View source for Models that are so vague they are barely an idea

{{stub}}

===Bathtub curve===

The bathtub curve refers to a theory about the likeliness of product failure, 
particularly applied to electronics but potentially to anything made of many components.


'''The idea''' is that 
: there is a '''certain likeliness the whole will fail almost immediately''', often due to a randomly faulty component.
: after this possibility of immediate failure, the '''probability stays low''', as components and design should last a few years.
: After maybe five years, the likeliness of failure '''slowly starts rising''' again due to wear, and minor wear/quality differences between components
:: if various critical components are rated for the same length of service, the probability of failure would start curving up noticeably after that interval
<!--
(on the order of magnitude of two to five years for most electronics) 
-->


'''"Why the name?"'''

If you plot a curve that drops rapidly, stays low for a while, then rises more slowly. 

This looks roughly like the profile of a one-person bathtub, with the curve one one side that you can lie on. Hence the name.



'''"Is it true?"'''

The bathtub curve is an interesting-sounding theory, but is it true?

Vaguely.
Depends.
Too crude to use as a general model, though.


Its two often-mentioned ingredients are based in something real, 
but that does not imply that either is used accurately, 
nor that the combination is complete.



Prime reasons it might be useful is often about estimating maintenance and replacement cost.

Say, when industry starts to use many copies of the same device,
you know you'll be needing to do more maintenance overall,
so you start doing calculations like [[Mean Time To Repair]].

That calculation will work better when you consider what all your critical components are,
and what their typical lifespan is.


At the same time, this reminds us we need to think about complexity. 

Even if the failure rate of each component is an extremely curve up so gentle that you can't see it for the first ten years, the risk of the device containing ''many'' of them is roughly all those curves multiplied.
This makes the overall curve both higher, and curve up faster.

So just describing that shape doesn't have a lot of predictive power.

Especially since the shape ''changes''.

<!--The most you can get out of that is an intuition such as 
"even if most components are good for ten+ years, 
use a lot of them and chances are you'll be fixing something years sooner" --
but doesn't tell you how to estimate long term maintenance for your boss.-->



Another thing it seems to point out is that, in certain fields, companies may have a Dead On Arrival policy 
(meaning "if it doesn't work and you send it back immediately, they'll just send out another"),
which is based on some cost-benefit analysis saying that 
it is cheaper for them to occasionally send out a complete replacement,
than to spend a ''lot'' more time testing each unit thoroughly for more exceptional faults.

...while in other fields, e.g. where things relate to safety,
that same approach is the best way to quickly stop existing as a company.


<!--
How true is it?

That may vary per field, manufacturer, seller, ''reseller'' (if any), and product.


In more detail, this supposes approaches like 
* that manufacturers have done a cost-benefit analysis where checking every component's history of quality is just not worth the effort, and as long as most products are fine
:: specifically, replacing a small number of them is cheaper than extremely tight control
:: basic QA will always be and stay a thing to guard this calculation

* sellers have done cost-benefit estimations and figured that if someone complains, it is worth replacing it with only a minimal check against fraud.
:: primarily because the time invested adds up

* ''re''sellers have done cost-benefit analysis how (a specific implementation of) 'buyer' protection


All this analysis ''itself'' varies with many things, e.g.
: how much they care about brand, and people sticking with it, 
: how costly the product is,
: how likely the repair is.
-->



<!--
The slow curve up is also related to why 'extended warranty' ''can'' be... not quite a scam, but frequently not worth it for consumers.

It depends a lot on the product, as the extra two years still lie in relatively safe territory - it's the difference between "three years warranty" 
altered to be "one year warranty, pay more for a expansion to three years isn't that great".

Chances are that the company did the math - that most devices will not fail before the extended warranty ends.
Is that occasionally going to give a few people something nice? Yup. 
Is the company going to lose money on that? Unlikely.


And sometimes there are other estimations.

For example, say that platter hard drives tend to last three to six years of powered time. 
For backup, or for hosting companies, that's not great, plus you want to know that and a lot more.

But for most of us? We tend to buy either a new computer or a larger hard drive after four years or so, 
and after that it falls in a mental 'it's not so useful / it was old' category and we don't consider failure that bad. It is useful to consider for long-term data storage.

-->



====See also====
* http://www.itl.nist.gov/div898/handbook/apr/section1/apr124.htm
* http://en.wikipedia.org/wiki/Bathtub_curve

===Gartner hype cycles===
<!--
The gartner hype cycle is a curve that looks like this, plotting expectation over time:



The idea is that when particularly technology is hyped, 
people promise to much and people expect to much, whether it is from that promise or their imagination, 
people get disillusioned and it goes from a peak of expectation to a crash,
until it slowly settles on the realistic level of what it can actually do.
The names on the curve even suggest wording like "until some more realistic people figure out what it can ''actually'' do".




There are various problems with this curve

* plotting things on this curve, anywhere except at the end makes some very strong assumptions about how useful something will be. If we knew that, we wouldn't need the curve. This isn't even self-fulfilling, it's just a wild guess.

* you are also assuming that everything settles at medium-useful
:: there is a ''strong'' selection bias ignoring everything that was hyped ''and then disappeared completely''

* this curve suggests multiple cycles won't happen and suggests that the second time right we ''will'' get it right
:: but e.g. we have already gone through two AI hypes and disillusionments. We are currently on a third hype, with strong indicators of another disillusionment. Anything truly practical done with the things are on the sidelines
:: people are very good at MS paint curves, ''but they all draw different ones''
-->


[[Category:Explanation]]


===Dunning-Kruger curve===

<!--

If you do an image search, you will find a plot, usually with competence on the X axis and confidence on the Y axis.

There are a few different takes on this plot, but it usually goes something like:
: that when you start learning you suddenly have new skill that you feel can do everything with,
: when you learn more you realize there was a lot missing and your confidence goes down,
: and as you get better you will eventually, slowly, get confident with broad knowledge


This resonates with a lot of people, because everyone has met some incompetent but also arrogant.
The peak in the curve is sometimes called mount stupid.  Yeah, those jerks.

Or it resonates with a feeling of "learning more you often realize how little your knowledge covers"


There's a whole spiel which creates a journey metaphor, where 
: the initial confidence is 'mount stupid',
: there is a 'valley of dispair' where you learn just enough to realize you were wrong but not yet past it,
: a 'slope of enlightenment' where you slowly build actual skill,
: and a 'plateau of sustainability'



It's also not what Dunning and Kruger wrote about, a decade before that plot ever emerged.

Dunning even pointed out that this plot is a not from their work, and a misunderstanding of their work.


It's a meme image. Its shape has little relation to what they measured - it would take some creative, opinioned re-imagination,
and that is where this curve's specific-seeming shape seems to come from: a creative fabrication - it's certainly not from their paper.


The paper's has a plot, but it is quite different: self estimation versus test scores, and found that people new to a topic overestimate,
there's a balance point, and the best scorers would structurally ''underestimate'' their score a little.

Instead of a curve, it is actually pretty linear. This suggests (but does not prove)
that people's confidence and skill are actually broadly related to skill (less if less, more if more) and maybe even linear-ish.
This alone sort of disproves the 'stupid people are arrogant' read.


It would be just as easy -- and probably just as incorrect -- to conclude that
people are perfectly realistic about their own confidence but just like to score things around seven,
as that would shift up the curve and make it look like some people are overconfident.


The paper(s) mainly points out there is a dual burden in the estimation that the test asks for:
if you don't know the answer, you are less likely to answer correctly, ''and'' less likely to estimate whether you got an answer correct.

That's it. 
Completely regardless of their attitude about it (and regardless of whether you call that metacognition, or just a numbers thing),
this alone would do a ''lot'' to produce most of that graph.


And yes, like any questionnaire style study, there are other effects entangled into this graph, 
to unstudied degrees, perhaps including 
focus on negative more than positive memory, 
optimism, 
pessimism, 
impostor syndrome,
and sure, a boastful overconfidence.

But that's the thing:
even if that's true, that's not at all shown by the paper. 
That's us reading into it what we want to. 
That's our biases.


'''Especially''' if you go on and use this to talk about
certain kinds of stupid people being poor performers ''despite'' not being beginners.

Because the effect seems to persist in any field new to any people - so it's about starters's current skill level, not their intelligence.



(it even persists if you tell people about the effect before answering)


See also:
* J Kruger, D Dunning (1999) "Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments"

-->