# Ministers’ incumbency ends after 1.79 failures with daily failing probability 0.11%

Today, Thomas Metz made me aware of a dataset about ministers in Eastern German federal states (Bundesländer) by Sebastian Jäckle. The dataset includes the variable “duration of incumbency” in days for 291 ministers between 1990 and 2011.

I was curious to look at the distribution of duration with the intention to be brave as a physicist and infer a simple stochastic model which reproduces that distribution. I copied the duration data into a matlab vector duration, made histograms, fits for different distributions and KS-Tests. As duration is a discrete random variable (days starting from inauguration), distributions living on the nonnegative integers are the natural candidates. The classical one-parameter distributions Poisson and geometric failed to deliver fitting distributions, but the negative binomial (NB) did surprisingly well. Normalized histogram for duration of incumbency and best-fit negative binomial probabilty mass function.

The best fit yielded parameters and . The Kolmogorov-Smirnov test did not reject that duration data came from the distribution with these parameters (p=0.32), but rejects under reasonably small changes of the two parameters. Thus, it is reasonable to assume What model does this imply? Looking at the days in the incumbency of a minister. Let us assume that every day can either be a success or failure which happens with probability . The negative binomial is the distribution of the number of successful days until failures occur (there is an extension to non-integer number of failures). Our model is thus, that a minister’s incumbency ends after a certain number of failures (what ever that means in practice). The best fit suggests that under this model 1.79 failures are allowed during a minister’s incumbency and that failures are relatively rare events happening with probabilty 0.11% every day, i.e. on average the first failure happens approximately at day 900.

Further notes:
Here ist the matlab code which delivers the results
% computation dist='nbin';bins=0:500:7500;hi=hist(duration,bins);par=mle(duration,'distribution',dist); % plots clf;bar(bins,hi/sum(hi)/(bins(2)-bins(1)));hold on;x=0:8000;plot(x,pdf(dist,x,par(1),par(2)),'r','LineWidth',3); % Kolmogorov-Smirnov test [h,p]=kstest(duration,[x' cdf(dist,x,par(1),par(2))']) 
You can test other two-parameter distributions by changing dist, e.g. to 'logn' or 'gam'. If you want to check one-parameter distribution you have to further remove ,par(2) from the code. It turns out that also the gamma distribution delivers a fit which is not rejected. This is reasonable because it is sometimes seen as the continuous-valued version of the negative binomial. Also the Weibull distribution was not rejected (although with much lower p-value), this shows that also other models might be appropriate. As always with statistics and real world data, I assume that the KS-Test would reject my theory when we have a larger dataset, as certainly some deviations from the negative binomial trend get dominant (e.g. caused by election cycles).

I hope this “finger exercise” on finding a simple stochastic model that fits is inspiring for political scientists, although every political theory would likely immediately reject it.

This site uses Akismet to reduce spam. Learn how your comment data is processed.