Slide 1The last video was about describing what exists. This video is about two more learning goals: explaining why something happens, and predicting what happens next.

Slide 2Description is essential to science and to decision-making related to needs and resources. The result suggests that more than half of married women of reproductive age were not using a modern method of contraception in 2010. This is useful to know if you work for the Ministry of Health and are concerned about promoting reproductive health.

Slide 3But description only gets you so far. You might want to go the next step and ask, why is uptake lower among younger women? And, what happens if we introduce a policy or program intended to promote contraceptive use? Can we explain patterns of modern method use?

Slide 4One way to explore a why question is to ask people. That's what one research team did in Nepal, to understand the reasons for low use of modern methods of contraception among young people. They organized six group discussions and conducted 25 interviews with a diverse mix of informants from one village, including teachers, youths, health workers, religious leaders, and government officials.

Slide 5One woman explained her non-use this way. Her husband works abroad. He came home during a festival, and three months later she learned she was pregnant. She was shocked. They already had three children, two of them unplanned. She did not have enough information about contraceptive measures in that situation. Had she known about them, she says, she would have used them. Her words are on the screen.

Slide 6The authors analyzed transcripts with passages like this and reported that young people face numerous barriers to initiating family planning, including a lack of awareness as described by this woman. They concluded the article with several ideas for designing new interventions, new programs, to promote modern method use among this group.

Slide 7Another common method for exploring why and what if questions is explanatory modeling, the use of statistical models for testing causal explanations. This work also falls under the label of causal inference, which one team defines as follows: what would happen to an outcome y as a result of a treatment, intervention, or exposure z, given pre-treatment information x.

Slide 8Sometimes researchers ask these questions in the context of an experiment that features random assignment to conditions and active introduction of the treatment, intervention, or exposure. But most often, research in the social sciences examines associations between variables in non-experimental data.

Slide 9Before the examples, some context. Research in low-income countries has found that women want to prevent or space their next pregnancy after they have given birth, but 60 percent do not start a method of family planning before their fertility returns. Precisely when women become fertile after childbirth varies, and is influenced by factors like breastfeeding, but in most cases ovulation returns before family planning is started. This puts women at risk of an unwanted or mistimed pregnancy. One option to reduce this risk is to have an intrauterine device, or IUD, inserted immediately after delivery of the placenta or within the first month postpartum.

Slide 10Let's consider two examples of causal inference, both on modern method use. In the first, researchers looked at the prevalence of modern contraceptive use in a rural municipality in one of Nepal's poorest districts, before and after a pilot program was implemented. In this program, community health workers conducted home visits with pregnant women and new mothers and offered them contraceptive counseling. The authors reported that overall modern method use increased from 29 percent pre-intervention to 46 percent post-intervention. The figure breaks that out by time since delivery: from 44 to 64 percent among women in the early postpartum period, and from 16 to 30 percent among women in the late postpartum period. But as we'll discuss in a later chapter, this non-experimental design requires strong assumptions for attributing the observed change to this program.

Slide 11In our second example, a team used a stepped-wedge design to estimate the causal effect of offering postpartum family planning counseling on the proportion of new moms who opt to have an IUD inserted following childbirth. In this design, which you'll meet more formally in a later chapter, the counseling intervention was turned on in six hospitals in a stepwise fashion. The three hospitals randomly assigned to Group 1 received the intervention first, in a staggered start. Approximately six months later, the remaining hospitals began offering the counseling intervention, one after the other.

Slide 12The figure shows that rates of uptake appeared to jump in each group after the intervention was introduced. The authors estimated that the intervention increased IUD uptake by 4.4 percentage points, with a 95 percent confidence interval from 2.8 to 6.4 percentage points.

Slide 13The third common learning goal is to use data and algorithms to predict new or future observations. You might know this as the domain of artificial intelligence, machine learning, or deep learning. I'll stick with the more general term of prediction modeling.

Slide 14Prediction modeling has gone full rocket-launch in medicine. Data scientists who might have once worked hard for incremental gains in the quality of Netflix's movie recommendations are now searching for ways to teach computers to outperform radiologists in detecting pathology. Increasingly, the same tools are being applied to solving public health challenges.

Slide 15Returning to the topic of contraception, prediction models help us to address a key public health challenge. More than 1 in 3 women in low- and middle-income countries stop using contraception within 12 months of starting, putting them at risk for unintended and mistimed pregnancies. What if there was a way to accurately predict who would eventually discontinue, and offer them additional support when starting a new method?

Slide 16One team took on this question, recruiting a cohort of more than 700 Kenyan women who did not want to get pregnant and following up with them over the course of 24 weeks to see who remained on a method and who stopped. The authors used data from 75 percent of the women to develop a discontinuation risk score, based on things like method choice, education, and marital status, and reserved the remaining 25 percent to validate their new tool. They found that discontinuation was almost 2 to 6 times higher among women labeled as high risk versus low risk. If you have a quantitative itch to scratch, now is the time to jump in.

Slide 17One caution before we move on. In papers that report regression results, you will often find references to predictor variables. These are the independent variables that predict the outcome. You can use regression to predict future, or unseen, values. But regression as often applied in the social sciences is mostly just fancy averaging with cross-sectional data obtained at one time point. So: describe, explain, predict. In the next video, we turn a learning goal into a question you can actually answer.