Side note: I understand "inherent" as the average level of somethin and not as "despite" but as "independent of" something else.
Main answer: As the influence of training is widely accepted, determining the inherent level of aggression for different breeds would need a study where the pairing of dog and owner is properly randomized because a strong positive correlation between certain types of owners with agression levels is to be expected. I don't know proper literature about this. And I don't think it likely that anyone puts up the money for such an expensive study of a question that is mainly academic, because the practical danger as a function of the breed is probably of far more practical interest, and I think that is answered far more easily (see below).
I did find a few studies about breed-specific aggression levels
- a dissertation in German Mittermann 2000: Assessment of Behaviour of Dogs of the Pitbull-type and Five other Breeds by Temperament Testing According to the Guidelines of the Dangerous Animals Act of Niedersachsen, Germany (GefTVO) of 5th of July 2000
The dissertation has a total of 415 dogs which fell into categories 1 (American Staffordshire Terrier, Bullterrier and dogs of the Pitbull-type) and 2 (Rottweiler, Doberman Pinscher and Staffordshire Bullterrier) according to state law. The study found overall 20 of 415 animals with inadequate (or disturbed) aggressive behaviour. At that time the legislation treated these two categories differently, legislation required per-dog assessment of behaviour. The thesis cites 4 studies finding hereditive aggressive behaviour, there are physiological reasons like density of opiate receptors (p. 33). This can be changed within few generations by breeding (compare the concerns of the American Veterinary Association wrt. legislation needing a stable basis; paper linked by Kathy). Section 22.214.171.124 discusses breed-specific disposition (i.e. breed-specific high risk for particular types of inadequate behaviour). Breed specific differences were observed (e.g. 1/68 Staffordshire bull terriers 0/38 bull terriers vs. 8/93 Am. staffordshire terriers, 4/63 pitbull type and 5/56 dobermans with >= inadequate agresssive behaviour). These differences are not significant at 5% niveau. But American staffordshire/doberman vs. bull terriers / rottweilers were around 10% significance niveau. Conclusion is that the categories do not warrant the legislative differences (note dobermans in the lower category and bull terriers in the higher!).
This study has the dogs handled by their owner (which is sensible as the legal test should show that the does is not inadequately aggressive when handled by its owner; but this may distort breed-specific influence).
- Ott 2008 compared those results to 70 golden retrievers, of which 1 showed inadequate aggression and concludes: not significant.
- Ott 2009 and Schalke 2010 compare 38 bull terriers of a blood line that was suspected to be more aggressive with the results of the previous 2 studies. None according to Schalke and 1 according to Ott (they are the same group) displayed inadequate aggression. No significant differences.
From a statistical point of view, the problem is that the differences would need to be huge in order to be significant with these sample sizes. The 95% confidence interval for the proportion of 15 dogs with inadequate aggressive behaviour of a total of 415 dogs is ranges from 3 to slightly above 7 % (point estimate: 4.8%). 1 out of 70 golden retrievers: 1.4 % (0 - ca. 7 %). The point estimates are more than a factor 3 apart, but it is not significant (p is about 0.10).
Or, huge numbers of dogs need to be tested. Significantly less aggressive animals at 5% niveau would mean: finding a breed of which 1 of 141 show aggressive behaviour vs. 20/415 (χ² test - that is an approximation that should work OK for these sample sizes).
The breeds with highest proportion of inadequately aggressive dogs were the dobermans (5/56) and Am. staffordshires (8/93). That is almost every 10th and 11th dog showed inadequately aggressive behaviour. Pooling these two breeds (for which I don't see any particular reason) would give 13/149 as reference for highly aggressive breeds. Significance at 5% niveau is reached with 0/105 or 1/134 animals.
Compared to the Am. staffordshires, the proportion of inadequately aggressive golden retrievers was about a factor 6 lower (8.6 vs. 1.4%). Sample size estimation at β=20% and α=5% (i.e. 1 out of 5 such studies missing the fact that another breed shows only 1/6th of the proportion of inadequately aggressive dogs) needs about 200 animals of the other breed. Or 134 dogs if there were (truly) a factor 10 between the different proportions inadequately aggressive dogs.
Now, the dissertation IMHO is a large study (considering the "experimental" effort). However, it is clear that the sample size per-breed is too low to detect reasonable differences in breed specific behaviour.
- hereditary aggression characteristics can change within few generations (Mittmann, p. 35 [giving further references]). Inadequate selection (too fast breeding, breeding according to outer characteristics without selection of mental characterists) can cause particularly fear-related aggression for fashionable breeds.
On the other hand, the fact that any breed of dog can be trained for low aggression, does not answer the question of breed specific dangerousness, neither. One would need to study how much training is necessary on average to get to low/high aggression levels for different breeds and then account for the damage they are likely to achieve.
So, I'd expect that different breeds have different levels of aggression and "dangerousness". In order to study this, I'd assume that such properties are related to target properties for breeding, and choose a comparison where a large difference can reasonably be expected. E.g. low agression potential for rescue dogs (e.g. Newfoundland; though note that today's breeding targets often developed far away from the working aspect, so that e.g. today's St. Bernhard dogs are usually not considered fit as working dogs any longer, so are German shepherds). Higher bred-in danger for working breeds that were/are used to guard herds also without human supervision against predators as well as theft (aggression potential against unknown humans, size & strength to be effectively dangerous enough to guard the hers).
That would probe the "hereditary" part in Kathy's answer. And the last two sentences in ratchet's answer (whose literature tells us to expect higher variance also for certain breeds). So the meta analysis of answers already given points strongly to a "yes". (Jolanta's answer deals with a different aspect of dangerousness which I'd not consider "inherent". Those findings can be in accordance with higher inherent danger, see below).
However, I doubt that incidence statistics (of any kind) allow to "measure" danger levels inherent to the breed. The reason is the already mentioned possibility of a positive correlation with respect to agression. It is quite possible (and IMHO probable) that certain dog owners who prefer an dangerous dog will both tend to have breeds that have the reputation of being more dangerous and also train them to exhibit more aggression rather than train them for low aggression (the other part of being dangerous, the level of damage the dog can achieve does depend on the breed. One bite of a St. Bernhard's can cause far more damage than a bite of a chihuahua - though ultimately a newborn may die from a chihuahua's nip into the nose).
The answer to the "Is breed a valid method of determining relative safety of a given dog?" sub-question may be yes (personally I think this very likely; but I cannot give "hard" scientific evidence; the studies cited above did not find significance at 5% level [discussion whether p-values are appropriate are off-topic here], but they also found a factor 6 between the proportions of dogs handled by there owner with inadequately aggressive behaviour between Dobermans or Am. staffordshire terriers and golden retrievers (or bull terriers)).
The breed of dog may be (personal expectation: is) a good predictor for the aggression level of the dog owner. Meaning that there is a positive feedback loop between alleged or existing aggressiveness and dangerousness and training target characteristic. Which makes the answer to the subquestion a stronger "yes".
Again, I have no hard evidence. But detailed studies for dog ban policies could give hints in favour or against this possibility:
If this were the case, one would expect
- that aggression studies on dogs from the breeds that fall under the typical ban policies would in fact be found to be aggressive on per-animal evaluation
- yet a dog ban on the basis of breeds should have only a small effect (if at all) on the overall danger of the dog population (incidents) as the potential of aggressively training dog owners is hardly affected, and there are enough dogs that can easily be trained to be aggressive.
- that banning of dog owners may have more impact, but I think it questionable whether anything the like could be practically enforced. So, again, this would be hard to observe.
Some more thoughts:
Huge critique point with the linked article: as there is not even an attempt to give approximate proportions of the respective breeds within the total dog population, the "danger" of a specific breed cannot be inferred.
On the other hand it is also quite clear that pit bulls (whatever they specifically are) are far less than 60% of the total US dog population.
Huge difficulty with the whole question: there are enough different (more or less) sensible definitions of danger that every opining can find supporting studies.
In Germany, about every town collects dog tax. For such countries, the relative frequencies of different breeds may be determined with enough precision for practical purposes. So the relative-to-the-general-dog-population difficulty can be addressed at least in some regions.
In general, googling news for sure isn't proper research. In the particular question of fatal dog attacks, however,
- I think that the risk of missing any fatal incident is low to negligible
- I also think it a fairly safe assumption that basically every fatal dog attack is reported.
- I also think that there are few enough fatal dog attacks that it is fairly easy to sort them out (i.e. not confusing two to be the same; nor double-counting)
- and I'd expect that it would be big headlines if the same dog is repeatedly responsible for fatal attacks.
Put together, I think fatal attacks may be treated as surrogate markers for severe aggressive incidents - under the usual caution you need with any kind of surrogate marker.
(more literature to come when time permits)