The relative effectiveness of the allied and axis armies in the ETO

[JonS]

On the statistics discussion, there are two things going on. One; people are misunderstanding what other people wrote. And two; people are enjoying the sound of their own keyboard.

A CEV difference of 1.2 is quite small, and fairly easily overcome by modest additions of reasonable forces. That, however, is not in dispute. Writing 1200 words on a topic not in dispute is an interesting use of time.

The armed forces of any nation are not uniformly capable (or incapable). Some will be better in easy to measure metrics such as supply and equipment. Some will be better trained than others. And some will be better in more intangible and difficult to measure dimensions such as leadership (with everything that means, including better planning, better CA integration, better decisions under fire, etc), cohesion, and morale.

On average, a CEV for the Germans of 1.2 provided the best fit for the dataset being used. However, within that data set the CEV for any given nation in different battles would vary quite a bit. Sometimes higher, sometimes lower. How much higher or lower we don't know, because it isn't reported. But all we get is the bald 1.2 CEV. If confidence intervals were reported, given what is involved, I expect that the CEVs for the UK, the US, and Germany all overlap each other, probably by quite a lot. In other words, I am asserting that the variation between the best and worst German forces (or US, or UK) would be higher than the reported variation of 1.2 between the Germans and the US.

That would mean that while the shift of the German CEV band to the right suggests they were 'better' tactically (and only tactically, mind) it is at best a suggestion, not a statement. This should be fairly uncontroversial, as a thought experiment pitting say the 82nd AB Div against the 71th Inf Div should show, or books like 'When The Odds Were Even' attest.

And that's aside from the issue of the dataset not being a random selection.

[JonS]

People overestimate the importance of the middle east for oil in the WW II era.

ME oil was pretty important, as far as the British were concerned, for the RN and Middle East ground forces. A lot of the oil and petrol used in Burma came from the ME too, along with shipments to the USSR. The ME was less important on the global scale, though, since forces in the UK got most of their oil from South America, or later from the US.

While small compared to production in the Americas, securing ME oil would have been a massive fill up for Germany (putting to one side the utter unreality of actually moving any of it to Germany)

Agree with the rest of the post.

[JonS]

For example take a look at the 7th Armoured Division - very different effectiveness levels in North Africa compared to Normandy.

The 7th Armoured Division is an interesting case study. Heginbotham, in his 1996 thesis, uses the 7th as an example of the instability of British divisional organisations, comparing it unfavourably with US practice in which once a division was created, it kept it's organic components together. And he's right - the 7th was highly schizophrenic as battalions and regiments came and went, and the division repeatedly reorganised to cope with losses and as it searched for an optimum structure.

But Heginbotham's example is problematic in two ways. Firstly, the 7th is waaaaaay out at the end of the bell curve. There is no other British division that even comes close to it in terms of length of active service, the number of reorganisations, or the number of battalion-sized units that came and went. Secondly, if we're going to cherry pick, there are plenty of US divisions who suffered the same kind of instability in terms of division structure. The US parachute divisions are obvious examples, but the US infantry divisions also saw plenty of battalions coming and going too. Not the nine core infantry battalions - they were generally static, but the attached tank, tank destroyer, AA, etc bns came and went with alarming frequency.

Anyway, in terms of your effectiveness example - the 7th in North Africa, vs the 7th in Normandy, the problem is that they were really very different divisions. About the only thing that remained constant was the name (and even that changed), which makes using it as an example a bit problematic.

[sburke]

Accountants make jokes about how boring actuaries are...

And then there is the exciting life of a Chartered Accountant

[JasonC]

"the CEV for any given nation in different battles would vary quite a bit"

This clause suffers from a misperception of its own concreteness.

CEV is a term in an equation, and in that context it has an actual, definite meaning. It is a coefficient in a linear model that minimizes a least squares error. It splits those errors better than any other choice for a different value for the named parameter in that equation. What you think it means otherwise, as a variable rather than a coefficient, remains a mystery, but it certainly isn't "CEV".

Who actually won each fight is a different matter.

You can posit an entirely imaginary CEV that was five times as high in fight A, but offset by a similarly unusual coefficient for air power or terrain that happened to be only one fifth its value in every other fight with similar terrain - and so on in any variety you please. But it is merely misunderstanding the word "average" to do so.

You can posit that in one of the 10% of fights that the model misclassifies, that the reason the outcome was different is that CEV was actually 2 or 0.6 in that specific fight. You can equally posit that terrain was different instead, or air power, or the relative weight to give to tanks rather than artillery, or whatever. Since every single one of the terms used in the model is being used in the same "average", not varying battle by battle, sense, it is completely arbitrary to pick one of them and regard it as the sole explanation of the residual unexplained outcome, around the model expectation.

In other words, CEV means average CEV or it doesn't mean anything; that is where the term came from and what its mathematically describes.

But if you hypothesize that CEV is random with a noise five times the size of the indecisive result band (mean 0, plus or minus 50%), then you won't get a big improvement in model fit by tweaking it by a 1.2 factor. That 1.2 factor makes 2/3rds of the model's mis-classifications disappear. A noise term doesn't do that - it would hurt you as often as it helped. Even with a mean that isn't zero, a very noisy underlying driver would not correct 2/3rds of your misses without a cost on your previous hits.

In short, the evidence is strong, not weak, that weapon for weapon the Germans fought marginally better than the Allies in Italy in 1943. We know that from reading decent operational histories, before we try to reduce it to math. It is not a strange conclusion or a difficult claim to believe.

And the reason that conclusion is resisted as much as it is, isn't the math.

[JonS]

In other words, CEV means average CEV or it doesn't mean anything; that is where the term came from and what its mathematically describes.

No kidding. And an average has a standard deviation. Which means the CEV for each nation lies within a range. And if those ranges overlap then you can only suggest things, to a certain % confidence. You cannot state them as absolutes.

[Broken]

People overestimate the importance of the middle east for oil in the WW II era. Saudi Arabia had started its development, true, but the first oil even found there was in the 1920s and investment there was still in its infancy. The world oil powerhouse was the US, which alone produced about half the world's oil in 1941. Anyone remember John D Rockefeller and Standard Oil?

You really need to write a book or seven, Jason. I don't always agree with you, but your prose is always highly readable. Better than Paul Kennedy's "Engineers of Victory" I am currently reading.

[JasonC]

JonS - that's fair. More precisely, we'd also like to know the standard error, T-statistic and p-value of the CEV line of the statistical model, not just the fitted value of the coefficient.

Broken - thanks for the kind comments. Now can I have more time, please, to write everything I'd like? OK, maybe not your department...