In grad school, my office had an ordinary Edison incandescent bulb, kept on 24/7. Being probability nerds, we kept statistics on when it burned out. We had been expecting something like an exponential, or a bathtub, but it was much closer to Gaussian, with a mean lifetime around 1 month and (rough) standard deviation on the order of a few days.
A Poisson variate always has mean = standard deviation. If these things lasted a month +- a few days, they couldn't be from a Poisson distribution.
(Or, with no calculation: The thing about a Poisson distribution is that it's what you get from a memoryless process. If the bulb almost always burns out within, say, a week of reaching 1 month, then it should also almost always burn out within a week of first being switched on. So its mean lifetime can't be much longer than the typical variation.)
It sounds to me more as if lifetime was proportional to some physical characteristic(s) of the bulb that were controlled in the manufacturing process. For instance, maybe lifetime was proportional to thickness of filament or something of the sort. (In which case, the vendors could have made the bulbs last longer, but perhaps only by also making them more expensive. Though it's tempting to speculate that they were designed not to last too long, so that you had to buy more.)
It fails when a certain proportion of the filament has evaporated.
The evaporation rate is pretty constant, assuming either a getter or large enough bulb that the evaporated tungsten doesn't change the environment.
So you would expect some form of bell curve distribution around the design lifetime.
Bath tub curves tend to come from large assemblies of reliable components
Yes, the results are consistent with that. Whatever stochastic effects there were, like electrical fluctuations or materials variations, were negligible.
http://www.centennialbulb.org/