. benefit transfer). The SAB also noted that VSL estimates from RP hedonic wage studies are derived from Marshallian willingness to accept a marginal increase in mortality risk (i.e. holding income constant), while VSL estimates from SP studies are derived from Hicksian WTP for a marginal reduction in mortality risk (i.e. holding utility constant). As has been shown in the literature, willingness to accept is often found to be larger than WTP (Tunçel and Hammitt, 2014[20]). The advisory board offers no specific solution to this issue, and for the purposes of this analysis no attempt is made to correct for this potential disparity, as this would not align with the existing meta-analysis literature.
← 13. Banzhaf (2022[21]) uses different assumptions about income adjustment of VSL, including assuming an income elasticity of VSL of zero for sensitivity analysis, stating that: “Some analysts may prefer this more conservative approach, especially given the uncertainty about the role of permanent income versus current income, cross-sectional variation in income and time series variation, and the measurement of income“ (p 194-195).
← 14. The following definitions were used risk causes: Disaster-related risk: man-made disasters; Climate-related risk: extreme weather, sea level change and depletion of fisheries, heat stroke; Virus: COVID-19, dengue fever, rabies.
← 15. Note that CE studies by design provide many risk change levels. For these studies risk change levels were averaged over the ranges used in CE designs. For RP studies it is often not possible or meaningful to derive risk measures from primary valuation studies. As discussed in e.g. Ginbo et al. (2023[22]), many RP studies do not report the risk differential between alternative and risky jobs, so that it is impossible to identify the exact level of risk change valued in RP studies.
← 16. Following the definition of Tukey’s fences, far-outs are considered to be values that falling outside of an interval delimited by a lower value equal to the 25th percentile minus three times the difference between the 25th and the 75th percentile, and an upper value equal to the 75th percentile plus this amount (Tukey, 1977[23]). Formally VSL ∉ . Note that it has not been assessed whether and how individual studies have addressed potential extreme values in their primary SP or RP data.
← 17. Approximately 2.5% of VSL estimates in the full dataset are equal to or less than zero.