Skip to content
 

Economic growth -> healthy kids?

Joe Cummins writes:

Anaka Aiyar and I have a new working paper on economic growth and child health. Any comments from you or your readers would be much appreciated.

In terms of subject matter, it fits in pretty nicely with the Demography discussions on the blog (Deaton/Case, age adjustment, interpreting population level changes in meaningful ways). And methodologically we were concerned about a lot of the problems that have been discussed on the blog: the abuse of p-values; trying to take measurement seriously; the value of replication and reanalysis of various forms; and attempting to visually display complex data in useful ways. There is even a bit of the Secret Weapon in Figure 2. In general, we hope that we built a convincing statistical argument that our estimates are more informative, interpretable and useful than previous estimates.

Would love to hear what your readers do and don’t find interesting or useful (and of course, if we messed something up, we want to know that too!).

Replication files are here.

Here’s their abstract:

For the last several years, there has been a debate in the academic literature regarding the association between economic growth and child health in under-developed countries, with many arguing the association is strong and robust and several new papers arguing the association is weak or nonexistent. Focusing on child growth faltering as a process that unfolds over the first several years of life, we provide new evidence tracing out the relationship between macroeconomic trends and the trajectory of child growth through age 5. Using two novel regression models that each harness different kinds of within- and between-country variation, and data on over 600,000 children from 38 countries over more than 20 years, our estimates of the association are relatively small but precise, and are consistent across both estimators. We estimate that a 10% increase in GDP around the time of a child’s birth is associated with a decrease in the rate of loss of HAZ of about 0.002 SD per month over the first two years of life, which generates a cumulative effect of around 0.04 SD by age 3 that then persists through age 5. Our estimates are small compared to most previously published statistically significant estimates, more precisely estimated than previous insignificant estimates, and relate to a broader population of children than previous estimates focused on dichotomous outcomes.

It’s a frustrating thing that this sort of careful, policy-relevant work (I have’t read the paper carefully so I can’t comment on the quality of the analysis, one way or another, but it certainly seems careful and policy-relevant) doesn’t get so much attention compared to headline-bait like pizzagate or himmicanes or gay genes or whatever. And I’m part of this! A careful quantitative analysis . . . what can I say about that? Not much, without doing a bunch of work.

But at least I’m posting on this, so I hope some of you who work in this area will take a look and offer your thoughts.

26 Comments

  1. D Kane says:

    Is this really “policy-relevant work?” Policy folks generally (always?) want causal information. If they spend $100 million on X, then children’s health improves by Y. That information allows them to make a better decision about whether or not to spend $100 million on X.

    But there is no policy here, no choice a policymaker can make, no manipulation to be performed.

    The “association between economic growth and child health” is an interest topic and worth studying. Learning about the facts of the world is cool! But there is nothing (directly) related to policy here since “economic growth” is not something that anyone knows how to (directly) manipulate.

    I am not complaining about the post. Indeed, I like it when you bring high quality work to our attention. I just want to make sure I understand what you mean by “policy-relevant.”

    • Andrew says:

      D:

      Indirectly policy-relevant, I think. Understanding the effects of existing factors can inform policy. Further thoughts on this general point here.

      • D Kane says:

        Then what would be an example of a paper discussed on this blog that is not “policy-relevant?”

        Again, I don’t deny that this paper teaches us something about the world and that, the more we know about the world, then the better our policymaking will be. I was just trying to understand why you would call this paper, relative to the hundreds of other papers you discuss, “policy-relevant,” when, if anything, it is probably less policy-relevant than many (most?) of the papers mentioned here, precisely because it lacks a lever for policymakers to pull.

        • Andrew says:

          D,

          An example of my own collaborative research that is indirectly policy relevant is our estimates of the effects of redistricting. An example that to me does not seem particularly policy relevant is Red State Blue State.

          • D Kane says:

            Red State, Blue State, among other things, provides “the basic facts on income and voting.” That sounds fairly policy-relevant to me, at least for anyone who wants to win elections, pass state ballots, supervise redistricting, et cetera.

            Again, my only claim is that, if “Age-Profile Estimates of the Relationship Between Economic Growth and Child Health” is “policy-relevant,” then everything discussed on this blog is “policy-relevant.” And that is OK! I love the blog.

            • Dustin says:

              From someone who works on policy stuff with most of my time, I’d consider this paper policy relevant, as it provides a better estimate of an outcome of interest that could be used in the input output models that are commonly used when deciding on which projects development banks invest in. Moreover, in attempting to forecast a variety of outcomes, it also seems like it would be a valuable estimate to have (given expected GDP growth, how much do we expect health outcomes to improve, and what’s the gap between where we want to see health outcome and where we project it will be. Follow up question, how much needs to be invested to close this gap given other good estimates of the effect of different interventions). From what I’ve seen at least, the numbers that are usually used for this type of modeling are way off. Hence, providing better estimates as this paper attempts to do seems highly relevant to improving policy decision making to me at least.

    • I’d say it’s policy relevant in the following way, the paper shows that GDP and growth in GDP both are associated with growth in a consistent way. The paper posits a mechanism, which is in part access to public goods such as markets for exchanges, health information, etc. People interested in affecting child health should who know this association is a true fact about the world should then investigate the actual mechanism that GDP is proxying for. It may be much easier to increase X which is some set of policy relevant variables associated with GDP than to increase GDP across the board for a whole country.

      Simply knowing that the world is a certain way affects your choice of what kind of experiments to do to understand why.

      So, by that measure I think it’s really important to know what is really going on, and that knowing what is going on is in fact policy relevant.

      So much of the bad stats/science we have seen on this blog are directed at explaining the mechanism of stuff that might not even be a thing (power pose? beauty and sex ratios, etc…) lots of people just claim the world is a certain way totally without merit.

      • Keith O'Rourke says:

        > really important to know what is really going on, and that knowing what is going on is in fact policy relevant.
        Good point.

        It really bothered me when the editor of the Canadian Medical Assoc Journal refused to consider our paper on prospective tracking of patients from initial concern by their GP to initial treatment for cancer as it had no comparison group. They said that it could not be research without a control comparison.

        Fortunately it was published in an American journal so that the information about the Canadian health system was made available – Diagnostic assessment of suspected cancer: Prospective cohort study of diagnostic delay intervals across three disease sites.

  2. Ed Hagen says:

    Any thought of controlling for some measure of within-household consumer/producer ratio? There is evidence in subsistence economies that, within households, more mouths to feed is negatively associated with child growth.

    I’m pretty sure the DHS has measures of household composition.

    • Joe Cummins says:

      Ed,

      We could certainly disaggregate by parity and repeat the analysis by birth order (at least for the first few children). Or we could even go Full Gelman and model some sort of (semi-)parametric interaction allowing each country to have its own HAZ-age profile for each birth order, and then allow the effect of GDP to vary in that dimension too.

      I’ve actually gone back and forth with Anaka and my own brain about whether we should do a bunch of disaggregations like that. I do think such work is interesting and important, but eventually decided this paper was long enough already and that any such disaggregated results probably deserved enough attention to be done separately. But part of what I hoped this paper would do is to provide people a new analytic-theoretical-metastatistical framework for thinking about how to analyze these kinds of effects. So hopefully someone (me, Anaka, someone else) will do the analysis across birth parity some day and give it the attention it deserves.

      • Joe Cummins says:

        … or if we wanted to do a better job of answering the exact question you asked, we could interact by both birth order and household size or number of children in the household (or maybe even household size by wealth quintile). But then we’d need to add some real structure of some sort, either with strong parametric forms for the interactions or some set of methods that are really good at partially pooling information and regularizing all the coefficients in some way…at some point even 700,000 observations gets sparse!

        • Ed Hagen says:

          OK, thanks. My basic thought is that there might be a couple of confounds/mechanisms: bigger households and shorter interbirth intervals in low GDP countries and smaller households and larger interbirth intervals in high GDP countries that go along with the fertility transition. So what you really might be seeing is a change in household size and interbirth intervals, and the negative effects of siblings, perhaps especially younger siblings, on child growth. Here is a recent paper that reviews some of this stuff:

          http://journals.plos.org/plosone/article?id=10.1371/journal.pone.0150126

          I don’t necessarily agree with the authors that the “biological significance of these results is negligible”, given the short timeframe they examined and the their -2SD criteria, but the paper will give you a start if you’re interested in this angle.

          • Joe Cummins says:

            Ed,

            Thanks for that link. I’ve been interested in parity/family size issues for a while, particularly regarding the debate on the India/Africa HAZ difference. Certainly differential fertility patterns (and the associated changes in household resource distribution) could be part of what is going on. In one sense, my argument to Daniel sort of applies here – if economic growth generates changes in fertility patterns that then generate improvements in health, then that is part of the effect of economic growth (or at least part of the relationship that we want to measure and include in the overall effect). But that kind of “mechanism” is certainly not what we model in our theory section (which relates to household budget constraint and public goods availability) and we should probably at least try to test it empirically…it shouldn’t be too hard. Maybe a first step is to include more household demographics in the regressions, but a second step might be to estimate differential effects across, say, average household size or number of children in the household (aggregated at the country level). I’ll have to think a bit more about it, but thanks for pushing me on the idea…I had probably dismissed it a little too quickly.

  3. Other thoughts I have: I’d really like to see “GDP” rewritten as “log_10(GDP)” in the graphs, that’d make the graphs much more quickly interpretable without having to read the text down to page 15 or so to figure out how “GDP” is defined.

    I’d also like to see a dimensionless ratio of the height other than HAZ used in addition to the HAZ. In particular I’d like to see Height(t) / RefHeight(2) used, where Height is the actual height of the child and RefHeight(2) is the average height of children in the WHO reference population at age 2.

    My preference for this is that it is more biologically interpretable, instead I think we’re looking at basically Height(t) / sd(ReferenceHeightData(t))

    and obviously the sd of the reference height population is changing in time, and is related to a bunch of factors outside of simply growth (such as variation across races, variation across geography, variation across cultural choices of diet, etc)

    Finally, I have thoughts about recasting this whole thing in terms of an ordinary differential equation for growth with a time-varying forcing function, but I’ll reserve that for if I can get an invite to Joe’s house in the near future ;-)

    • Joe Cummins says:

      Hey Daniel,

      1. We will fix the graph labels, thanks.

      2. You aren’t the first person to suggest alternative units, and that probably means I should think more about it. But I do think there are a lot of good reasons to focus on HAZ. First, just about everyone else does, and that makes effect sizes more comparable to other estimates out there, and the field has a good sense of general effect sizes when HAZ is the outcome. Most of the rest of my arguments also apply to other normalizations, so your suggestion also works, but I do think that using well-understood units as the primary outcome makes a lot of sense. Quick note though: the sd of the reference population is NOT changing over time, in the sense that it is fixed by when the WHO gathered their reference norm data – perhaps the “right” standards would, if secular trends are strong, but the standards the WHO provides don’t change over time.

      3. Look forward to hearing your differential equations pitch over coffee or cocktails soon.

      • I meant sd changing by age not that a given age sd changes as new WHO data is collected. For example undoubtedly there is a smaller variation in height at birth compared to variation at 3 years yes?

        I think HAZ is important to use due to the reasons you mention, I just think that an alternative unit with a constant by age denominator should be presented as well.

        • Joe Cummins says:

          “undoubtedly there is a smaller variation in height at birth compared to variation at 3 years yes?”

          Yes, of course. But isn’t that sorta desireable here? I mean, we don’t want to use centimeters right – 1 cm at birth means something much different than 1cm at age 4. The use of a changing SD allows you to ask about where these children fit in the distribution of health kids… I’m inclined to think that being at the 10th percentile of the WHO reference at age 1 and age 4 means those children are comparable in some sense. That is obviously a problematic thing to say (rank shifting happens all the time, and it isn’t clear that rank relative to WHO pop means the same thing at each age), but I’m struggling to see how “fraction of median child’s height” really helps that much… if variation in the population is increasing with age, isn’t that a part of biology we’d want to keep in our measure?

          What is it specifically you want as a property of a measurement that HAZ doesn’t have and Height/Median_Height does have? Or you just want another measure to make sure it isn’t a measurement thing?

          • HAZ is a nonlinear transformation of growth that makes it easy to answer one specific comparison question. If that’s the only question you have then it’s fine by itself, but if you have questions like how much energy towards growth deficit there is, or something, then 1 unit of haz doesn’t represent a biologically interpretable unit of growth. After all, if you are a median first world child your HAZ stays constant at 0.

            So I think displaying a unit that directly linearly relates to physical quantities in addition to HAZ helps inform a wider range of questions. I’m not saying HAZ is bad, just limited purely to a typical economists viewpoint ;-)

            • It’s particularly true that if you want to model growth as a function of various biological inputs, and then model inputs as a function of GDP, and GDP growth rate, then you wouldn’t write that model in terms of HAZ. And that’s precisely what I think the next step should be.

            • Joe Cummins says:

              Good points. I’m not sure if/how I’ll implement them, but at the very least I’ll try to add some discussion about alternative measurement possibilities when I discuss how to interpret HAZ. Thanks for the time and effort – I really do appreciate it.

  4. Joe Cummins says:

    Andrew – thanks for posting this. Appreciate the kind words, but more than that appreciate the opportunity to get feedback from the best comment section on the internet.

    I’ve been running around a bit this morning, but I will read through comments and respond as best I can in a couple of hours. Thanks to everyone who has commented so far, and of course Anaka and I would appreciate any thoughts, comments, criticisms or mistakes you all find/have found. Thanks for offering us your time (especially since the only reward is helping us make the paper better).

    • Here’s the first question I have, is the effect of economic growth any different from the effect expected because of GDP changes integrated throughout the duration of the measurement period? (say 2 years)?

      For example, suppose that you recognize that there is growth, then you may choose to do things differently now knowing that later you’ll have more resources. On the other hand, in the absence of knowledge of growth, you’re always doing thing now based only on resources now…

      In the first case, there are “second order” effects, whereas in the second case it’s just a changing value of “now” that’s constantly altering what’s going on now.

      • Joe Cummins says:

        Daniel,

        Agreed that expectations regarding future resource/income streams could certainly affect a household’s optimal health input streams now. But that mostly seems to me part of the “causal effect” of growth, right? At least in the sense that you couldn’t imagine having an effect of economic growth on child health without also getting the corresponding effect on people’s expectations, so even though those might be two different mechanisms, they are both part of the effect GDP growth. It isn’t immediately clear to me why we’d want to separate out the effects unless we were really trying to use the GDP analysis to say something about how other changes to resource availability might affect child health… but maybe I’m missing something (or just wrong).

        • I think it would be useful to do the calculation for a “future blind” model and compare to reality because it would tell us something about the mechanism of how decision making is being done, and then that could inform things like policy interventions. For example policy that reduces future risk for parents would have no effect if people are just making short term decisions. I do suspect that future expectations would play a bigger role for general decisions at higher GDP levels, but probably the result is negligible on HAZ because at higher GDP you’re closer to the reference group. For low income situations if future expectations play an important mechanistic role, that seems inherently interesting

Leave a Reply