Stuff you cannot randomize...


This page in:

I have been thinking about marriage recently. No, not about my own marital status, but marriage among school-age girls and its effects on future outcomes… While many arguments are made to curb teen marriages (and pregnancies), it is not clear whether these events themselves are the cause of poor future outcomes or they are simply correlated with other background characteristics that are prognostic of future outcomes. A brief survey of the literature indeed suggests that the evidence is mixed; especially when it comes to the effects of teen childbearing on future outcomes.

The first thing to note about the literature on the effects of teen marriage on future outcomes, such as, say, poverty status, is that the implicit channel that the researchers have in mind is almost always stunted educational attainment as a result of the marriage decision. In this sense, delayed marriage and increased schooling are treated as two sides of the same coin: if you are in school then you are not married; and if you’re married then you’re not in school. This is true in many cases, especially in Sub-Saharan Africa: marriage and schooling are mutually exclusive.

However, it is not true that if you’re not married then you’re in school. Findings from a recent study by Baird, McIntosh, and Özler pose somewhat of a policy conundrum. An intervention that reduced marriage rates by nearly 50% among adolescent girls in Malawi barely improved enrollment rates and had no effect on learning. On the other hand, while it is true that if you deter an adolescent schoolgirl from dropping out of school then you will have a good chance of delaying her marriage, because many programs are only marginally effective in improving school enrollment and only a portion of dropouts actually get married as teens, the effects of programs intended to improve schooling can be small on delaying marriage. Hence, it may be best to decouple these two policy objectives and go after them using different levers tailored to each.


Let’s look at some of the literature below -- not only to find causal evidence of the decision to get married early on future outcomes, but also to look for these impacts to plausibly go through channels other than increased education.

Dahl (2010) exploits the time and state variation in the legal age of marriage laws in the U.S. to examine the impact of early teen marriage on likelihood of being poor as an adult and finds very large impacts. Girls married before the age of fifteen are 31 percentage points (pp) more likely to be living in poverty than those who got married later. Field and Ambrus (2008), using age of menarche as an instrument for age at first marriage in Bangladesh, find that delayed marriage is also associated with increased use of preventive health services during pregnancy, has an effect on reducing the likelihood that women are prevented from leaving the village for work and increasing the likelihood that they participate in major household decisions. They argue that these are age effects, rather than the effect of increased years of education. Hence, increased age at first marriage may improve health and agency (or intra-household bargaining power) through channels other than education.

In these papers and others (e.g. Goldin and Katz 2002; Jensen and Thornton 2003), there is a common thread: teenagers are making a mistake they might regret later, their capacity to give consent is in question, their preferences may not be fully formed, or there may be incomplete altruism on the part of the parents. But, protecting people from themselves or their parents is not always straightforward, so it helps to have some credible evidence that exogenous variation on age at first marriage is linked with important future outcomes for adolescent girls: the evidence from Field and Ambrus suggests that delaying marriage is a (surely imperfect) substitute to increased education when it comes to women’s mobility and their ability to influence household decisions.

Even then the evidence cited above come from the US and Bangladesh, while I am interested in Sub-Saharan Africa. The debate on whether marriage is super risky (see slide 2 on this presentation), or protective (see this paper by Bongaarts) against HIV infection is one of the most frustrating ones I know, mainly because credible evidence is hard to come by. With the emotions running as high as they do in the HIV/AIDS debate, advocates just pick and choose from observational studies to argue their side. Whoever produces credible information on (a) whether and under what circumstances marriage can be protective or dangerous for HIV infection, and (b) suggest what kinds of programs we can employ to intervene will have done a great service to the field.

Which brings me to "speed dating as an instrument for marriage", as promised at the end of my last blog (note to self: do not promise anything for next week’s blog). The problem with the identification strategies used in the papers above is that they will not work to isolate the impact of delaying marriage on HIV (or other outcomes for that matter) in sub-Saharan Africa. I suspect that district/time variation in age of marriage laws are out (please correct me if I am wrong). Age of menarche is also likely to be correlated with age at first sex, so will not isolate the effect of delaying marriage on HIV (I suspect premarital sex and teen pregnancy outside of marriage were not highly relevant factors in the Bangladesh setting). Providing uniforms or giving cash transfers (unconditional, or conditional on schooling) do not satisfy exclusion restrictions. Neither do cash transfers (or providing cooking oil as in this project) conditional on not getting married. So, what then?

One of my colleagues joked that you could do an encouragement design for using dating websites. While that may be OK for adults who live in the West, it’s not going to work for teenage girls in sub-Saharan Africa. Forgetting the issue of access to such websites for a second, it’s simply not desirable to encourage teens to get married: what we ideally need is a deterrent rather than an encouragement that will affect nothing else that is prognostic of future risk of HIV infection. Barring that, I am reduced to dreaming about conducting speed dating events at churches or community centers for girls above a certain age (such as legal age of marriage)…

Can you think of a solid identification strategy for this question? Even if a handful of grad students around the world are now interested in exploring this question carefully, the few hours I spent on putting this post together will have been worth it. Of course, always check with your thesis advisor first!


Berk Ozler

Lead Economist, Development Research Group, World Bank

Join the Conversation

April 20, 2011

Sorry, I'm new to RCTs so I perhaps don't quite understand the meaning of an `exclusion restriction' - so perhaps treat this question as a demand for education! Why wouldn't conditional cash transfers work? Couldn't you give the control an unconditional transfer and the treatment a transfer conditional on not getting married? And then compare the subsamples of people that didn't get married in each?

Berk Özler
April 20, 2011

The Baird et. al. study cited in the original post is an experiment exactly like that with the conditionality being on attending school regularly instead of marriage. The problem is that the condition affects many other things than just school enrollment. So, you would not be exactly identifying the causal effect of schooling on an outcome of interest by comparing that outcome in the CCT and the UCT groups, but rather the effect of the condition that someone not get married for a while: the two are different.

April 20, 2011

The link may be very weak but I'm wondering if you can do something with the bride price. Assume the bride price is two cows for a woman, the likelihood of getting married will be lower the years cows are expensive. It's a kind of natural experiment. Data would be hard to get and it does not work if bride price is adjusted when the price of cows varies (but I don't think it's the case).

Good luck, that's an interesting topic.