There have been heroic efforts to produce thorough, high-quality, publicly available climate data, which has provided new opportunities for researchers interested in climate demography. The wealth of climate resources we now have provide solid, well-organized, and fine-grained data on temperatures and rainfall and how these conditions vary over time. What is more challenging is getting the personal histories from surveys to line up with those data.
If researchers are interested in the relationship between climate and reproductive issues such as contraceptive use, sexual activity, and fertility preferences, data analysis can be relatively straightforward. Questions answered at the time of the survey give us this current information. We know where respondents live at the time of the survey and when the survey took place, so space and time are easily addressed.
If we want to unravel how climate might influence reproductive events, such as pregnancies, we need to look back across time at what happened, when, and where. This post addresses some common limitations and issues that can be tricky to work with when aligning climate and reproductive health data longitudinally.
Some of our focus below is on analyzing pregnancy history data collected prior to DHS Phase 8. For surveys from DHS Phase 8 forward (released after 2023), when pregnancy histories were routinely collected, analyzing reproductive events over time may be more straightforward. But any analysis of reproductive events using longitudinal data must grapple with issues related to migration, defining the period at risk of fetal loss, and making sound choices about covariates, all topics discussed in this post.
Why look at reproductive histories and climate?
Research at the intersection of climate and reproductive health has already demonstrated important and concerning issues. In general, we know that rainfall and temperature conditions influence birth rates.1–3 Underlying this finding, however, are many distinct processes we are only beginning to understand. Weather can influence whether pregnancies are carried to term4 and increase the risk of miscarriage.5,6 Weather may also affect whether pregnancies happen, as we know thermal stress from heat can reduce fecundity.1,7
Beyond these direct, immediate effects, climate and weather events may have longer-term, indirect influences on reproductive health and outcomes. Food insecurity and infectious disease prevalence are both influenced by heat and rainfall and can impact women’s reproductive health.5,8,9 Birth spacing can shift depending on how breastfeeding is impacted by weather.10
If we consider individual agency, climate and weather may change childbearing preferences, including ideal family size.11 When agricultural production or labor needs shift in response to weather and climate, resources and income are affected, which can shape the demand for children.12,13 Libido may decline in response to hotter temperatures.14,15 Adaptation strategies to weather may also increase physical labor, which can reduce short-term fecundity.16–19
In sum, the pathways through which reproductive histories are affected by weather and climate are remarkably varied, and we have only begun to map these pathways.
What we need
When looking at factors that influence events over a reproductive life course, we can think in terms of time to an event, where we need to fill in the blanks before the event. The event of interest could be menarche, conception, a pregnancy outcome, or menopause. Along with having information about past pregnancies and births, for example, we need to know where women lived and, consequently, what they were exposed to, during the time preceding these events.
Getting the exposure correct is key to systematic and rigorous longitudinal analysis. We must ensure we are aligning individual and climate/weather data at the exact right moment in time and space. This alignment is generally facilitated through GPS location information provided at the level of the community at the time of the survey. But what happens if someone has not been living in that community long? Because we cannot correctly assign climate data before an individual moved to their location at the survey date, we must know something about the person’s residential history and adjust our sample accordingly.20
Note that “exact” is relative. We do not need exact home addresses; rather, we need an idea of the environmental conditions around where people lived and worked. The spatial detail around this information varies according to the population-environment process we are investigating, just as the detail in the health survey data varies.
Besides having GPS location and knowing length of time at that location, we need full reproductive histories. Ideally, we want information about pregnancies as well as about the timing of live births. Knowing when a pregnancy began and ended (even if not in a live birth), as well as the nature of the ending, is essential for sorting out the different factors that link climate change and reproductive health.
What is available in DHS
The Demographic and Health Surveys (DHS) remain the best source of individual- and household-level health data for poor and middle-income countries. The scope of the DHS program and the consistency of this survey enables standardization of results, which is key for comparing and generalizing findings.
In addition, The DHS Program has made some key decisions that increased the surveys’ usefulness to climate and health researchers. First, DHS began providing GPS coordinates that allow researchers to match climate measures to low-level locations (e.g., the approximate site of a village where the sample was drawn). Second, the survey began to collect information on how long individuals have resided at their current location. Many of us dream of full migration histories being collected by DHS, which would help us deal with bias caused by climate migration. At the moment, we have to assume that those who stayed were least affected by past weather events and climate shifts. Nevertheless, with information about the person’s length of time at their current residence, we can at least be sure we are correctly assigning the contextual measures to a location for the right amount of time.
These two pieces of information—GPS coordinates for the primary sampling unit and respondents’ length of time at their current residence—are not available in all DHS datasets. Although most recent data sets have this information, researchers must check whether older samples have variables such as V104 (RESIDEINTYR in IPUMS DHS), which reports how many years a woman has lived in her place of residence.
The narrowest time intervals available in DHS data are generally months and years. For example, we have the month and year of birth for all women of childbearing age and for their offspring. We also have information on time of moving to the current location and on the date of the woman’s first cohabiting union. For data on reproductive events, DHS surveys conform to one of two types:
Live birth histories: The year and month of each live birth to a woman is collected.
This information is recorded regardless of whether the child is still alive and whether the child lives with the mother. Although the birth history collection strategy misses pregnancies that do not lead to a live birth, it is superior to the common strategy in surveys and census sources of only documenting the ages of children in the household. Most DHS surveys from Phases 1 through 7 only collected birth histories from women of childbearing age.
Pregnancy histories: The year and month of every pregnancy outcome is collected.
Pregnancy histories are more complete than live birth histories because they also capture any known conception that ends in a miscarriage, abortion, or stillbirth. Whether the specific outcomes of miscarriage, abortion or stillbirth are distinguished varies across these samples, with abortions and miscarriages commonly combined into one category.
These two types of histories relate to the entire life of the woman. Until recently, fewer than 25 DHS surveys collected pregnancy histories. For DHS Phase 8 (with data distributed around 2023 or later), a pregnancy history rather than just a birth history was included the standard questionnaire that serves as the starting point for each country’s survey.
Besides the birth history or pregnancy history data, the DHS often collects monthly reproductive history data for women for the five years preceding a survey. The simplest 5-year reproductive calendars only report whether a woman experienced a birth, pregnancy, or pregnancy termination in each month. Most reproductive calendars also collect monthly information on family planning, including the type of contraception used and, often, the reason for its discontinuation. Combined with other information that situates the last five years in a woman’s reproductive history, the calendar data can be a rich resource. IPUMS DHS has made it easier to analyse calendar data, through providing integrated variables accessed by choosing “Woman month” as the unit of analysis.
But for maximum statistical power, which quickly becomes important when working with covariates and splitting the data into comparison groups, and to avoid left truncation of data (not knowing when contraceptive use or a pregnancy began), full birth or pregnancy histories are often preferable. The rest of this post focuses on strategies for working with full reproductive histories.
Making it work
Defining the window of observation
The first step is establishing the window of observation, that is, the time period when we know a woman is living at her current location. In the absence of detailed migration histories, we can begin this observation period with the exact month and year she moved to her current place of residence (when such detail is available). Alternatively, we can choose a standard reference period, such as the first January following the year she moved to her location. Since the most commonly available DHS migration variable covers “Years lived in place of residence” (V104/RESIDEINTYR), this last approach is often the most practical. Although we do not analyse births or pregnancies before that starting point, these unobserved events are used when creating information about the woman, such as parity or years since last birth/pregnancy outcome. For women who reported “always” living in their current location, the full reproductive history can be analyzed (i.e., events from age 15 until the time of interview).
Once the window of observation is constructed for each individual, and the sample selected to include only those observations, then climate data can be merged on to the unit of analysis (i.e., woman or woman month).
Defining the timing of a pregnancy
A particularly salient period during which climate and environment can affect reproduction is the time between conception and the end of a pregnancy. When working with live births, including data from the widely available DHS birth histories, the conception month is usually calculated by subtracting 9 months from the date of birth.
When working with pregnancy history data, the situation is sometimes more complicated. DHS samples with pregnancy histories always report the year and month a pregnancy ended. For this blog post, we focus on the 22 standard DHS surveys that collected pregnancy history data before (the current) Phase 8 survey. The Phase 8 approach is very different than the pre-Phase 8 approach and as a result data analysis approaches will need to be different. Strategies to use both the pre-phase 8 and the phase 8 data are useful to allow for longer-term investigations of climate impacts on pregnancy.
For the pre-Phase 8 data, the variable names relating to pregnancy history are sample-specific. In general, country- and year-specific variable names in DHS women’s surveys begin with the letter s followed by the number of the relevant question on the questionnaire. Thus, for example, on the women’s questionnaire for the 2012-2013 Pakistan sample, question number 226 records the month and year that each pregnancy ended. Accordingly, variables s226m_XX and s226y_XX respectively report “Month of pregnancy lost” and “Year of pregnancy lost,” with XX referencing the number of the pregnancy, from the most recent (01) to the twentieth most recent.
Figuring out the date of conception for pregnancies in pre-Phase 8 pregnancy histories can be difficult. Some of these surveys, such as the Pakistan 2012-13 sample, directly asked how long the pregnancy lasted. For many of the samples with early pregnancy histories, particularly from Eastern Europe and East Asia, this direct information about pregnancy duration (and thus conception date) is lacking.
In such cases, establishing the month of conception based on the month of miscarriage, abortion or stillbirth requires imputation. Knowing the month of conception is required for investigating climate impacts on pregnancy outcomes. One approach uses information derived from the five-year calendar data.21 In those data one can observe the month a woman said she was pregnant and the month when the pregnancy ended. By taking the difference, we can get an estimate of how long pregnancies lasted on average before the specific outcome in a specific culture. We can also calculate the mean and standard deviation for pregnancy outcomes for a given calendar sample and use those statistics as a structure for imputation. Note that reports from the calendar data may be sensitive to context—for example, the extent to which stillbirths are called miscarriages and vice versa. In addition, the length of gestation before abortion may vary by abortion legislation, which obviously varies across time and space. Thus, no clear rules of thumb, like “pregnancies ending in a live birth usually last 9 months,” can be stated for pregnancies ending in miscarriage, stillbirth, or abortion. Such estimates for terminated pregnancies should be context-specific (e.g., access to pregnancy tests) and empirically grounded.
Results from calendar-based imputed pregnancy duration data for one study are shown in Table 1, which details the pregnancy outcome timing from Armenian and Tajik DHS data.21 Only 65% of reported pregnancies ended in a live birth in Armenia during the five years preceding the survey (2015-16) and 84% in Tajikistan (2017). To impute the timing of conceptions ending in stillbirth, miscarriage, or abortion, the authors applied a normal (Gaussian) distribution using the mean and standard deviation observed in the five-year calendar data within the specific DHS data set.
| Live Birth | Stillbirth | Miscarriage | Abortion | |
|---|---|---|---|---|
| Armenia, 2015-2016 | ||||
| Mean | 8.56 | 6.77 | 2.57 | 2.24 |
| Median | 9 | 7 | 2 | 2 |
| SD | 1.35 | 1.89 | 0.88 | 0.8 |
| Min. | 2 | 2 | 1 | 1 |
| Max. | 10 | 9 | 7 | 9 |
| Number of pregnancies | 2,084 | 17 | 305 | 783 |
| Tajikistan, 2017 | ||||
| Mean | 8.63 | 6.61 | 2.22 | 1.9 |
| Median | 9 | 7 | 2 | 2 |
| SD | 1.28 | 1.93 | 1.18 | 1.02 |
| Min. | 1 | 2 | 1 | 1 |
| Max. | 10 | 9 | 9 | 9 |
| Number of pregnancies | 6,890 | 70 | 538 | 715 |
Selecting appropriate covariates
A final consideration when working with retrospective longitudinal data is which variables can be used as covariates. Unless information comes in the form of histories, where we can track changes in a characteristic for an individual over time, we should only use information that does not change. Examples of non-changing variables are birth cohort and ethnicity/nationality.
The DHS provides a wealth of household and individual-level information that is unfortunately problematic when analyzing climate effects over time. This is generally unsatisfying, since we know socioeconomic factors play an important role in health outcomes. But ascribing back characteristics—such as household wealth quintile, woman’s employment status, her contraceptive knowledge, or her decision-making power at the time the survey—to the respondent earlier in time is risking a few problems. First, such ascription introduces error into the data, whereby a characteristic is assigned at a time for which it may not be true. Although it may seem relevant that the person eventually ended up with that characteristic, assigning the characteristic before it is true introduces estimation inconsistencies due to “anticipatory analyses”.22
While assuming that current characteristics held earlier is never completely reliable, error could be minimized by restricting analysis to recent pregnancies—for example, those occurring in the three years before the survey or focusing on only the most-recent and second-most-recent pregnancies (indexed _01 and _02), if they occurred recently.
One option for gaining some measure of socioeconomic status that varies over time is to construct time-varying covariate variables. For example, from the variable that tells us the woman’s completed number of years of education (V133, or EDYRTOTAL in IPUMS DHS), we can produce a reasonable estimate of her last year in school. We can code the following years according to the highest level of education the respondent achieved. Similarly, we do not have partnership histories in DHS, but we do know the woman’s age at her first marriage or coresident union (V511, or AGEFRSTMAR in IPUMS DHS). From this information, we can construct a time-varying covariate for whether the woman had ever entered a co-resident union.
Conclusion
Some of the pathways by which climate change affects health relate to reproductive events such as pregnancies and births. Analyzing such events over the life of a woman demands care on the part of researchers. One must, for example, “define the window of observation,” by excluding reproductive events that occurred before a woman migrated to her survey location. One must choose the type of longitudinal reproductive data: calendar data from the previous five years, birth history data, or pregnancy history data from across the woman’s life. Pregnancy history data in the IR (women’s) files from before and after DHS Phase 8 differ in variable naming conventions and scope. While dates for the end of pregnancies are consistently known, determining month of conception may need to rely on data imputation for some pre-Phase 8 samples. Covariates should reflect time-invariant characteristics, be restricted unless analyzing very recent pregnancies, or be constructed to reflect variation over time (e.g., in educational attainment).
Climate change and health research yields valid conclusions only when we avoid glaring errors—for example, assuming that environmental conditions were the same in a woman’s previous location as in her current residence, or that current characteristics (e.g., employment, household wealth) were static across time. We hope this post helps researchers avoid such errors and guides them in solving problems that arise with retrospective longitudinal reproductive histories.
Getting Help
Questions or comments? Check out the IPUMS User Forum or reach out to IPUMS User Support at ipums@umn.edu.