- Review
- Open Access
- Open Peer Review
Are infant mortality rate declines exponential? The general pattern of 20^{th} century infant mortality rate decline
- David Bishai^{1}Email author and
- Marjorie Opuni^{1}
https://doi.org/10.1186/1478-7954-7-13
© Bishai and Opuni; licensee BioMed Central Ltd. 2009
- Received: 26 November 2008
- Accepted: 23 August 2009
- Published: 23 August 2009
Abstract
Background
Time trends in infant mortality for the 20^{th} century show a curvilinear pattern that most demographers have assumed to be approximately exponential. Virtually all cross-country comparisons and time series analyses of infant mortality have studied the logarithm of infant mortality to account for the curvilinear time trend. However, there is no evidence that the log transform is the best fit for infant mortality time trends.
Methods
We use maximum likelihood methods to determine the best transformation to fit time trends in infant mortality reduction in the 20^{th} century and to assess the importance of the proper transformation in identifying the relationship between infant mortality and gross domestic product (GDP) per capita. We apply the Box Cox transform to infant mortality rate (IMR) time series from 18 countries to identify the best fitting value of lambda for each country and for the pooled sample. For each country, we test the value of λ against the null that λ = 0 (logarithmic model) and against the null that λ = 1 (linear model). We then demonstrate the importance of selecting the proper transformation by comparing regressions of ln(IMR) on same year GDP per capita against Box Cox transformed models.
Results
Based on chi-squared test statistics, infant mortality decline is best described as an exponential decline only for the United States. For the remaining 17 countries we study, IMR decline is neither best modelled as logarithmic nor as a linear process. Imposing a logarithmic transform on IMR can lead to bias in fitting the relationship between IMR and GDP per capita.
Conclusion
The assumption that IMR declines are exponential is enshrined in the Preston curve and in nearly all cross-country as well as time series analyses of IMR data since Preston's 1975 paper, but this assumption is seldom correct. Statistical analyses of IMR trends should assess the robustness of findings to transformations other than the log transform.
Keywords
- Gross Domestic Product
- Infant Mortality Rate
- Rate Ratio
- Logarithmic Transformation
- Exponential Decline
Background
The assumption that infant mortality rate (IMR) declines are non-linear as an economy develops is enshrined in the Preston curve [1]. The curve depicts the relationship between life expectancy and income and shows that the relationship is non-linear with life expectancy in wealthier countries being less sensitive to variations in average income. Many analyses of IMR declines over the last three decades have used the logarithmic transform of IMR [2–7]. And many studies of infant mortality inequalities among sub-population within countries have done the same [3, 8].
A moment's reflection reveals that an infant mortality rate cannot keep declining linearly forever, since at some point it would reach zero, and it can never be negative. Although the logarithmic transformation is convenient, there is no evidence that logarithmic transformations are the most appropriate transformations for infant mortality rate time trends.
Whether or not IMR time series require logarithmic transformations has important implications for analyses of convergence of IMR both across countries and among sub-populations within countries. If IMR decline is exponential and requires a logarithmic transform, this implies more rapid IMR convergence. In contrast, if IMR decline is closer to linear, IMR convergence is expected to be slower.
In this paper, our objectives are twofold. First, we assess whether IMR decline as experienced in the 20^{th} century was a process of exponential decline in which case past analyses were wise to use logarithmic transformations of IMR in linear regressions. This objective will be met in the process of determining the best transformation to fit time trends in IMR reduction from 18 countries that span most of the 20^{th} century. Second, we examine how important it is to use the proper transformation of IMR trends in models. We do this by comparing regressions of 1n(IMR) on same year GDP per capita against appropriately transformed models. We demonstrate that the logarithmic transformation of IMR declines is seldom appropriate. And we show that assuming that IMR declines are logarithmic without checks can induce biases.
Methods
We use two separate datasets for this analysis. First, we obtain IMR time series from the Human Mortality Database (HMD) http://www.mortality.org/. A collaborative project between the University of California, Berkeley and the Max Planck Institute for Demographic Research, the HMD provides detailed mortality time series for populations where death registration and census data are virtually complete. The methods used to compute the mortality rates and life tables in the HMD are described elsewhere [12]. We note only that in the absence of complete data, demographers frequently rely on interpolation to patch together annual mortality estimates from a collection of overlapping household surveys. An analysis of heavily interpolated data would yield limited insight on the best transformation to fit time trends in infant mortality reduction. The second dataset we use is the historical GDP per capita time series derived by Angus Maddison http://www.ggdc.net/maddison/. The data are in 1990 International Geary-Khamis dollars. Details on the data and the sources for the time series are available elsewhere [13–17]. The selection of countries for this analysis is driven by data availability. All countries with IMR data and GDP per capita data for at least half of the 20^{th} century are included.
arguing that the transformation on a dependent variable could make the residuals more closely normal and more homoskedastic [18]. Because the transformation embeds a number of popular functional forms including the reciprocal, logarithmic and square root transformations, it has also been used as a method for testing functional form [19]. When λ = 1, the Box-Cox transformation amounts to a simple linear model. As λ approaches -1, 0 and 1/2, the Box-Cox transformation converges to reciprocal, logarithmic, and square root transformations, respectively. (To see the limit of (a^{ λ }-1)/λ as λ → 0, it is necessary to first Taylor expand both log(a) and BoxCox(a) and then observe their identity in the limit.) One disadvantage of the Box-Cox transformation is that it breaks down when zero or negative values must be transformed but this is not an issue in the analysis of IMR data.
Results from Box-Cox models in which (IMR^{ λ }-1)/λ is regressed against time
Country | Years | λ | Chi-squared for null that λ= 0 (Logarithmic model) | Chi-squared for null that λ= 1 (Linear Model) |
---|---|---|---|---|
Whole sample | 0.077 | (22.050) *** | (2131.57) *** | |
Australia | 1921–1999 | 0.249 | (37.750) *** | (153.650) *** |
Austria | 1947–1999 | 0.102 | (9.200) *** | (165.330) *** |
Belgium | 1900–1913 1918–1999 | 0.379 | (67.180) *** | (130.450) *** |
Canada | 1921–1999 | 0.151 | (50.290) *** | (288.830) *** |
Denmark | 1900–1999 | 0.260 | (40.530) *** | (169.090) *** |
Finland | 1900–1999 | 0.356 | (58.640) *** | (130.600) *** |
France | 1900–1999 | 0.354 | (52.030) *** | (118.600) *** |
Italy | 1900–1999 | 0.587 | (122.850) *** | (63.540) *** |
Japan | 1947–1999 | -0.321 | (62.000) *** | (238.210) *** |
Netherlands | 1900–1999 | -0.092 | (10.680) *** | (342.410) *** |
New Zealand | 1947–1999 | 0.591 | (49.640) *** | (27.030) *** |
Norway | 1900–1999 | 0.364 | (90.730) *** | (169.360) *** |
Portugal | 1940–1999 | 0.523 | (74.340) *** | (64.170) *** |
Spain | 1908–1999 | 0.450 | (71.570) *** | (82.540) *** |
Sweden | 1900–1999 | 0.192 | (39.310) *** | (234.810) *** |
Switzerland | 1900–1999 | 0.133 | (41.060) *** | (352.000) *** |
UK | 1900–1999 | 0.136 | (34.220) *** | (310.440) *** |
USA | 1933–1999 | 0.040 | (0.990) | (168.680) *** |
Results
Table 1 shows the results of the Box-Cox models where (IMR^{ λ }- 1)/λ is regressed against year for each country. The third column in Table 1 includes the values of λ for each country and for the pooled sample that yield the most appropriate transformations. The fourth and fifth columns in Table 1 include the results of chi-squared tests comparing the best fit λ for each country and for the pooled sample for the null that λ = 0 (logarithmic model) and for the null that λ = 1 (linear model). Infant mortality rate decline during the 20^{th} century is best described as an exponential decline only for the United States. For the remaining 17 countries included in our study, the best transformations of IMR decline are neither linear nor logarithmic. The best fit for the pooled sample of IMR decline occurs with a transformation of (IMR^{0.077}-1)/0.077.
A comparison of coefficients on GDP per capita by IMR transformation
Country | Years | Box-Cox IMR transformation λ = "Best Fit" | Log IMR transformation λ = 0 | |
---|---|---|---|---|
Coefficient on GDP per capita | Coefficient on GDP per capita | Coefficient gap | ||
Whole sample | -0.065 *** | -0.078 ** | 0.013 | |
Australia | 1921–1999 | 0.217 *** | -0.034 ** | 0.251 |
Austria | 1947–1999 | 0.151 *** | 0.079 | 0.072 |
Belgium | 1900–1913 1918–1999 | -0.070 *** | -0.09 ** | 0.02 |
Canada | 1921–1999 | -0.038 *** | -0.051 ** | 0.013 |
Denmark | 1900–1999 | -0.057 ** | -0.06 ** | 0.003 |
Finland | 1900–1999 | -0.088 *** | -0.106 ** | 0.018 |
France | 1900–1999 | -0.084 *** | -0.11 ** | 0.026 |
Italy | 1900–1999 | -0.151 *** | -0.132 ** | -0.019 |
Japan | 1947–1999 | -0.020 *** | 0.026 | -0.046 |
Netherlands | 1900–1999 | -0.023 *** | -0.004 | -0.019 |
New Zealand | 1947–1999 | 0.035 | 0.068 * | -0.033 |
Norway | 1900–1999 | -0.004 | -0.046 ** | 0.042 |
Portugal | 1940–1999 | -0.366 *** | -0.233 ** | -0.133 |
Spain | 1908–1999 | -0.127 *** | -0.154 ** | 0.027 |
Sweden | 1900–1999 | -0.062 *** | -0.062 ** | 0 |
Switzerland | 1900–1999 | -0.009 | -0.031 ** | 0.022 |
UK | 1900–1999 | 0.145 *** | -0.031 ** | 0.176 |
USA | 1933–1999 | -0.009 | -0.009 | 0 |
Discussion
For the countries and the years studied, the logarithmic transformation of IMR is seldom appropriate. Furthermore the results show that the logarithmic assumption can lead to biases in estimating the coefficient of IMR on covariates like GDP per capita. IMR decline in 18 countries during the 20^{th} century was for the most part neither linear nor an exponential decline. Only one of 18 IMR time series was best transformed as logarithmic. Although the logarithmic transformation of IMR is convenient and commonly used, we caution analysts against habitually taking logs. Statistical analyses of IMR time series should assess the robustness of findings to transformations other than the log transform and such assessments should be undertaken for all IMR time series separately.
This analysis is also relevant to the debate about what to expect when comparing IMR decline across multiple populations. As Figure 2 shows, historically, infant mortality decline occurred at a variable pace. The transition from IMR of 200 to IMR of less than 10 appears to have been the result of a heterogeneous process that was neither always linear nor always exponential throughout the 20^{th} century. A linear process would have meant that societies engaged in improving infant health found it possible to achieve the same size decrements each year regardless of whether IMR was low or high. An exponential process would imply that societies could generally lower IMR more when IMR was higher and less when it was lower. The history of IMR decline suggests that the timing of the process is heavily influenced by punctuated innovations in public health that disseminate laterally across countries [23, 24]. Innovations in public health can provide permanent alterations in mortality levels and alterations in the pace of decline. By the same token, new epidemics such as the HIV epidemic can also punctuate the process.
The present state of the world reveals vast global disparities in infant mortality despite the wide availability of public health knowledge. Differences in material resources certainly provide a large part of the explanation for how international populations can share the same knowledge but achieve disparate mortality rates [4]. Differences in the efficiency of social institutions and health systems can also enable countries with similar resource levels to register disparate mortality levels [2]. What remains particularly perplexing is those intra-national disparities where populations share similar resource levels and health technology but achieve different health outcomes in various regions of the same country.
Tracking progress in health equity requires an objective measure of disparities that will accurately reflect progress when progress occurs. The analysis in this paper indicates that tracking the rate ratio in infant mortality rates (a relative measure) may fail to reflect progress. As shown in Figure 1, the rate ratio can increase asymptotically even though there is a constant gap and equal IMR decrements each year in the linear case or the rate ratio can remain constant even though the gap is narrowing and there are equal percentage drops in IMR every year in the exponential case. This study of the general pattern of IMR decline revealed that the average of 18 countries was for IMR to decline neither linearly nor exponentially. The best fit for data encompassing most of the twentieth century occurred overall when IMR^{0.077} was linear in time. If there were one universal pattern of IMR decline, and it behaved like this best fit model, then the rate ratio of IMR rate ratios (higher divided by lower) would never decline. However there is substantial heterogeneity in the exponent on IMR across countries, and furthermore it is unlikely that the best fitting exponent for any single country remains constant across the span of a century. Assessing the degree of mortality convergence across populations in the 21^{st} century will require more sophisticated measures than rate ratios.
Conclusion
This analysis warns of two hazards in the study of infant mortality rates. The first hazard is to assume that the best transformation of IMR for regression analysis is logarithmic. The second hazard is to assume that the rate ratio is a complete reflection of health disparities between populations. Our recommendations would be for future studies to proceed by always checking a Box-Cox transformation of IMR and using the best-fit Box-Cox parameter. Furthermore, an accurate reflection of disparities would need to supplement relative measures like rate ratios with absolute gap measures as well as measures of how the various aspects of the health system differentially affect the health of subpopulations.
Appendix
The general formula for the derivative of a quotient f(t)/g(t)
Is given by: (f/g) [f'/f - g'/g].
Consider then the quotient of two declining linear functions: (B-bt)/(A-at)
Which is positive as long as B>bt and A > at and aB > bA indicating that the quotient will be increasing for the situation shown in Figure 1 where B > A and the slope a= b.
Now consider the ratio of two declining exponential functions: R= [e^{(B-bt)}/e^{(A-at)}]
The derivative of this quotient equals R [-b + a]. The sign of this derivative is positive if a> b. If a = b, the derivative is zero and the quotient never changes from its equilibrium value of e^{(B-A)}.
Declarations
Acknowledgements
This study received financial support from the Gates Foundation through Grand Challenge 13. Acknowledgements also go to Population Health Metrics and to the universities of Harvard, Johns Hopkins, MIT, and Queensland. Helpful comments from Vladimir Canudas are gratefully acknowledged. All errors are our own.
Authors’ Affiliations
References
- Preston SH: The changing relation between mortality and level of economic development. Popul Stud (Camb) 1975, 29: 231-248. 10.2307/2173509View ArticleGoogle Scholar
- Jamison D, Sandbu M, Want J: Why has infant mortality decreased at such different rates in different countries? Disease Control Priorities Project Working Paper. Washington, DC 2004.Google Scholar
- Kleinman JC: Further comments on 'State trends in infant mortality'. Am J Public Health 1986,76(11):1361-1362. 10.2105/AJPH.76.11.1361-bView ArticlePubMedPubMed CentralGoogle Scholar
- Pritchett L, Summers LH: Wealthier is healthier. Journal of Human Resources 1996,31(4):841-868. 10.2307/146149View ArticleGoogle Scholar
- Schalick LM, Hadden WC, Pamuk E, Navarro V, Pappas G: The widening gap in death rates among income groups in the United States from 1967 to 1986. Int J Health Serv 2000,30(1):13-26. 10.2190/8QMH-4FAB-XAWP-VU95View ArticlePubMedGoogle Scholar
- Singh G, Yu SM: US Childhood Mortality, 1950 through 1993: Trends and Socioeconomic Differentials. AJPH 1996,86(4):505-519. 10.2105/AJPH.86.4.505View ArticleGoogle Scholar
- Baird S, Friedman J, Schady N: Macroeconomic Shocks and Infant Survival: A Global Perspective. Paper presented at the 2006 Population Association of America: 2006 2006.Google Scholar
- Lee CH: Regional Inequalities in Infant Mortality in Britain, 1861–1971: Patterns and Hypotheses. Population Studies 1991,45(1):55-65. 10.1080/0032472031000145086View ArticlePubMedGoogle Scholar
- Low A, Low A: Health inequalities under New Labour: relative rather than absolute gaps are important over time and place. BMJ 2005,330(7506):1507. author reply 1507–1508 10.1136/bmj.330.7506.1507View ArticlePubMedPubMed CentralGoogle Scholar
- Low A, Low A: Importance of relative measures in policy on health inequalities. BMJ 2006,332(7547):967-969. 10.1136/bmj.332.7547.967View ArticlePubMedPubMed CentralGoogle Scholar
- Singh GK, Kogan MD: Persistent socioeconomic disparities in infant, neonatal, and postneonatal mortality rates in the United States, 1969–2001. Pediatrics 2007,119(4):e928-939. 10.1542/peds.2005-2181View ArticlePubMedGoogle Scholar
- Wilmoth JR, Andreev K, Jdanov D, Glei DA, Boe C, Bubenheim M, Philipov D, Shkolnikov V, Vachon P: Methods Protocol for the Human Mortality Database.2007. [http://www.mortality.org/Public/Docs/MethodsProtocol.pdf]Google Scholar
- Maddison A: Monitoring the World Economy. Paris: OECD Development Centre; 1995.Google Scholar
- Maddison A: The World Economy: a Millennial Perspective. Paris: OECD Development Centre; 2001.View ArticleGoogle Scholar
- Maddison A: The World Economy: Historical Statistics. Paris: OECD Development Centre; 2001.View ArticleGoogle Scholar
- Maddison A: Contours of the World Economy, 1–2030 AD: Essays in Macroeconomic History. Oxford and New York City: Oxford University Press; 2007.Google Scholar
- Maddison A: Background Note on "Historical Statistics" in.2007. [http://www.ggdc.net/maddison]Google Scholar
- Box GEP, Cox DR: An Analysis of Transformations. Journal of the Royal Statistical Society Series B-Statistical Methodology 1964,26(2):211-252.Google Scholar
- Greene W: Econometric Analysis. New York: MacMillan; 2000.Google Scholar
- Bloom DE, Canning D: Commentary: The Preston Curve 30 years on: still sparking filres. Int J Epidemiol 2007,36(3):498-499. 10.1093/ije/dym079View ArticlePubMedGoogle Scholar
- Ruhm CJ: Macroeconomic Conditions, Health, and Mortality. In The Elgar Companion to Health Economics. Edited by: Jones A. Cheltenham: Edward Elgar Press; 2006:5-16.Google Scholar
- Ruhm CJ: Are Recessions Good for Your Health? Quarterly Journal of Economics 2000,115(2):617-650. 10.1162/003355300554872View ArticleGoogle Scholar
- Bishai D, Opuni M, Poon A: Does the level of infant mortality affect the rate of decline? Time series data from 21 countries. Econ Hum Biol 2007,5(1):74-81. 10.1016/j.ehb.2006.10.003View ArticlePubMedGoogle Scholar
- Bishai DM: Infant mortality time series are random walks with drift: are they cointegrated with socioeconomic variables? Health Econ 1995,4(3):157-167. 10.1002/hec.4730040302View ArticlePubMedGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.