From: kymhorsell@gmail.com   
      
   EXECUTIVE SUMMARY:   
   - We run a new s/w over UFO sightings of different types and see which   
     state demographics predict local UFO activity.   
   - Some factors occur in several UFO models. It seems state temperature   
     and distance from the sea figure in several key "types".   
   - Some key factors may relate to the behaviour of UFO witnesses.  But   
     almost all seem to relate "mostly" to the reported behaviour and/or   
     imputed interests of UFOs.  We can reason in some cases a factor can   
     not reasonably be associated with state residents in the way or to   
     the extent the stats procedure claims it does.   
   - Finally, we look at a couple of key UFO types and note some   
     interesting and some unpleasant associates they may have.  In many   
     cases the attributes found by the s/w using only relatively   
     un-imaginative data science techniques and a wide range of data are   
     the same suggested by many UFO witnesses and/or researchers over a   
     number of decades. Seemingly, no matter how crazy some of the claims   
     may sound, some of them seem to have some backup in the hard data.   
      
   The AI's have been chattering away and come up with some new takes on   
   the patterns of UFO sightings across N America.  They've had some new   
   tools to play with in the past few months and this report is based on   
   state by state "lasso regression" that validates the model found.   
      
   A lasso regression involves not only trying to find the "best fit" of   
   a line across a dataset to follow a given target variable (in our   
   case, here, some kind of UFO count state by state) but it also tries   
   to make the resulting model "as simple as possible" by also including   
   a special "regularisation term" to the fitting value to be   
   optimised. Normally this means as many coefficient as possible in the   
   model get set to 0 and only the "key variables" are left to explain   
   the target data as well as they can.   
      
   If that wasn't enough of a thing to grasp in one sitting, the   
   "validation" part gets added on. It turns out just statistical tests   
   are not always the best way to establish a model is "good".  It may   
   appear to fit the data very well (typically meaning the so-called   
   "explanation power" statistic -- the R2 -- is reasonably close to 1)   
   yet have no predictive power at all. A new set of data -- despite the   
   claims of statistical tests -- will fail to be well-estimated by the   
   model usually because "something changed" between the data you had   
   when the model was created and the time you got the new data to use to   
   make a new estimate. To better ensure a stat model will work with new   
   data at some time later we have to at least run a test for that one   
   time.  And the way most stats-heads and s/w packages do it is to hold   
   some of the data aside when building the model to test whether it   
   actually explains the "hidden data" as well as the data used to build   
   the model.   
      
   So in the procedure used here the model is created using just a random   
   1/2 of the data to hand, then it is tested against the other 1/2 of   
   the data and if both halves are reasonably-well explained by the model   
   then it is accepted as have not only statistical robustness, but   
   predictive power.   
      
   So these are the models we will now look at. Using all the state   
   demographics and other data to hand we will try to build robust and   
   predictive models for UFO activity of various types over each state.   
   In the UFO dataset I normally use -- the set from the NUFORC -- the   
   types of sighting are normally broken down by "shape" of the object or   
   objects seen. The NUFORC have a long-established set of shapes and I   
   normally follow those fairly closely. Some new shapes tend to get   
   added over time, following "fashion sense" among witnesses.  E.g. a   
   "tictac" was added at some point. But I also have chosen to sometimes   
   add yet other shapes my own fashion sense dictates.  E.g. I look in   
   the comment section of reports for certain keywords like "faint",   
   "bright", "cube", or "pyramid", among others.   
      
   And so we will look here at how a series of validated lasso   
   regressions explained sightings of a collection of UFO types across N   
   America.  Viz:   
   	pink   
   	purple   
   	Egg   
   	Disk   
   	Oval   
   	Circle   
   	Sphere   
   	bright   
   	faint   
   	gold   
   	red   
   	yellow   
   	black   
   	Light   
   	notLight   
   	Triangle   
      
   In general those keywords starting with a capital are NUFORC's chosen   
   shape names while the lower-case words are ones I've chosen for one   
   reason or another to locate in the comment section of reports.  They   
   are typically colours but also can indicate other things like "bright"   
   and "faint". Most of these can be applied to Lights type UFO's.  But   
   they also may apply to lights seen *on* some other kind of UFO.   
   Typically a witness reports whether there were lights on the bottom of   
   some otherwise dark object, or sometimes they see what they think are   
   "port-holes" or "domes" on object through which some coloured light is   
   visible as well. And, finally, the notLight category is all of those   
   UFO's not in the Light category. It make sense to split up what people   
   typically see as "little lights going across the sky at night" and   
   everything else they might see typically in the daytime and oftentimes   
   reasonably up close and personal.   
      
   There are of course many other shapes and colours. But these are the   
   boundaries we're using today. :)   
      
   I currently have 915 state-by-state statistics. These cover a very   
   wide range of things including social factors, economy, environment,   
   weather, medical, law, crime and many other things.  All of these are   
   thrown into the mix and all the above UFO types are cross-checked   
   against all 915 factors and the best ones "lassoed" and validated.   
      
   We then can check to see whether some factors are common across all   
   the selected UFO types. And there are a few.  Ignoring the factors   
   that only apply to one type we find the common factors are:   
      
   Variable	Number of UFO types with this var as a relevant factor   
   		in NUFORC sightings since ~1950   
   coffeerank		5   
   maxtemp			5   
   murder2019pc		5   
   coastdist		4   
   diversity		3   
   infgasgatherint		3   
   medage			3   
   birthrate		2   
   carspc			2   
   casespc20210408		2   
   cattledeathspredpct2015	2   
   coalconsint		2   
   coviddailycasespc20211121 2   
   dope			2   
   evpct			2   
   lung			2   
   maskapr2021		2   
   minpre			2   
   missingdoenetworkpc	2   
   pctgenrenew		2   
   pctnonwhite		2   
   planeaccidents-personal	2   
   pre6			2   
   suiciderate		2   
   totdeathspc14feb2021	2   
   under18			2   
   usoutagesvandalismpc	2   
   wind			2   
      
   Hopefully most of these things are self-explanatory. Some factors are   
   related to conditions in the US at a certain time. Some are factors   
      
   [continued in next message]   
      
   --- SoupGate-Win32 v1.05   
    * Origin: you cannot sedate... all the things you hate (1:229/2)