Conflict occurrence
As we consider conflict in or the five years prior to the year of marriage, our effective period of analysis reduces slightly to 1994 to 2021. Within this period of time, our sample includes 2,317,396 marriages, 660,762 of those before age 18 (weighted incidence of child marriage of 29.2%). 238,259 of child marriages in our sample occurred at times when the cluster was within a conflict zone in the year of marriage or the five years prior to that. Across our whole sample (not controlling for any other factors, but weighted by individual sampling weight), the share of women married before age 18 was 36.5% in such conflict-affected contexts versus 26.2% in non-conflict settings.
Table 1shows the regression results for all four regressions for models (A), exploring the impact of conflict occurrence on the incidence of child marriage. We find a positive and significant increase in the probability of a girl being married before age 18 if the cluster was located within a conflict zone in or the five years prior to the year of marriage. We found that the risk of child marriage increased by 16.4 percentage points for girls affected by conflict compared to women not living in a conflict zone in regression (1). When we restrict the sample to marriages that took place in the five years preceding a survey in regression (2), the effect remains significant and positive, but the effect size decreases to 5.9 percentage points. In regression (3), considering only observations from surveys where we know that women are still living in the same place as at the time of marriage, we find that conflict occurrence increases the incidence of child marriage by 4.5 percentage points.
Table 1
Results of linear probability regression for conflict occurrence on incidence of child marriage
|
(1)
|
(2)
|
(3)
|
(4)
|
|
First marriage by age 18
|
Conflict in or five years prior year of marriage
|
.1638***
|
.0586***
|
.0451**
|
|
(.0036)
|
(.0159)
|
(.0187)
|
|
State-based armed conflict in or five years prior year of marriage
|
|
|
|
.1438***
|
|
|
|
(.0046)
|
Non-state conflict in or five years prior year of marriage
|
|
|
|
.089***
|
|
|
|
(.0072)
|
One-sided violence in or five years prior year of marriage
|
|
|
|
.0685***
|
|
|
|
(.0046)
|
Highest level of education (compared to lower than primary)
|
|
|
|
|
Primary education
|
− .0416***
|
− .0137
|
− .0328**
|
− .0414***
|
|
(.0032)
|
(.0092)
|
(.0144)
|
(.0032)
|
Secondary education
|
− .1747***
|
− .0603***
|
− .17***
|
− .1755***
|
|
(.0033)
|
(.0095)
|
(.0155)
|
(.0033)
|
Higher education
|
− .324***
|
− .1132***
|
− .3082***
|
− .3263***
|
|
(.004)
|
(.0118)
|
(.0202)
|
(.004)
|
Wealth quintile (compared to poorest)
|
|
|
|
|
Poorer
|
− .0013
|
− .0183**
|
− .0004
|
− .0013
|
|
(.0032)
|
(.009)
|
(.0132)
|
(.0032)
|
Middle
|
− .008**
|
− .0228**
|
.0094
|
− .0077**
|
|
(.0035)
|
(.0101)
|
(.0148)
|
(.0035)
|
Richer
|
− .0236***
|
− .0316***
|
.0052
|
− .0232***
|
|
(.0038)
|
(.011)
|
(.0173)
|
(.0038)
|
Richest
|
− .0427***
|
− .0384***
|
− .0047
|
− .0419***
|
|
(.0044)
|
(.0129)
|
(.0216)
|
(.0044)
|
Age at time of survey
|
− .0026
|
− .0828
|
− .0467
|
− .0044
|
|
(.0185)
|
(.0553)
|
(.0751)
|
(.0186)
|
Constant
|
.462
|
2.2364*
|
1.7795
|
.5205
|
|
(.5194)
|
(1.2714)
|
(2.1795)
|
(.5224)
|
Observations
|
2314697
|
546859
|
461315
|
2314697
|
R-squared
|
.7772
|
.908
|
.8994
|
.7767
|
Robust standard errors are in parentheses |
*** p < .01, ** p < .05, * p < .1 |
Furthermore, in regression (4), we do find that all three types of violence increase the risk of child marriage, however, the effect is particularly large for state-based conflicts compared to non-state conflicts or one-sided violence.
Or to illustrate the findings differently: a coefficient of 5.9 percentage points in regression model (2) – being the model with an effect size somewhat in the middle between (1) and (3) – and a base child marriage rate of 29.2% across our sample, means that children within conflict zones (as defined above) are on average 20.1% more likely to be married as a child compared to children not affected by conflict.
The findings of the whole sample do hide regional differences though. We run separate regressions for each world region (see coefficient results for the main variable of interest in Fig. 2). We find consistently significant and positive effects across all three regression models in East Asia & Pacific and Latin America & Caribbean. In South Asia and Middle East & Northern Africa, two out of the three regression models are significant and show a positive impact of conflict occurrence on child marriage. In Sub-Saharan Africa and Europe & Central Asia, only regression model (1) with the full sample is significant and positive, however results are no longer statistically significant when we restrict the sample to five years prior the survey or use a smaller sample for which we can control for residency at the time of marriage. This suggests that migration and displacement in those regions prevent us from getting conclusive findings using our methodological approach.
While regression model (1) is statistically significant in 40 countries, country-specific regressions often lead to inconclusive and statistically insignificant findings once we restrict the sample to control for possible migration or displacement. This likely driven by the significantly smaller sample size (either due to the focus on marriages in the five years prior to the survey or by restricting the sample to those surveys where we can control for the residence at the time of marriage) and the potentially limited variation in conflict zones over time depending on the type of conflict. We find significant effects for 11 countries in regression model (2) and 10 countries in regression model (3). In most countries (Angola, Burundi, Cambodia, Chad, Democratic Republic of Congo, Egypt, Guatemala, Liberia, India, Peru, Sierra Leone, Togo, Timor-Leste, and Uganda) children affected by conflict were more likely to be married than their peers not living in a conflict zone five years prior to marriage, a result that aligns with our global and regional estimates. However, in a smaller group of countries (Bangladesh, Madagascar, Mali, Nigeria), we did find significant and negative effects of conflict on child marriage, suggesting that children in conflict zones were less likely to get married than children unaffected by the occurrence of armed conflicts. These finding align overall with those found by Krafft et al. (2022), who also found mixed results across the 19 conflict-affected countries included in their study. However, more context-specific work is needed to explain the variance between different countries.[3]
For the other variables included in our models (especially education and wealth), we do find the expected effects: the higher the level of education a woman has and the wealthier the household she lives in, the lower the risk of child marriage. The age of women at the time of survey is not significant in any of our models, which was expected given that we control for year of birth in our fixed effect model.
Finally, we also estimated the effects of experiencing conflicts on child marriages taken place below the age of 15 (see Table 3 in Annex). We do find a significant and positive effect for regression models (1) and (4), suggesting that children affected by conflict are more likely to marry below the age of 15 than their peers not affected by conflict. However, those results are no longer significant once we restrict the sample to control for possible migration effects.
Conflict severity
The previous results estimated the effect of any conflict event taking place on child marriage. However, it can be assumed that this effect is more pronounced as more severe conflicts are (measured by a higher record of deaths). Therefore, regressions (5) to (8) presented in Table 2 analyse the effect of conflict deaths (in logarithm) over the last five years before the year of marriage. Similarly, to the observation for the occurrence of conflicts, we find the largest effects in regression (5) without any restrictions of the sample. In this case an increase of conflict deaths by 1% leads to an increase of the incidence of child marriage of 3.1 percentage points. The regressions (6) and (7) find lower effect sizes, once we restrict the sample to marriages within the five before the survey, and marriages, where we have information that the women are still living in the sample village. Again, mirroring the results in model (A), regression (8) finds larger effects for increases in deaths due to state-based than for increases due to non-state conflicts and one-sided violence. All control variables are significant and have the same direction of effects as described above.
Table 2
Results of linear probability regression for conflict severity on incidence of child marriage
|
(5)
|
(6)
|
(7)
|
(8)
|
|
First marriage by age 18
|
Number of conflict deaths in or five years prior year of marriage (log)
|
.0314***
|
.0159***
|
.0166***
|
|
(.0005)
|
(.0029)
|
(.0035)
|
|
Number of deaths due to state-based armed conflict in or five years prior year of marriage (log)
|
|
|
|
.0197***
|
|
|
|
(.0007)
|
Number of deaths due to non-state conflict in or five years prior year of marriage (log)
|
|
|
|
.009***
|
|
|
|
(.0013)
|
Number of deaths due to one-sided violence in or five years prior year of marriage (log)
|
|
|
|
.0073***
|
|
|
|
(.0007)
|
Highest level of education (compared to lower than primary)
|
|
|
|
|
Primary education
|
− .0412***
|
− .0137
|
− .0327**
|
− .0413***
|
|
(.0032)
|
(.0092)
|
(.0144)
|
(.0032)
|
Secondary education
|
− .1728***
|
− .0604***
|
− .1695***
|
− .176***
|
|
(.0032)
|
(.0094)
|
(.0155)
|
(.0033)
|
Higher education
|
− .3215***
|
− .1131***
|
− .3069***
|
− .3299***
|
|
(.004)
|
(.0118)
|
(.0202)
|
(.004)
|
Wealth quintile (compared to poorest)
|
|
|
|
|
Poorer
|
− .0012
|
− .0182**
|
− .0006
|
− .0013
|
|
(.0032)
|
(.009)
|
(.0132)
|
(.0032)
|
Middle
|
− .0077**
|
− .0227**
|
.0092
|
− .0075**
|
|
(.0035)
|
(.0101)
|
(.0148)
|
(.0035)
|
Richer
|
− .0234***
|
− .0316***
|
.0052
|
− .0229***
|
|
(.0038)
|
(.011)
|
(.0172)
|
(.0038)
|
Richest
|
− .042***
|
− .0384***
|
− .0048
|
− .0413***
|
|
(.0043)
|
(.0129)
|
(.0215)
|
(.0044)
|
Age at time of survey
|
− .0031
|
− .0841
|
− .0486
|
− .0052
|
|
(.0185)
|
(.0549)
|
(.0752)
|
(.0187)
|
Constant
|
.4698
|
2.2546*
|
1.815
|
.5523
|
|
(.5196)
|
(1.2608)
|
(2.1805)
|
(.5241)
|
Observations
|
2314697
|
546859
|
461315
|
2314697
|
R-squared
|
.778
|
.9081
|
.8996
|
.776
|
Robust standard errors are in parentheses |
*** p < .01, ** p < .05, * p < .1 |
Similar to the findings earlier, we do find regional variation underlying the results. As before, we do find consistently significant and positive effects across all regression models in East Asia & Pacific and Latin America & Caribbean. Apart from Europe & Central Asia, at least two out of three models show significantly positive effects in all other regions. Using severity instead of occurrence of conflicts confirms the country-specific effects presented earlier.
Equivalent to our findings for conflict occurrence and child marriages below the age of 15 earlier, we only find a significant and positive effect of conflict severity in regression model (5). Other models restricting the sample to marriages in the five years prior the survey or controlling for residency at the time of marriage are no longer statistically significant (see Table 4 in Annex).
Robustness checks
As mentioned earlier, we run various sensitivity checks to test the robustness of our findings. First, reducing (increasing) the time period for which we consider a conflict event before the year of marriage, decreases (increases) the effect size in both models (A) and (B) – as one would expect –, with the effects of the variables of interest remaining significant in both cases.
Second, we run both logistic regression as well as probit model to confirm the significance and general direction of the findings in a non-linear model. While we cannot follow the same methodological approach of focusing on variations within each cluster given the limitations of both models, the logistic and the probit regression are highly significant for the variables of interest and confirm the direction of the effects found in the linear model.
Third, our findings show that women with higher education levels or in wealthier households are overall less likely to be married as a child. While we believe that such socio-economic variables improve the estimation, there is also a risk of multicollinearity, as girls who get married as children might subsequently drop out of school and are poorer as a result. When we test the robustness of our overall estimates by excluding those variables, our different regression models remain highly significant and effect sizes even increase.
Finally, we include two DHS surveys from India (2015/16 and 2019/21), with a combined sample of 1,042,445 marriages. This represents 41.5% of marriages in our sample, and consequently it is possible that the overall findings are driven by this one country. Repeating the global estimates without India, we indeed find no significant effect of conflict in regression model (2), however strongly positive and slightly larger effects in regression models (1), (3), (5) and (7), and a significant, but smaller effect in regression (6). Furthermore, our regional estimates presented above show that most other regions also experience significantly positive effects of conflict on child marriage.
Limitations
Several limitations characterise this kind of analysis. First, due to data limitations, our study only focuses on the impact of child marriage on girls, not investigating similar effect for boys. Second, while the use of cohort data allows us to analyse time trends within clusters, this depends largely on information about the age at first marriage, which could be subject to recall bias. Third, using this quasi-longitudinal data also relies on the implicit condition that women have not moved in or out of the cluster over time, with information on residential history only being available in a subset of DHS surveys. As discussed previously, we employ multiple sensitivity analyses to control for this fact, but more work is needed to control for cross-cluster migration and displacement. Fourth, while the DHS household surveys provide good background information on risk factors within the current household, they do lack historical data, especially on the household situation prior to marriage. This would in fact require either additional questions on household conditions prior to marriage (although this would potentially be severely impacted by recall bias) or the existence of longitudinal panel data, which could then reveal information on asset wealth, education of a woman’s parents, violence in the family etc. Finally, comparable household survey data is often lacking in some of the most conflict-affected countries. Out of the 17 conflict-affected countries in 2022/2023 as categorised by World Bank (2022), georeferenced DHS data is missing for 8 countries and older than 2015 for another 3 of those.
[3] More research is particularly needed on the link between shocks and social norms related to bride price or dowry (Corno, Hildebrandt, and Voena 2020; Chort, Hotte, and Marazyan 2022). Data on those practices is not consistently available for all countries or individual clusters (and we would expect this vary over time within the same cluster) and was therefore not included in this analysis. However, the social norms seem to vary across countries with very different outcomes, as we see positive effects in both countries with bride price (Burundi, Liberia, Sierra Leone, Uganda) as well as dowry (India), and vice versa. (Corno, Hildebrandt, and Voena 2020; Mansoor 2011)