Recording, Counting and Estimation of War Deaths

A Wall of the Pinkas Synagogue

Each of these Holocaust victims is memorialized by a name on a wall inside the Pinkas Synagogue in Prague.

Muslim Burial Ground and Peace Garden

This memorial in Woking, UK is a short walk from my house.

The Names on the Wall

The memorial lists names of Muslims who died for the UK in WW1 and WW2

Casualty Recording

The walls of the Pinkas Synagague and the Muslim Burial Ground and Peace Garden are exemplary works of casualty recording.

Monument to the Unknown Civilian

This is an initiative of the NGO Humanity and Inclusion

This need to memorialize an unknown civilian underscores the limits to casualty recording - in most wars it is not (yet) possible to list every single person killed.

Casualty Counting

Casualty counting is a form of guesstimation as exemplified by Enrico Fermi’s famous piano tuner problem.

Most high-qualty figures on sizes of whole wars come from casualty counting.

Casualty Counting can Generate Important Real-Time Insights into Ongoing Wars

Posts on the Mystics and Statistics blog (see this and this) demonstrate that Ukrainian claims for Russian losses in the war in Ukraine are greatly exaggerated.

The Statistical Estimation of War Deaths

Statistical estimation is of little help in memorializing individual victims and it struggles to provide practical real-time information about ongoing wars.

But the statistical estimation of war deaths can, potentially, inform the historical record about total deaths in war and their patterns better than casualty recording or casualty counting can.

A Word of Caution on Estimation

Several of the most high-profile statistical estimates of war deaths have badly distorted the historical record.

Iraq - A Graveyard of Surveys

Three separate surveys dramatically overestimated violent deaths in the Iraq War

A Good Idea does not Imply Good Execution

My point not that all surveys are bad - I argue, simply, that we should take seriously the concept of survey quality.

In other words, I merely wish to challenge a (sadly) common line of “thinking” along the lines of: “surveys are a scientific technology - this thing right here is a survey - therefore, its findings are entirely reliable and trump all other information.”

A Corrective on Multiple Systems Estimation (MSE)

Here is a lovely and well-known result based on multiple random samples. It is reprinted in the book “Studying Human Rights” by Todd Landman.

MSE - The Triumph of Alchemy?

Unfortunately, Landman goes on to claim:

“The key difference between the statistical estimations used in public opinion research [surveys] and MSE is that where public opinion research uses random samples of the population, MSE uses multiple non-random samples of the population. Both forms of analysis produce statistical estimates with associated margins of error…Endogenous sources of bias such as lying, timidity, and political mobilization of testimony can be controlled …”

Non-random Sample - Part 1

But with non-probability samples we can have a fixed overlap structure underpinned by a total of 100 deaths...

Non-random Sample - Part 2

.…or we can have the exact same overlap structure underpinned by 1,000 (or 10,000 or 100,000) total deaths.

MSE does not Solve all the old Problems of Alchemy

That is, MSE does not magically transform non-probability samples (Alchemy’s wood) into valid estimates (gold).

To the contrary, MSE works by imposing a probability structure on the data.

Quality Out Requires Quality In

When MSE yields good estimates then the data sets employed in the estimation process must be worthy of close attention in their own right.

There is no magical garbage-in-quality-out transformation.

The End