Age-Period-Cohort Effects in Child Soliders

I was recently introduced to a method called an age-period-cohort (APC) model. These models are often used in medical research to determine to affect of age and age-cohort on disease or drug treatments; however, they have also been used in studies of violence and criminology. After being introduced to this model I was immediately struck by its applicability to the study of low-intensity conflict, particularly with respect to child soldiers. That is, are there upward trends in violent acts among children in war-torn areas? APC analysis might be able to illuminate such a trend by disaggregating data by birth cohorts.

Then yesterday, as if by some mystical act grace by the data universe, Chris Blattman, Assistant Professor of Political Science and Economics at Yale University and prominent scholar of child soldiers, posted five years worth of data (h/t Andrew Little for pointing it out to me) from his Survey of War Affected Youth (SWAY). From the description, “SWAY is a research program in northern Uganda dedicated to understanding the scale and nature of war violence, the effects of war on youth, and the evaluation of programs to recover, reintegrate, and develop after conflict.”

What luck! This new data set could provide exactly the resource needed to conduct an APC analysis on child solider, and this morning I decided to peek at the data and see what it had to say.

Using the Blattman data on males (part I) and a STATA package that implements the intrinsic APC estimator proposed by described by Yang, Fu and Land (Sociological Methodology, 2004), I estimated a simple APC model of violence among youth in Uganda. Specifically, the dependent variable in the model is Blattman’s measure of total violence which is a, “Sum of 25 violent acts received, witnessed, upon family, or perpetrated,” by the survey respondent. The model then controls for respondent’s reported job skill (categorical variable 0-4), level of education (categorical variable 0-16), health score (continuous) and whether the individual had ever been abducted (binary). The APC effect is estimated using respondent’s age and the calendar year the survey was conducted. Results are reported below (note, I am only reporting the coefficients from control variables and the cohorts, though age and period coefficients are generated):

At first glance, unfortunately, these result seem destined for the Journal of Null Results. None of the cohort effects carry any statistical significance, and their large standard errors indicate considerable weakness. There are, however, some items of note. First, the independent variables show a strong and significant positive influence on violence for health and abductions. This is logical, as it is likely children are forced into violent position based on some perceived physical ability, and health may be a proxy for that. Also, if a child has been abducted the probability of being forced to view or perpetrate a violent is logically higher. These variables likely also show considerable correlation (i.e., the healthier a child, the more likely they will be abducted).

Next, while weak, the coefficients on the birth cohorts do show “clumps” in the positive and negative effect of cohort on violence. For example, cohorts starting in the late 1970′s to the mid-1980′s appear to have the highest propensity for violence, which then generally trends downward into the 1990′s. Without any context this is difficult to parse, however, this is an interesting springboard for future research—with serious caveats. Clearly there is much more to be done here, and none of these results should be taken as anything more than a simple exercise of the data.

Finally, I encourage other to experiment with Blattman’s data and design competing models that are able to better explain APC effects in child soldiers. This example only touches on the numerous explanatory variables packed into the dataset, providing ample opportunity for additional models to be built. For STATA users, simply type findit apc_ie at the command line to download and install the APC intrinsic estimator algorithm, then check the help file with help(apc_ie) for the model syntax. I look forward to reading others’ results!

Photo: Chris Blattman

Automatically Generated Related posts:

1. Selection Effects, Causality, and Bad Policy
2. Measuring Network Effects with No Network Analysis
3. Is Human Altruism Evolved or Rationalized from Conflict?
4. Variation of the Frequency of Fatal Quarrels With Magnitude
5. Heatmap of Predicted Probability of Homicide in Philadelphia