The Causes, Consequences, and Management of Civil Wars

030:178, Section 1
Guide to Reading Statistical Articles

1) Don’t skip the tables! Even though the numbers and statistics might be difficult to work
through, it is important to figure out what their content implies. The authors should try to
convey in the text what the tables/figures are showing.
2) Unit of analysis: what does each case represent in the data? In this class, typical units of
analysis will include state-year, duration of peace (number of years since last war), and
civil war outcomes. Figure out what years and countries are covered in the study, which
will be useful for making comparisons across studies with different results.
3) Measurement: how are the key dependent and independent variables measured? Are they
nominal (e.g. male/female), ordinal (e.g. low democracy/high democracy), or continuous
(e.g. number of battle deaths)? Are they counts of events, such as number of war
participations? Dependent variables that are nominal and dummies (1/0) will typically be
analyzed using logit or probit models. Models that look at the duration of civil wars or
peace between wars will count the number of time periods (months or years). These
models are often called survival analyses because they are looking at the time to death
(e.g. the end of a civil war). If the dependent variable was ordinal, then the authors might
employ an ordered logit model. If the dependent variable was continuous, then a
regression model would be appropriate.
4) Statistical significance: Scholars are typically trying to determine if the factors they have
identified as independent variables (e.g. primary commodity exports, income inequality,
regime type) have a significant effect on the dependent variable (e.g. civil war onset).
The null hypothesis is that the effect is zero, while the alternative hypothesis is that the
effect is different from zero. One-tailed hypothesis tests specify the direction of
difference (positive, negative), while two-tailed tests specify only difference but not
direction. Most statistical packages report two-tailed tests, so it is fairly common for
authors to employ them. Social scientists typically employ a 95% confidence interval for
their tests, which means that if we drew 100 random samples from the population, 95
would contain the population parameter being estimated. Articles we read will
sometimes report the exact level of significance (p=.07), or the range (* will indicate
p<.10). There are two basic rules of thumb here. If the estimated parameter (b or β) is
twice as large as its standard error (or more), or b/s.e. > 2, then the effect of the variable
is statistically different from zero. For one-tailed tests, the threshold would be lower for
significance (t>1.65 for 95% significance). Authors might also report the p-value, which
is the area under the standard normal curve (or student t distribution) beyond the
estimated z(or t) score. If that p-value is less than .05, then the effect is statistically
significant.
5) Model fit: Scholars will also report statistics about the overall goodness of fit. These give
us a benchmark for how well the independent variables in the model explain the variance
in the dependent variable. One common measure is R2 which is a measure of the
percentage of variance in the dependent variable explained by the dependent variables. If
R2 equals zero, then the independent variables do not help us explain variance in Y; If R2
equals one, then the independent variables perfectly predict the variance in Y (the data
would fit on a perfect line in a bivariate model with two variables). Of course, most
 models in Political Science never approach R2 equal to one, so scholars will discuss the
percentage of explained variance. In the civil war context, a good model might explain
80% of the variance. Scholars might also report a chi-square or F-test for the overall
fitness of the model. If the p-value for these tests is less than .05, then we can conclude
that the model is a good one overall. Scholars may also report their improvement in
predictions for logit & probit models. These help to determine how much better we can
guess the actual value of the dependent variable (e.g. 1 for civil war onset, 0 for peace)
based on knowledge of the independent variables (predicted value based on model).
6) Substantive significance: In addition to determining whether the independent variables
have a statistically significant effect on the dependent variable, scholars are also
interested in the size of that effect. If we estimated the effect of class attendance on
grade, we could imagine you would come to class more often if each class raised your
grade by 5 percentage points than if the effect was only 0.2 percent. Scholars assess
substantive significance in regression models by looking directly at the size of the
coefficient, while in logit or probit models, they calculate the predicted probability that
the dependent variable equals one. This entails setting all independent variables at their
mean or mode and then moving one variable at a time across some range of possible
values (e.g. minimum to mean to maximum). You might see the actual probabilities
reported, or sometimes you will see the chance in predicted probabilities. Suppose that
the probability of civil war onset was 0.05 when democracy was high and 0.35 when
democracy was low. The change in probability would be 0.35-0.05/0.05 = 6, or a 600%
increase in the likelihood of war.
7) Multicollinearity: scholars may discuss this issue and it refers to the possible high
correlation between independent variables. Suppose that democracy and wealth are
correlated at 0.80. This means that typical democratic countries are rich and typical non-
democratic countries are poor. The problem this creates is that we cannot determine
whether civil wars occur because countries are poor or because they are non-democratic.
It is hard to assess the independent effects of each variable because they share so much
variance in common. What often happens is that the overall model fit looks good (R2
might be 0.70) but the t-scores for each variable are insignificant (t< 2). In these cases,
scholars sometimes put in one variable at a time or use statistics to assess the overall
effects of the correlations. It is important to think about correlations between variables
because these can have big effects on our findings in multivariate models.
8) Heteroskedasticity & Autocorrelation: scholars might discuss these issues and they create
similar problems to multicollinearity by increasing the standard errors and making it
more likely to think the independent variables have no significant effect. What happens
in the case of heteroskedasticity is that the model fits differently across the cases.
Imagine that at high levels of income, the model does a good job of explaining the
incidence of civil war, while at low income levels, some countries have civil wars, while
others do not. The model fits worse at the low range of income, which creates bigger
variance in the standard errors. Autocorrelation refers to temporal correlation between
cases over time. For example, presidential approval in one month is correlated with
presidential approval in the previous month. Typical solutions to these problems include
running fixed effects (putting in dummy variables for each country or each year) or using
statistical fixes for the problem (Generalized Least Squares, panel corrected standard
errors, cubic splines, lagged dependent variables, etc.)