Fri 2026-Nov-20

Ukraine War: 1.6 Million Russian Dead

Tagged: MathInTheNews / Politics / Sadness / Statistics

Russia has stubbornly plowed on, unfortunately reaching another milestone of human sacrifice of their own citizens in Ukraine.

Russia Just Keep Showing They’re Crazier Than We Thought

Porter @ MSN/Daily Beast: Putin says he didn't start war, Ukraine did We note sadly that Putin said at the Shenzen summit between Putin/Xi/Trump that he didn’t start the war, but by invading Ukraine he was trying to end some war or other that, in his mind, Ukraine had started! [0] This is so blatantly counterfactual that it barely even counts as a lie, being more in the way of an invitation to a hallucination.

Putin’s claim comes as a surprise to everyone else on the planet, who watched Russia start the war with Ukraine. Even on this Crummy Little Blog That Nobody Reads (CLBTNR) we have, for some regrettably long years, been expressing quantitative but severely negative opinions about the Russian invasion of Ukraine. [1] [2] [3] [4] [5] [6] [7] [8] [9] [10] [11] [12] [13] [14] [15] [16] [17] [18] [19] [20]

Our quantitative modeling of the whole mess started out as a way to collect objective evidence of the evil of Russian leaders, and the evil complicity of the west in failing to defend, or at least arm, Ukraine. Then it was just my ball of white-hot anger at the human sacrifice for stupid gods of power. Now… it’s just my continuing sadness that the mortal world is as it is, and resists change.

And here we are, repeating a gruesome but tired analysis. As you can see below, Russia has attained a new milestone of 1.6 million dead:

Ukraine Ministry of Defence @ X/Twitter: Russian losses reach 1.6mln

(And yes, that link goes to X/Twitter, despite my absolute revulsion for the pit of Nazis and fascists. It appears they’ve tied upxcancel.com and nitter in court for the apparent crime of quoting them. So, reluctantly, a link to X/Twitter. Please don’t bother following it.)

All previous quibbles about using the Ukrainian Ministry of Defence for casualty statistics, given their obvious interest, still stand. All previous responses also still stand: they seem to be middle of the pack in these matters, as the 19 posts above will attest in rather more detail than you probably want to know.

Suffice to say: this is a credible starting point for deciding what to think about the state of the world, and the state of the world is decidedly grim.

The Dreary Demographics of Russian Men

Once again, let’s make the repeated, dreary point that Russia is hollowing out the middle demographic of Russian men of military service age. As we’ve pointed out many times now, there started out about 24 million of them:

\[144\ \mbox{million Russians} \times 0.5\ \mbox{males} \times 1/3\ \mbox{military age} = 24\ \mbox{million}\]

If 1.6 million are dead in Ukraine, then that’s $100\% \times 1.6\ \mbox{million} / 24\ \mbox{million} = 6.67\%$ of the military-age male population, now gone.

The Russian emigration & COVID-19 deaths (see older post [14]: 900,000 men of military age emigrated and about 67,000 dead of COVID-19), costs about an extra million, for a total loss of 2.6 million. That’s $100\% \times 2.6\ \mbox{million} / 24\ \mbox{million}$, for a total loss about 10.83% of the Russian men of military service age.

So almost 11%. Russia is now paying tithing, but doing so “in kind” via the human sacrifice of their young- to middle-aged men.

Our Segmented Regression Model

To repeat our previous explanation of the segmented regression model:

Our “segmented” regression model is basically a linear model of the number of soldiers killed versus time, except that it can have a kink in it where the killing rate changes suddenly. As we explained previously, this means the number of soldiers looks like:

\[\mbox{Soldiers}_t = \beta_0 + \beta_1 \mbox{DayNum}_t + \beta_2 \theta(\mbox{DayNum}_t - \psi) (\mbox{DayNum}_t - \psi) + \epsilon_t\]

where:

  • $t$ indexes data points in time (here measured in days since 2023-Jan-22; the war itself started earlier, but this is when we started paying quantitative attention),
  • $\psi$ is a parameter for estimating the day number at which the kink occurs, and
  • $\theta()$ is the Heaviside step function (0 for negative argument, 1 for positive argument, here used as an indicator for “after day number $\psi$”, i.e., the fancy-pants modeling language for an “if” statement),
  • $\epsilon \sim N(0, \sigma^2_{\mbox{Soldiers}|\mbox{DayNum}})$ is the error usual error term, normally distributed around 0 with the conditional variance shown.

If we consider expectation values, the interpretation is immediately obvious for the mean number of casualties at a given time:

\[E[\mbox{Soldiers}] = \left\{ \begin{align*} \beta_0 + \beta_1 \mbox{DayNum}, & \mbox{ if } \mbox{DayNum} \lt \psi \\ \beta_0^* + (\beta_1 + \beta_2) \mbox{DayNum}, & \mbox{ if } \mbox{Daynum} \ge \psi \end{align*} \right.\]

Here:

  • $\beta_1$ is the slope before the kink,
  • $\beta_1 + \beta_2$ is the slope after the kink, and
  • $\beta_0$ is the intercept before the kink & $\beta_0^*$ is the post-kink intercept (depending on the slope difference and the location of the kink in a way about which we generally do not care, for most purposes).

The 2 slopes $\beta_0$ and $\beta_0 + \beta_1$ are the things about which we care most: did the rate of killing Russian soldiers actually go up after the kink?

Original regression: number of Russian soldiers over 120 days near the beginning of the war First, recall the regression we did several years ago, on about the first 120 days of the war:

  • Points:
    • The blue points are the early data points, collected daily.
  • Axes:
    • The horizontal axis is time, measured in days since I started measuring (not quite the start of the war).
    • The vertical axis is the number of Russian soldiers killed, as reported by Ukraine.
  • Fitted line:
    • The black line is the least-squares line.
    • The gray bands around it are 95% confidence limits, though as mentioned previously, I don’t quite trust these (they seem too tight to believe).

As you can see, the blue points wobble around the fitted line a bit, as one expects from real data. However, the fit is excellent. I won’t rehash all the old statistics (e.g., $R^2 \gt$ 99%), just look at it and see the upward linear trend in casualties.

Original regression: number of Russian soldiers, compared with subsequent data, indicating need for kink About a year later, I began to check in again on the casualty figures at each increment of 100,000 deaths. That’s what’s shown in this plot:

  • Points:
    • The blue points are the early data points, collected daily.
    • The red points are more recent data points, collected approximately every 100,000 deaths.
  • Fitted line:
    • The black line is line fitted to the early data only, i.e., using just the blue points.
    • The gray bands around it are 95% confidence limits, though as mentioned previously, I don’t quite trust these (they seem too tight to believe).
    • The dotted line shows where the initial data fit, without a kink, would have continued in a world with no change in killing rate.

I wanted to see if the line fitted to the initial 120 or so days of data had any predictive power. What I found was that the killing rate had suddenly taken an upward turn, and the data from early in the war was a significant underestimate. If the early war data had been a good predictor, the red points would be on the dotted line. Obviously, this is not that world.

Russian casualties in Ukraine over time: a segmented regression model So then I moved on to a segmented regression model, which is the fancy-pants statistician way of saying “line with a kink in it”. The results are shown here.

The best estimate, using the latest dataset, is that the kink occurred on or about 2024-Feb-17, with a 95% confidence limit of 2024-Jan-21 to 2024-Mar-15. This is well in line with previous analyses, indicating that this result is stable with respect to the arrival of new data.

At the kink, either the Ukrainians figured out something about fighting efficiently with drones, or the Russians started doing human wave attacks without a care for casualties, or the reporting changed, or… something. Something bad, if you’re a Russian; something hopeful if you’re a Ukrainian hoping the Russians will think better of all this and withdraw.

So far, the Russians have not thought better of much of anything here.

Statistical Significance & The Need for a Kink

The box in the lower right of the plot shows the result of Davies’ Test

for the statistical significance of a kink. At $p \sim 4.65 \times 10^{-65}$, it is obviously and absurdly significant. There is definitely a kink in the data; this is not just our opinion (though it is, indeed, also our opinion).

The Killing Rates Before & After the Kink

Davies’ Test above tells us the kink is statistically significant, i.e., likely to be a real thing that would reproduce if we could (somehow! horribly!) run the whole experiment again.

Having found a real result that is statistically significant, the next thing a statistician must ask is whether the presumed real effect is large enough to bother about, or small enough to ignore. This is a strength of effect measure, and I’ve become enough of a crabby old statistician to be irked by people who stop at significance and don’t assess strength.

For this, we’ll consider the change in killing rates before & after the kink. In the box at the lower right of the plot, we see these rates are 699.58 soldiers/day before, and 1183.30 soldiers/day after. (Also reported are 95% confidence limits, if you want to model the uncertainty here. It seems quite clear, so we won’t do that, though probably we should, to be excruciatingly thorough.)

That gives us a percentage change in killing rate of:

\[100\% \times \frac{1194.00 - 699.58}{699.58} = 70.7\%\]

Previously we’ve gotten changes from 69% to 74%, so this is broadly consistent, and a large change indeed. We have a large effect size!

3-Fold Crossvalidation: Is This Real?

Another reality-test statistic asks: have we really learned something about the world, or have we just memorized this one particular dataset? After all, if you only ask me to “predict” the casualties on a date already in the data, I can be 100% correct by looking it up, without having learned anything!

That’s why we always do $k$-fold crossvalidation. That means we divide the data up into $k$ subsets, train on $k-1$ of them, and then predict on the withheld data. Do that all possible ways, extract averages, and contemplate.

Here we divided the data into $k = 3$ folds, which is about as much as we can get away with given that there are not so many late time points.

3-fold cross-validation results and entire dataset fit, compared with simple linear (no kink) model This table shows the result for withholding each of the folds, and one more using all the data with nothing withheld.

  • The fitted parameters – kink position and the before & after slopes – are all pretty consistent, both across folds and using the whole dataset.
  • The standard deviations are also pretty consistent, saying that there’s nothing peculiar about any of the folds, i.e., we chose them fairly. The standard deviation does go down when using the whole dataset (last row), as expected.
  • The real test is the RMS error: this measures the root-mean-square error of prediction on the withheld data (except the last row, which uses all of it). So we’re measuring prediction error on data the training of the model never saw. It’s all pretty consistent across folds and shows the addition of the kink definitely reduces error.

In other words, the kinked model performs well on out-of-sample tests, and always outperforms the model with no kink. We should definitely believe the kink model.

The Weekend Conclusion

Everything necessary for peer review of the above is available, including the scripts, the data files, transcripts of runs and so on for your inspection. [21]

What in the world should we think of all this?!

Honestly, it reminds me of a story by British SF writer Charles Stross, called Scratch Monkey, available both in print and on Stross’s web site. It starts out with a nation killing its own inhabitants on an almost unimaginable scale, apparently entirely senselessly. Though the killing eventually makes villain-ish sense of a sort, it still boggles the mind. I can’t help but think Russia is doing something of this sort to itself (and Ukraine, about which they simply do not care, not even enough to conquer).

Hains @ MSN: Rubio quotes Pompey the Great, understanding nothing Verma @ Yahoo: Hegseth also quotes Pompey the Great, understanding nothing

Rather more chllingly, with the growth of fascism in the US around Trump, Russian attitude comes to America. The senseless worship of what amounts to state murder was articulated first by Secretary of State Rubio [22] and then by Secretary of Defense Hegseth. [23] Both of them quoted Roman general Pompey the Great, warning the International Criminal Court that they will not be bound by international law:

Stop quoting laws to men with swords.

It seems clear now that administration officials are not well read, and are fond of quoting classical sources which they have not read. Like their previous exhibition of ignorance when invoking the Melian Dialog, this invocation of Pompey shows they understand absolutely nothing about Pompey nor about how to interpret that quote. Someone programmed them with this snippet because they thought it sounded masculine and powerful; the truth is otherwise.

Let us sum up, with just a cursory stroll through Wikipedia that would have warned them:

  • Gnæus Pompeius was a Roman general in the final days of the Roman Republic, before it finally gave in to imperial strong-men.
  • He was given the cognomen ‘Magnus’, or ‘the Great’, by the dictator Sulla for some military successes. But his contemporaries called him adulescentulus carnifex, “teenage butcher” for his brutality and lack of sense of proportionality.
  • This particular line is from Plutarch (Lives, Pompey 10.3.2.), when the Mamertines complained that Pompey’s conquest was illegal; it is a straightforward evil of the “might makes right” sort:

    “Οὐ παύσεσθε,” εἶπεν, “ἡμῖν ὑπεζωσμένοις ξίφη νόμους ἀναγινώσκοντες;” (‘Will you not give up,’ he said, ‘reading laws to us men girt with swords?’)

  • He lost the Battle of Pharsalus, despite having a numerical advantage, and had to flee to Egypt, where he was beheaded. His body was burned, but the head was kept as “evidence”.
  • His end was summarized by Cicero as, “His life outlasted his power.”

So Pompey was a pompous (forgive me!) and brutal man who thought armies would spring up whenever he demanded. He was mistaken. He lost his final battle due to his near-total lack of perspective, died an ignominious death, and had a humiliating Ciceronian epitaph.

This is not the sort of person to quote if you wish to appear powerful! Doing so without understanding will, however, make you look ignorant and liable to fail gruesomely.

Quoting laws to men with swords is the essence of the rule of law, that might does not make right. Laws, in fact, form all of us in society into a giant man with a sword to enforce our laws. The laws determine which way the sword points.

Trump and his acolytes can temporarily defy this only because they have corrupted and perverted the laws and their enforcement at the Department of Justice. Perhaps the International Criminal Court can do us a favor and arrest them the next time they travel where it has reasonable jurisdiction?

(Ceterum censeo, Trump incarceranda est!)

(Et ceterum censeo, index Epsteiniani divulganda est!)


Notes & References

0: T Porter, “Putin says he didn’t start war 4 years after starting war with Ukraine”, MSN, 2026-Nov-19. ↩

1: Weekend Editor, “Another Grim Anniversary”, Some Weekend Reading blog, 2023-Mar-02. ↩

2: Weekend Editor, “Do the Ukrainian Reports of Russian Casualties Make Sense?”, Some Weekend Reading blog, 2023-Apr-15. ↩

3: Weekend Editor, “Update: Ukrainian Estimates of Russian Casualties”, Some Weekend Reading blog, 2023-May-01. ↩

4: Weekend Editor, “Updated Update: Ukrainian Estimates of Russian Casualties”, Some Weekend Reading blog, 2023-May-09. ↩

5: Weekend Editor, “Updated${}^3$: Ukrainian Estimates of Russian Casualties Hit 200k”, Some Weekend Reading blog, 2023-May-17. ↩

6: Weekend Editor, “Ukraine Invasion: 250k Russian Dead”, Some Weekend Reading blog, 2023-May-17. ↩

7: Weekend Editor, “Tacitus in Ukraine”, Some Weekend Reading blog, 2023-May-25. ↩

8: Weekend Editor, “Casualties in Ukraine: Grief Piles Higher & Deeper”, Some Weekend Reading blog, 2024-Apr-10. ↩

9: Weekend Editor, “Post-Memorial Day Thought: 500k Russian Dead in Ukraine “, Some Weekend Reading blog, 2024-May-31. ↩

10: Weekend Editor, “Ukraine: 600k Russian Dead”, Some Weekend Reading blog, 2024-Aug-27. ↩

11: Weekend Editor, “Ukraine War: 700k Russian Dead”, Some Weekend Reading blog, 2024-Nov-04. ↩

12: Weekend Editor, “Ukraine War: 800k Russian Dead”, Some Weekend Reading blog, 2025-Jan-08. ↩

13: Weekend Editor, “Ukraine War: 900k Russian Dead”, Some Weekend Reading blog, 2025-Mar-21. ↩

14: Weekend Editor, “Ukraine War: 1,000,000 Russian Dead?!”, Some Weekend Reading blog, 2025-Jun-12. ↩

15: Weekend Editor, “Ukraine War: 1,100,000 Russian Dead”, Some Weekend Reading blog, 2025-Sep-22. ↩

16: Weekend Editor, “Ukraine War: 1,200,000 Russian Dead”, Some Weekend Reading blog, 2025-Dec-31. ↩

17: Weekend Editor, “Ukraine War Casualties: A Segmented Regression Model”, Some Weekend Reading blog, 2026-Jan-28. ↩

18: Weekend Editor, “Ukraine War: 1,300,000 Russian Dead”, Some Weekend Reading blog, 2026-Apr-02. ↩

19: Weekend Editor, “Ukraine War: 1.4 Million Russian Dead”, Some Weekend Reading blog, 2026-Jun-28. ↩

20: Weekend Editor, “Ukraine War: 1.5 Million Russian Dead”, Some Weekend Reading blog, 2026-Sep-11. ↩

21: Weekend Editor, “Updated R Script for Analysis of Russian Casualties in Ukraine, 2026-Nov-16”, Some Weekend Reading blog, 2026-Nov-16. You will also need the R pacakges PerformanceAnalytics and plyr, both available from CRAN.

The data driving the original analysis, in .tsv format, is also available for review by the persistent skeptic.

The data driving this analysis, in .tsv format, is also available for peer review.

There is also a transcript of running the script, for your peer review and skeptical scrutiny.

This whole mess depends on some idiosyncratic subroutine libraries of our own construction, graphics-tools and pipeline-tools. We are happy to supply these to our most exceptionally persistent skeptics who wish to reproduce these results. :-) ↩

22: T Hains, “Rubio warns ICC on weaponization of international law: ‘Stop quoting law to men with swords’”, MSN, 2026-Nov-15. ↩

23: I Verma, “Pete Hegseth’s Roman General Quote Comes With a Gruesome Twist”, Yahoo News, 2026-Nov-16. ↩

Published Fri 2026-Nov-20

Gestae Commentaria

Comments for this post are closed pending repair of the comment system, but the Email/Twitter/Mastodon icons at page-top always work.