Summer  >>  2012  

The Future of Counterinsurgency

Altogether Quantitative, Scarcely Analytical ...

Computational Spatial Social Scientist
Data Tactics Corporation
McLean, VA

Mr. Richard Heimann (@rheimann) recently
returned from Afghanistan where he supported the 82nd Airborne Division in
Regional Command South as a quantitative assessments analyst. He currently works
for Data Tactics Corporporation focusing
on advanced analytics, big data and cloud computing. He teaches Human Terrain
Analysis at George Mason University, is adjunct faculty at The University of Maryland, Baltimore County, and also
supports DARPA in Arlington Va.

The military sets conditions that either inspire a native population to embrace the efforts of a counterinsurgency (COIN) in an exhibition of solidarity, and thereby rejecting the legitimacy of an insurgency, or that motivate a population in an entirely different, inauspicious direction. Thus, operationalizing and analyzing key variables of combat effectiveness and, to a greater extent, the human terrain, aid commanders in producing acutely sensitive understanding of the battle space – one of often complete and otherwise certain ambiguity in irregular warfare.

Successful counterinsurgencies, both tactical and strategic, require certain computational rigor. Without such rigor, decision makers inadvertently support quite incorrect conventions. These incorrect conventions lead to seriously misplaced and spurious conclusions. As a community, data scientists and social scientists alike must develop a framework that uses data to rationally generate, test, discard, and modify operational alternatives, theoretical constructions, and conventions of all types.

The burden of COIN however, is the measurement of largely misunderstood human, social and political factors of attitudes and behaviors over time and space. Any lack of computational rigor is exacerbated by poor measurement, vis-a-vis, one never validating the other. Without good analysis, the value of measurement is never known, while misplaced value on measurement marginalizes analysis. The efforts on behalf of the International Security Assistance Force (ISAF) and ISAF Joint Command (IJC) in measuring, analyzing and ultimately understanding the tepid human terrain (Heimann, 2011) and insurgency landscapes continue to be suspect.

It has been said that those who do not learn from history are ultimately doomed to repeat it. If true, what ought to be learned after ten years and two concurrent wars? Lessons of previous wars, especially shadows cast by the more recent ones, no matter how unsuitable, overwhelm commanders and policymakers alike. Compounding matters is data collection of vast proportion, at varied scale and complexity both within and among campaigns, presenting certain challenges for decision makers.

The unabated blending of these cognitive pitfalls and computational hurdles leads to so-called “big data” problems of consequential significance, the basis of which is the lack of a sound theoretical COIN framework. A notional problem is that“... we (often) have more data than [we]think, but (often) need less data than[we] have,” (Hubbard, 2010). Hubbard is suggesting, quite evidently, that data lacks uniform utility or explanatory power. This fact informs measurement and analysis in ways that have not yet been realized.

COIN is a data-rich, but theory-poor environment. The lack of a theoretical framework is the paradigm for big data and necessitates exploratory analysis and large-scale pattern discovery and recognition.

Theoretical Framework

Henry Kissinger poignantly summarized Vietnam when he said, “We fought a military war; our opponents fought apolitical one. We sought physical attrition;our opponents aimed for our psychological exhaustion. In the process,we lost sight of one of the cardinal maxims of guerrilla war: the guerrilla wins if he does not lose. The conventional army loses if it does not win.”

FIGURE 1. IED attacks are a rather common tactic of insurgents and represent one kind of SIGACT. Trends are reported to commanders and decision makers in a fashion similar to Figure 1, ignoring spatial patterning and any de-trending of time-series data.
While many deny the parallels between the wars in Vietnam and Afghanistan, certain parallels do exist.One striking example is the strong desire for measurement and quantitative analysis, underscored by the lack of a theoretical framework. The Military Assistance Command Vietnam and ISAF measured the battlespace of irregular warfare and rightfully so.The number of engagements by third-party counterinsurgencies has been plentiful over the past 100 years,and given the military superiority of the U.S., it seems plausible that the irregular model will be used with feverish frequency.

Historical third party counterinsurgencies include the French in Algeria, the British in Malaya, and the U.S. in the Philippines,Vietnam,Iraq and Afghanistan.The empirical perspectives are rich and noteworthy, but may also be deterministic.COIN suffers from physics envy; it is a data-rich, but theory-poor environment. The lack of a theoretical framework is the paradigm for big data and necessitates exploratory analysis and large-scale pattern discovery and recognition.

Tommy Franks noted how counting bodies would not be used as a metric for success. Major General Michael Flynn too noted that old-school metrics like body counts were being focused on too much, perhaps remembering the credibility problems of measuring body counts during the Vietnam War or the dubious inference associated with the analysis of body counts and other kinetic events.One certainly has to wonder what is to be measured if body counts are not. It seems that in the absence of body counts,commanders instead rely on Significant Activities (SIGACTS). SIGACTS are the counting of violent attacks by or on coalition forces. The most notable perhaps are Improvised Explosive Devices (IEDs).

Measurement errors do exist with SIGACTS. The system that stores the data, Combined Information Data Network Exchange (CIDNE) is certainly not perfect. However, the source provides an unusually objective and consistent base of information. The error inherent in the data seems rather systematic. Put differently, the errors appear to be irregular, and over and under counts seem to be random and lack any structural regularity.

The most important acknowledgement however, is that measurement occurs only when coalition forces are present to observe the act. This type of observer bias relies on an inductive inferential model built upon uncertain ground. The Joint IED Defeat Organization (JIEDDO) estimates that 50%of IEDs go undetected (Erwin, 2010), adding additional complexity with forms of a selection bias to an already enfeebled approach to measurement.

The real issue with SIGACTS is the weak inferential model. The United States spent $3.63B in 2006 on a largely technical, engineering-based Counter IED effort, most of which went to JIEDDO. JIEDDO has certainly saved lives and this article does not contend with that fact. However, the ability to understand the true nature of violence has been remarkably elusive. JIEDDO relies on rather primitive statistics to justify its claims of success. See Figure 1.

These metrics do not effectively capture or accurately reflect performance of COIN. IEDs prevail due to their ease of assembly, small support structures as well as their psychological impact on both friendly and civilian populations. For an organization that originally funded human terrain, little is known about the relationship between SIGACTS and COIN effectiveness.More attention ought to be paid to violence as a process rather than a rather broad and imprecise act. As a process it would be understood in the context of counterinsurgency rather than a result of an insurgency and comparatively to other types of violent acts. Counter-IED research draws too often off of its own dependent variable.

George Box said that all models are wrong, but some are useful. The best models in Afghanistan rely on primitive statistics and ignore any spatial structure.Worse, analysts seem impotent to adapt. Certain isotropic properties are exhibited in all time-series analyses. The isotropic properties of time-series analysis are easy to understand but do not contain the merits of spatial analysis,most notably anisotropy. The unidirectional nature of time-series analysis is a simpler model, but not the most useful,as it fails to account for the true multi-directional nature of COIN. Spatial data,as it turns out, is special.

FIGURE 2. This figure reflects the utility of localized statistics. Variables that show little regional change (right) inform global COIN, and variables that show significant regional change (left) inform local COIN.
Stathis Kalyvas wrote in The Logic of Violence in Civil War (2006) of the non-monotonic function of violence(i.e. decreases in violence don’t always require mission effectiveness). An in surgency is simply more than violence and COIN is more than counting violent acts. The measurement of SIGACTS in COIN has value; these measures are not an elixir, however, and are misused with a frequency close to their rate of exploitation.For example, interdicted weapon caches give a false sense of success; the insurgency can still win the population,as the French learned from their interdiction efforts in Algeria. On the other hand, COIN cannot win the war and lose the country, akin to Vietnam where tactical success did not equate to strategic success. The insurgency wishes to out-administer the counter efforts, not necessarily defeat efforts tactically.

Understanding SIGACTS requires acknowledgement of a bidirectional causal model. In a rather counter-intuitive manner, coalition forces’ mere presence accounts for the variability in the data, despite intuition suggesting that increased patrols will have a negative relationship with SIGACTS. An increase in SIGACTS may simply be a reflection of our own increase in activity if only influenced by force strength.

COIN operations,in this case, are not exogenous. This is instinctively known by commanders but misunderstood by data analysts. Sadly,commanders falsely accept this defective model because itis what is reported.Broadly, the loop of causality between the independent and dependent variables of a COIN model leads to endogeneity. The simultaneity between the two factors highlights the weak causal model and may suggest reverse causality. The near complete disregard for endogeneity indicates a fundamental failure in empirical COIN research. The model is measurably complex; useful models, however,should remain the goal, and SIGACTS should be used with caution.

Computational Counterinsurgency - Spatial Pattern Recognition and Spatial Statistics

David Kilcullen (2009) explains that today’s conflicts are a complex hybrid of contrasting trends that counterinsurgencies continue to conflate, blurring the distinction between local and global struggles, and thereby enormously complicating the challenges faced.Kilcullen steps through local and global struggles and outlines the importance of commensurate policy. This process can be characterized roughly as useful spatial models whose statistically significant global variables exhibit strong regional variation to inform local policy, and as statistically significant global variables that exhibit little regional variation to inform region-wide policy. See Figure 2.

Unfortunately, computational COIN relies on time-series data almost exclusively,thus collating all actors into one fighting force operating on an assumed homogeneous population base. Operational analytical workflow consists of mere plotting of temporal patterns and describing discrete time {Xt-1} over time{Xt) change. The nature of insurgency,like most phenomena, is change overtime and space. Kalyvas examines how strategies vary temporally and spatially,focusing on the spatial variation of control on the part of the counterinsurgency and the spatial variation of violence. Kalyvas concludes that violence is nonrandom and non stationary. Insurgency, he concludes does not resemble a Hobbesian world. In other words, violence does exhibit spatial structure and commensurate spatially explicit theory should follow a blended idiographic and nomothetic methodology.

FIGURE 3. Global regression models represent an average COIN. One best of fit line must be made whereas local regression fits multiple local averages to better represent the counterinsurgency landscape.
Time-series analysis shares some similar challenges as spatial analysis.Shared challenges include the modifiable time unit problem, analogous to the well known modifiable areal unit problem,intrinsic heterogeneity, and others. The most interesting perhaps is Simpson’s Paradox and the spatial companion,Spatial Simpson’s Paradox (SSP). SSP represents the analytical equivalent to Kilcullen’s central thesis (see Figure 3), where poor COIN operations for one group or location and poor COIN operations for another group or location are good for everyone, if analysts just collapse over the grouping variable or over space.COIN operations that achieve a significant positive effect on average might still be undesirable because they leave a large fraction of the population worse off.

Uprooting insurgency, for example from city centers, has shown to have a negative correlation with civilian casualties as well as an impact on insurgent tactics. Spatially explicit theory, e.g.proximate casualty hypothesis, (Gartner, Segura, and Wilkening 1997) is in greater need for COIN operations and provides theoretical evidence to support certain operations. Temporal patterns just fail to have the same impact on COIN planning and policy as does spatial data, driven by measures of spatial dependence and spatial heterogeneity. Statistically significant local statistics prevent the Lake Wobegon Effect of a uniformly above average battle space,decomposing patterns and exposing spatial spillovers, spatial externalities, structural instability, spatial drifts and spatial regimes.

Big COIN Data

Certain computational complexities are evident in COIN, whether small data or big data. Due to turnover caused by multiple deployments, one wonders ifthe same lessons have been learned one year at a time for ten consecutive years versus an enemy with better memory.The reality is that no one knows more on their last day of a deployment and lesson their first. Big COIN Data provides an expedited learning curve and fluid understanding of the human terrain from campaign to campaign, deployment to deployment, and region to region.

Direct observation and blink (Gladwell 2005) qualitative assessments by battle space owners will always outpace the timeliness of computational counterinsurgency. These qualitative assessments,however, tend to be empirically deterministic and thus inappropriate for the construction of COIN theory. The transition to computational COIN is not one of mere desire but of absolute necessity.

The journey, however, has not been without missteps. The efficacy of computational COIN has seen only limited success. That limitation is due to a number of factors, but mainly that results often describe known variation, in a choropleth map or time-series, for example, due to known factors. If Big COIN Data is to be successful, then results must relate the unexplained components of variation,particularly spatial variation for analysts and decision makers. Our visual-cognitive system is ill-suited to the task of mentally removing known components of variation, envisioning the patterns in residuals, and relating these patterns to other variables.

“You have to understand not just what we call the military terrain…the high ground and low ground. It’s about understanding the human terrain,really understanding it.”

- Gen. David H. Petraeus, U.S. Director of CIA

Furthermore, certain social processes cannot be explained without greater incorporation of space. The inflexibility of traditional social theory and COIN theory, regardless of space, is slowly giving way to new spatially explicit COIN theories. The current developments and interest in computational COIN are aided by broader trends in computational social science, data science, and spatial statistics.

Kurt Vonnegut, a 20th century American writer, is quoted as saying that any search for the one will ultimately prove to be incomplete. Computational COIN is built upon the idea of heuristics. The pursuit of an exhaustive singular metric to understand a complex space is not realistic. Spatial data analysis offers a balanced blend of idiographic and nomothetic laws, offering local and global understanding of an insurgency and of the native population. The emergence of big COIN data allows all available data to be analyzed. It may challenge the traditional laws of parsimony, finding elusive omitted variables causing the Simpson’s Paradox, and calibrating local spaces causing Spatial Simpson’s Paradox.

Big Data deviates significantly from traditional methods of analyzing geographic data, the birth of which dates back to the quantitative revolution in geography during the 1960s. Perhaps the emergence and merging of these efforts solve a number of enduring problems in calibrating the battles pace of irregular warfare. The times of quite idiographic surveying of local tribes and inventory of body counts, SIGACTS, and other kinetic and combat-related metrics are waning.

Sensors & Systems | Monitoring, Analyzing and Adapting to Global Change | Stay in tune with the transformation. Subscribe to the free weekly newsletter.