Findings



First Data Set

Effect Likelihood Ratio Tests 



Source
L-R ChiSquare
Prob>ChiSq
COidv
4.04135133
0.0444*
COmas
9.60139333
0.0019*
COltowvs
0.82440907
0.3639
COivr
8.68558171
0.0032*
USNuke
2.20938042
0.1372
COpec
3.31489888
0.0687
USmilex
0.27008684
0.6033
COidv*COltowvs
23.0949931
<.0001*
COidv*COivr
7.6963466
0.0055*
COidv*USmilex
13.935887
0.0002*
COmas*USNuke
13.9181373
0.0002*
COivr*USmilex
8.86646576
0.0029*
COpec*USmilex
9.68384586
0.0019*

Table 1 - Variables included in the model for the first data set. 
Variables significant at the p < .1 are bolded.


The analysis of the model generated for the first data set shows a level of fit that is statistically significant (Chi Squared Value = 0.8165) and it has a misclassification rate of 20%. Of the three models analyzed, this model has the best fit and has exceptional specificity.

The confusion matrix indicates the following:
  • If the coercive attempt was not successful, the model successfully predicted it 91% of the time (specificity)
  • If the coercive attempt was successful, the model successfully predicted it 61% of the time (sensitivity)
  • The model has a positive predictive value of 76%
  • The model has a negative predictive value of 82%


Second Data Set

The initial analysis of the model generated for the first data set applied to the second data set shows that it has both statistically significant poor fit and a high misclassification rate (45%). Note that the Chi Squared test assumes that the model prediction is independent of the host nation. Breaking it apart by country reveals the weakness of the model—it only generates correct results only for the United States—the model’s successfulness does depend on the host nation. Thus it would not be appropriate to apply this model to other nations.

Further statistical analysis was unable to create a model that fit the data better than the model created for the first data set.


Third Data Set

The third data set excludes Germany in an attempt to capture the variables that are relevant only to world hegemons. Once again, the model built for the United States has poor fit and a high misclassification rate, so we abandon it and attempt to create a new model with better fit and greater predictive power. The resulting model still has significantly poor fit and a misclassification rate of 32%. The model’s predictive power is worse than before, but the specificity is still high.
The confusion matrix indicates the following:
  • If the coercive attempt was not successful, the model successfully predicted it 76% of the time (specificity)
  • If the coercive attempt was successful, the model successfully predicted it 52% of the time (sensitivity)
  • The model has a positive predictive value of 71%
  • The model has a negative predictive value of 58%
Because of the difference between the countries in terms of variables like Gross Domestic Product (USGDP) and Power Energy Consumptions (USpec), we instead consider the comparative scores (CompGDP, Comppec)—scores that give a ratio of the host country and the coerced country as a way to capture data from both countries in a single score. At first analysis, the fit is good with a misclassification rate of 28%. However, the sensitivity goes way down—the model is more likely to be wrong than right if it is predicting for a successful case.


Effect Likelihood Ratio Tests


Source
L-R ChiSquare
Prob>ChiSq
CompPDI
0.03317517
0.8555
Compmas
3.85730299
0.0495*
Compuai
2.6315265
0.1048
Compltowvs
1.9177664
0.1661
Compivr
0.26338428
0.6078
Compirst
0.43828583
0.5080
USNuke
0.23917085
0.6248
CompGDP
9.56186796
0.0020*
CompPDI*Compuai
0.76232262
0.3826
CompPDI*Compivr
5.07974193
0.0242*
CompPDI*USNuke
0.4325219
0.5108
Compuai*Compivr
2.80852889
0.0938
Compltowvs*Compirst
0.86148677
0.3533
Compivr*USNuke
4.025164
0.0448*
Compirst*USNuke
4.48072246
0.0343*
Compirst*CompGDP
2.92369884
.0873

Table 2 – Variables included in the model  for the third data set. 
Variables significant at the p < .1 are bolded.

The confusion matrix indicates the following:
  • If the coercive attempt was not successful, the model successfully predicted it 89% of the time (specificity)
  • If the coercive attempt was successful, the model successfully predicted it 39% of the time (sensitivity)
  • The model has a positive predictive value of 73%
  • The model has a negative predictive value of 65%

Fourth Data Set

We now consider the United States, the USSR and the United Kingdom only for the period in which they were considered world hegemons. Again, the model built for the United States has poor fit and a high misclassification rate. We abandon it and attempt to create a new model with better fit and greater predictive power. The resulting model has good fit but a misclassification rate of 35%. This ends up being the worst model yet.

The confusion matrix indicates the following:
  • If the coercive attempt was not successful, the model successfully predicted it 85% of the time (specificity) 
  • If the coercive attempt was successful, the model successfully predicted it 26% of the time (sensitivity) 
  • The model has a positive predictive value of 68% 
  • The model has a negative predictive value of 47%