4 e mpirical analysis the theoretical narrative has

4 E mpirical analysis
The theoretical narrative has two core empirical implications: reprisals
should be less intense where partisans interdicted external resources at a
high rate, and more intense where partisan coercive violence was high. I
now test these predictions with archival data on Belarus, in several steps.
First, I use semi-parametric regression analysis to gauge the overall re-
lationship between partisan violence and reprisals, while accounting for
other covariates, and fixed differences across geographic locations and time
periods. Second, I conduct a more focused, matched analysis of a single
major derailment campaign – Operation ‘Rail War’ in August 1943 – to
more rigorously account for partisan target selection and covariate imbal-
ance. Third, I use a marginal structural model to account for post-treatment
bias and repeated exposure to interdiction. Unless otherwise indicated, I
conducted each test on two levels of analysis (district-month and village-
month), separately for three outcome measures of German violence: people
killed, houses destroyed, and overall number of reprisals.
4.1 O verall patterns
I examine the general relationship between partisan activity and subse-
quent reprisals using a semi-parametric Generalized Additive Model with
a quasi-Poisson link. The quasi-Poisson model – an extension of the Pois-
son GLM with an unrestricted dispersion parameter – is appropriate be-
cause the outcomes are overdispersed event counts. To account for long-
term geographic variation and temporal shifts in baseline intensities of
reprisals (e.g. east-west trend in Figure 1 b, monthly fluctuations in Fig-
24

ure 2 ), the model includes a spatial spline and time fixed effects.
79
Table 3 shows regression coefficients and model fit statistics at the district-
month level, estimated separately with absolute and proportional levels of
rebel interdiction as covariates.
80
Table 4 reports the same results at the
village-month level.
81
Figure 3 shows simulations from these models.
The theoretical model’s predictions find strong support in the data. As
Figure 3 shows, German reprisals were most severe where partisans de-
railed few trains, but destroyed many garrisons. An increase from 0 to
10 derailments in a district-month was associated with 14 fewer houses
burned (95% CI: -9.88,-20.09) in retaliatory attacks. At the village-month
level, the negative relationship was even starker: where partisans carried
out at least two derailments per month, there were subsequently zero Ger-
man reprisals on average. Coercive partisan violence had the opposite
impact. For each garrison the partisans attacked per district-month, the
Germans razed 19 additional homes (95% CI: 12.69,28.5).
A similar story holds with a composite measure of partisan violence –
interdiction as a proportion of overall partisan violent activity (rightmost
panes of Figure 3 ). In district-months where Soviet partisans focused on
attacking supply lines rather than personnel, reprisals by occupying forces
were less frequent. Compared to localities where partisans used only coer-
cion, interdiction-only districts saw 30 fewer homes (95% CI: -19.97,-46.35)
79
The core specification is
y
it
=
g
-
1
(
α
+
γ Z
i , t - 1
+
θ y
i , t - 1
+
β X
it
+
s
(
long
i
+
lat
i
) +
ζ
t
+
e
it
)
(1)
where g
- 1
( · )
is an inverse quasi-Poisson link, Z
i , t - 1
is partisan tactics during the previous
month (with separate models for absolute and relative levels of interdiction and coercion),
y
i , t - 1
is a one-month time lag of the outcome,