From: [d--i--z] at [ix.netcom.com](Dan Z)
Newsgroups: talk.politics.guns,rec.humor,alt.tv.snl
Subject: Re: Saturday Night Live Special?
Date: 2 Jun 1998 15:33:27 GMT

In <[3573568 c 374972] at [news.pacific.net.sg]> [wysiwyg 000] at [hotmail.com]
(Raymond) writes: 
>>>Gee, really Ray?  My mother has an old shotgun gathering
>>>dust in the back of a full closet, and she doesn't own any
>>>ammo.  No one in town even knows she owns it (other than me),
>>>and she lives alone.
>>>
>>>So tell us, Ray, by what magical mechanism does her risk of
>>>homicide increase by 2.7?  We await elucidation.  Either that,
>>>or we await an admission that you're talking nonsense.
>>Actually, if there was no risk of homicide to start with then
>>multiplying it by 2.7 gives a grand total of no chance
>
>Shotguns also make good clubs.  Still 2.7 times


Now on to the claim that having a gun triples your homicide risk.

They supposedly determined this by going to the homes of people in a
three-urban-county area and interviewing the survivors of homicide
victims who had died in or around their own homes.  They then found
randomly selected people who lived within a few blocks and who were
similar to the victims in age, sex, and race and interviewed them
similarly, testing a few dozen different types of variables (living
alone, drug use, etc.).  They then compared the statistics of the
homicide victims against the statistics for the supposed control group. 
They discovered that firearms were somewhat more common in homes where a
homicide had taken place than in the control homes (44% vs 36%).  Note
that this is the *only* indisputable hard fact in the study.  Everything
else is conjecture based on statistical massaging and "it seems
reasonable to us" conclusions stemming from that one sole original fact.
 Note that saying "an NEJM study found that people who were killed were
somewhat more likely to own firearms" doesn't have much punch, however,
and brings to mind too many "what ifs" which don't support the admitted
anti-gun positions of the author and publisher.  So they worked on that
statistic a while, and after some statistical massaging, they claimed to
have demonstrated that homes where firearms were present were 2.7 times
as likely to have a homicide (usually reported as "three times likely").

Problems:

1.  Correlation does not equal causation.  As an example, the same study
determined that renting the home has an even greater correlation (4.4)
to homicide than guns (2.7).  From this, is it safe to conclude that if
you purchase your rental home from your landlord, your homicide risk
suddenly drops to a quarter of its previous level?  No, of course not. 
Obviously, there is another factor involved, probably that people who
rent are poorer than people who own homes (or less responsible, or
whatever), and this has a lot to do with their lives, their jobs, their
neighbors, their habits, and the kinds of things that make someone a
higher homicide risk.  Likewise, by far the most likely explanation for
the gun correlation is that violent people are more likely to acquire
guns, and at the same time violent people are more likely to become
homicide victims or kill someone else in the home.  Another example:  A
study conducted exactly as this one was which examined drowning victims
and compared them to randomly picked controls would turn up the
interesting fact that life jackets are highly correlated with drowning. 
Would this mean that buying a life jacket will increase your risk of
drowning?  No, exactly the opposite.  It means that people who drowned
were much more likely to be boating or water-sport enthusiasts than the
landlubbing control group, and thus were more likely to own life
jackets.  Now, does this mean that people who live in Arizona who go out
and buy a life-jacket just for the heck of it are suddenly more likely
to drown?  No.  Does this mean that people who fish for a living will be
safer if they sell their life-jacket?  Hell, no.  In short, and as even
Kellermann admits, the study has only shown that guns are a "marker" for
homicide, which means that if you chose a home at random and found a
firearm already in it, it may indicate that something about the home
(perhaps the type of people who live there) is (only somewhat) more at
risk for homicide than a randomly- chosen house which didn't have a
firearm.  It says *nothing*, and this type of study is not *capable* of
determining, that going out and buying a firearm will in any way change
the risk of homicide for your home.  It might even help, as in the life
jacket example, but again this kind of study would be incapable of
determining that fact.  See "Designing Clinical Research" by Hully and
Cummings, Williams and Wilkins, 1988, chapter 8 and appendix 8.B, which
deal with case controlled studies.

The real irony here is that with the kind of correlation observed in
this study, it is in theory possible for acquiring a firearm to always
*decrease* the risk in a given home, yet for the study to still show
firearms to be a marker for *increased* risk.  As an example, consider a
drug which is used to treat violent people and which *always* makes them
less violent, although it works very well on some patients (making them
almost as non-violent as an average person) and works not so well on
other patients (making them less violent than they were without the
drug, but still dangerous).  A case control study of the type under
discussion would find that violence (and homicide) occurred more
commonly in homes in which people were taking the anti-violence drug
than in homes where people were not taking the anti-violence drug.  Why?
 For the obvious reason that violence occurs most often in homes where
violent people reside, and these are the very homes where the
anti-violence drug would likely be in use. The general population being
used as a control group would have almost zero instances of the
anti-violence drug being used, so the correlation would show that the
anti-violence drug was a strong "marker" for violence, EVEN THOUGH IT
DECREASED RISK IN EVERY HOME IN WHICH IT WAS USED.
 
  [Read that last sentence again; it's crucial, and totally invalidates
   Kellermann's key assumption in his study -- that if the factor
   being studied is actually a risk-decreaser (i.e. offers a protective
   benefit) then it will not be possible to find it correlated with an
   increased risk.  Kellermann's entire premise is fatally flawed.]

2.  Besides the third-factor causality mentioned above, there's
a possible direct reverse causation.  People who are aware that
they are at a greater risk for homicide (bad neighborhood, risky
job, psycho acquaintances, violent spouses or ex-spouses) are
more likely to buy firearms for self-defense than the average
population.  Note that the NEJM study didn't study only people
who had been killed by the firearm in their own home, but people
who had been killed who happened to own a gun, period.  Thus,
the fact that the homicide homes were more likely to have guns
than the control homes may just be an indication that many of the
people who were killed *knew* they were in danger, and thus
bought guns.  Again, the NEJM study did not attempt to determine
the circumstances of the homicide or the gun ownership.

3.  The two possible causations listed above (violent people are more
likely to get guns and kill, threatened people are more likely to
get guns and be killed), along with the old "if a gun's handy you
might be tempted to use it", as well as other factors, all certainly
contribute to the correlation to some degree or other.  The question is
how much does each contribute?  Unfortunately, the study was not
conducted in such a way that we could decide the importance of the
various factors, and any attempt to conclude that the very act of
acquiring a gun will increase one's homicide risk are totally unfounded.
Common knowledge strongly suggests that it won't, and that the
other causal relationships account for most of the correlation.
99.98% of gun owners will not misuse their guns this year, which
suggests that the "have a gun and you'll go wild" idea just doesn't
happen with any frequency.  It *is* common for violent people (criminals,
gang members, drug dealers, etc.) to acquire guns, and it's also
common for people who feel in imminent danger of homicide to get guns
for self-defense, so it's not a stretch to believe that these are
the factors which best explain the correlation.

4.  If "getting a gun will turn an argument at home into a shooting",
why did the study discover that people who live alone were even *more*
likely to be killed (3.7 correlation) than people who did not live
alone who had a firearm (2.7)?  Again, it looks like some other social
factors are the true cause of the higher risk for the various groups,
and not the act itself of living alone, renting, or owning a firearm.
The social factors which encourage homicide result in the living
alone, or the renting, or the owning of the firearm, not the other
way around.  Likewise, the study found that previously getting into
a fight in the home (4.4) or using illegal drugs (5.7) were more
associated with homicide than were firearms.  Does this mean that if
you pick a fight once or smoke one joint you'll suddenly increase
your homicide odds?  No, again these things are indicators of
something else wrong in the household (like unstable people), not
direct causes of the homicide.  It's important to note that the
great majority of all homicide victims (that's *victims*, not murderers)
have a history of drug abuse, violent behavior, or a criminal record.
Undoubtably, the same is true of the homicides examined in this study,
so the applicability of the study's findings to the average, normal
home is questionable.

5.  The "3:1" figure only came out after a great deal of statistical
massaging, in which a "model" of relationships was created by the
study's authors, and a mathematical attempt (based on the model's
assumptions) was performed in order to separate out the various
(possibly interdependent) factors.  The "3:1" figure is much higher
than what the raw data indicates (the raw data supports 1.6:1 instead
of 3:1), indicating that "separation" of the measured risk factors was
necessary to pump the figure that high.  This separation relies on
the assumptions of the model in question, and the fact that the model
"determined" that problems with alcohol were *not* a risk factor (despite
the fact that in the raw data various problems with alcohol were
high risk factors in the range of 7-20:1), indicates that something
is seriously wrong with that model -- alcohol is a known facilitator
of homicide and other kinds of violence.  Furthermore, the high risk
associated with alcohol in the raw data had to be apportioned
somewhere in the final results, and would cause the other, "separated"
risk factors (including gun ownership) to register artificially high.

6. The very concept of "separating out" sociological risk factors is 
probably flawed from the start.  Trying to separate out the various factors
in order to come up with numbers that could be said to meaningfully capture
what would happen if "all else were equal" is perhaps impossible to
do in a meaningful way.  I don't believe that you *can* do so.  The
best that could be done would be to break all the cases down into somewhat
comparable groups and try to extract some meaning from each one, the
results of which would only be applicable to that subgroup.  I'm afraid
that doing numerical massaging on a lumped group that likely includes the
deadly results of mob hits, gang wars, drug dealing, domestic disputes, the
acts of deranged individuals, people killed in self defense, and innocent
crime victims is likely to produce numbers that are inapplicable to *all*
the subgroups, not to mention the general population.  It's like trying to
find vehicle accident risks for a group that includes bus drivers and race
car drivers, when the only thing you do to try to factor in their driving
styles is ask them whether they've ever exceeded the national speed limit.
Whatever number you come up with will be useless for bus drivers *and* for
race car drivers.  Similarly, is a statistic which includes drug dealers
being killed by rivals going to be of any use to non-drug-dealing "regular"
people who are trying to lower their own homicide risk?

7. The risk factors which were the raw data for the study's analysis
were measured in a binary way ("yes or no"), as opposed to being ranked
by severity.  This introduces several problems.  Let's take the prior
arrests, for example.  Whatever criteria you use when considering what
kind of arrests should be counted as a marker for violence (one assumes
an arrest for unpaid parking tickets shouldn't count, but even if it
does the following still applies), if your measure is "yes" someone in
the household was arrested versus "no" no one in the household was arrested
(as Kellermann did), you've woefully failed to properly capture the very
tendencies towards violence you're trying to measure.  For example, 23.4%
of the control households had experienced the arrest of a member of the
household.  52.7% of the case households (that is, the ones where murders
occurred) had had such an event.  By Kellermann's measure one has to assume,
for lack of any other option, that we're comparing apples and apples, and
that the nature of the arrests in both types of households are comparable.
When you run your multivariate analysis with this assumption (as Kellermann
did), you end up with a given risk weighting (Kellermann ended up with a
2.5 adjusted odds ratio).  However, the situation is obviously not as
simple as Kellermann made it out to be, and his model is flawed.  What if
the arrests which occurred in the control (i.e. "normal") households
were mostly petty crimes (trivial fistfights, childhood shoplifting,
etc.) whereas the arrests in the case (i.e. murderous) households
were for serious crimes (severe wife beating, felonious assaults,
drug dealing, attempted or previously successful homicide)?  The
case households are, after all, households in which a murder occurred,
and it's pretty obvious that the people living there aren't going to
be saints.  Various studies show that domestic murderers, for example,
*very* often have a history of serious violence in their recent past.
 
So here we find that by using a "yes or no" arrest measure, Kellermann
sadly undermeasured the amount of violence found in the case households.
He assumed that since there were twice as many arrests in case households,
that they were only twice as violent, or twice as likely to have a violent
member in the household, or however you want to describe the relationship.
However, it might actually be the case that only, say, 5% of the control
households had a significantly violent (or criminal, or whatever) person in
the household, whereas say 45% of the households in which a murder took place
might have had such a person.  So Kellermann's using a 2:1 factor, when
the reality might be 9:1.
 
As such, he not only ended up with results which seriously underestimated
the contribution of violent histories to a risk of home homicide, but
since his multivariate model attempts to determine how much the various
possible factors influence each other, he has also understimated how much
the relationship "violent/criminal tendencies leads to both gun ownership
and murder" contributes to his final results.  Put another way, when he
tried to factor out violent tendencies and figure out how much of the
remaining risk factor is attributable to gun ownership by itself, his
inadequate measure of violent tendencies means that he left a lot of the
risk factor directly attributable to violent people in the final measure
of how much guns (supposedly) alone contributed to homicide risk.
 
The "yes or no" choices on the questions of "is there an illicit drug user
in the household" or "has a member of the household been hit or hurt in a
fight" in the home are equally flawed, for "yes or no" answers do not
properly measure the relative severity of the incidents that are observed
in the case versus control households.

8.  The level of honesty of the answers to the questionnaires would
be different in the two types of homes.  Survivors would be more
likely to volunteer incriminating information about the deceased
(both because it's about someone else and because it can no longer
harm the deceased) than the control group would, since the controls
were responding to questions about *themselves*.  In particular,
the survivors would have no reason or even credible ability to
lie about the presence of a firearm in the home when the firearm
was used in a homicide and is a matter of public record, whereas
some in the control group may well not feel comfortable admitting
to a firearm in the house to an interviewer.  Even if firearms were
completely immaterial to homicides and were present in equal
percentages of homicide and control homes, it would only take
30 of the control respondents (out of 388) saying that they didn't
have a firearm in the home when in fact they did to produce the
results found in the NEJM study.

9.  The study found gun ownership *only* correlated to "increased risk"
of being *shot* to death, not being strangled or whatever.  Since
this was a study of homicides that took place *in* the home, we
find that the study discovered that people who were shot in their own
homes had a gun around the house more often than did a home picked at
random, and the study took this as alleged evidence that firearms in
the home "increased" the risk of homicide in general.  More likely, they
just uncovered the totally unremarkable finding that if you get shot in
your own home, there's likely to be a gun in the home.  And drowning
victims are usually found near water.  Big deal.

10.  Finally, everyone who quotes the 3:1 figure as gospel is missing
the point; even ignoring the flaws mentioned above, it simply isn't
true that a 3:1 figure for a population as a whole is applicable to
individual cases, or even to a majority of individual cases.  Let's
say that indeed statistically speaking a gun in the home increases
the overall homicide risk for the home to three times the previous level.
This does *not* mean that if you go out and buy a gun then *you* are more
at risk for homicide.  This is because no population is homogeneous --
because of circumstances, some people will be more at risk for some factor,
some will be less, and the overall numbers will average out somewhere in
the middle.  As the car window stickers say, "your mileage may vary".  For
example, in a given year "a person" has about a 1 in 50000 chance of
death by drowning.  However, this is the mythical "average person".
Whether or not *you* have those odds, or worse odds, or better odds,
depends on your circumstances.  If you are a good swimmer, and don't
spend much time in or near the water, your odds are much better.
If you're a so-so swimmer but spend a lot of time on small boats out
at sea, your odds are much worse.  As a more extreme example, 1 of N
people may die a year from cervical cancer, but that doesn't mean
*I'm* that much at risk -- I'm a man.  Likewise, there's a small but
very high-risk group for "killed by a gun in the home." -- violent people,
criminals, drug dealers, gang members.  While the NEJM study didn't
break it down, or even bother checking, it's almost certain that most
of the home homicides in the NEJM study were committed by people who
were criminals themselves, or by people who were a "murder waiting
to happen" because of their violent natures or unbalanced mental
state; other studies have shown that this is indeed the case.
By and large, "regular people" don't kill each other.  Contrary to what
TV would have you believe, gun homicides are so rare (1:15000 per
person-year) that it only takes a few nutcases to run up the *overall*
homicide risk odds -- nonetheless, the truth is that for the vast
majority of regular people, a gun in the house is more help than harm.
If you're violent, unbalanced, or involved in a life of crime, you are
indeed
much more likely to use your home gun unwisely (and by far higher odds
than
just 3:1).  If you're stable, normal, non-suicidal, and not prone to
extreme violence (and this includes the vast majority of Americans),
your odds will be far *less* than 3:1, and some back-of-the-envelope
calculations show that it's likely so close to 1:1 (no increased risk)
as to be down in the noise, even leaving aside the issue of possible
protection.  There is a very small population of very high-risk
individuals among the very large population of low-risk individuals,
and the high-risk group is skewing the apparent averages.  Don't make
the mistake of thinking that averaged overall figures apply to
everyone equally -- they don't.

(Thanks to Dan Day)