# Dangers of Forcing Regressions Through the Origin

Posted on the 17 October 2017 by Bradshaw @conservbytes

I had an interesting ‘discussion’ on Twitter yesterday that convinced me the topic would make a useful post. The specific example has nothing whatsoever to do with conservation, but it serves as a valuable statistical lesson for all concerned about demonstrating adequate evidence before jumping to conclusions.

The data in question were used in a correlation between national gun ownership (guns per capita) and gun-related deaths and injuries (total deaths and injuries from guns per 100,000 people) (the third figure in the article). As you might intuitively expect, the author concluded that there was a positive correlation between gun-related deaths and injuries, and gun ownership:

__

Now, if you’re an empirical skeptic like me, there was something fishy about that fitted trend line. So, I replotted the data (available here) using Plot Digitizer (if you haven’t yet discovered this wonderful tool for lifting data out of figures, you would be wise to get it now), and ran a little analysis of my own in R:

Just doing a little 2-parameter linear model (y ~ α + βx) in R on these log-log data (which means, it’s assumed to be a power relationship), shows that there’s no relationship at all — the intercept is 1.3565 (± 0.3814) in log space (i.e., 101.3565 = 22.72), and there’s no evidence for a non-zero slope (in fact, the estimated slope is negative at -0.1411, but it has no support). See R code here.

Now, the author pointed out what appears to be a rather intuitive requirement for this analysis — you should not have a positive number of gun-related deaths/injuries if there are no guns in the population; in other words, the relationship should be forced to go through the origin (xy = 0, 0). You can easily do this in R by using the lm function and setting the relationship to y ~ 0 + x; see code here).

Sounds reasonable, and if you take it on face value, it justifies the purported positive relationship shown in the original article. Suddenly, the slope is now positive (1.0308 ± 0.1236) and supported statistically.

But slow-down there, matey! There’s a problem here, and it means everything for the interpretation of the relationship. Just because it’s intuitive and you can force the relationship to go through the origin, it doesn’t mean that the data support this approach.

If you look closely, you’ll see that the origin is far away from the nearest real data — in other words, we’re taking a huge leap by assuming the relationship continues from the lowest values in the dataset back to the origin. You’ll also notice that the outlier (the USA: point farthest to the right of the plot) seems to be driving much of this relationship.

What we need to do is examine the relative support for the 2-parameter (intercept + slope) model versus the origin model. We can do this easily by comparing the sample size-corrected Akaike’s information criterion (AICc) for each model, and then dividing the AICc for the full model by the AICc of the origin model; this is called the information-theoretic evidence ratio.

As you can see from the code, the evidence ratio is 56.8, which means that the 2-parameter model is 56 times more likely than the origin model after accounting for different the number of parameters in each.

As such, there is no statistical justification whatsoever to support the forced-origin model, even though it seems to make intuitive sense (see another useful discussion here). This is because we have overstepped the bounds of our inference capacity based on the small sample of data available.

Now, in this particular case, I admit that even I wanted this relationship to be real, for I am a firm supporter of extensive gun control (the many other data supporting why this makes sense are overwhelming). I am, in fact, a gun owner myself (for scientific and farming purposes), and I am in absolute favour of the tightest gun-control legislation possible.

But just because I want it to be true, doesn’t make it so. We have to stick to the evidence.

You Might Also Like :

• ## You Belong: Shawn Austin Interview

We caught up with Shawn Austin at the SiriusXM Canada offices in Toronto on a busy media day for the Vancouver singer and sat down to talk about his career and... Read more

The 16 October 2018 by   Phjoshua
LIFESTYLE, SELF EXPRESSION
• ## Mattel Introduces Doctor Who Barbie

Most of what I’ve seen from Mattel recently has been underwhelming. However, the announcement of their newest doll, the 13th Doctor based on the long running... Read more

The 16 October 2018 by   Ashley Brooke, Kewpie83
CULTURE, FAMILY, SELF EXPRESSION
• ## UNIBEN Admission List Out – 2018/2019

UNIBEN Admission List for the 2018/2019 academic session is now out on the school portal. All candidates who applied for admission in the university are hereby... Read more

The 16 October 2018 by   Naijapary
SOCIETY, A-CLASSER
• ## Doll Shows 101: Etiquette

For a doll collector, attending a doll show can be as exciting as stepping through the gates of Walt Disney World. Even the most narrowed minded collector is... Read more

The 16 October 2018 by   Ashley Brooke, Kewpie83
CULTURE, FAMILY, SELF EXPRESSION
• ## Neanderthal Cave Art is Fake News?

Life in a Neanderthal cave was surprisingly human. They built structures, created jewellery, and cared for each other. To top it all of, we recently learnt... Read more

The 16 October 2018 by   Reprieve
BIOLOGY, SCIENCE
• ## Is the Tide Turning in Australian Dietary Advice?

Has the Dietitians Association of Australia (DAA) had a change of heart? Just last month, we reported that the DAA was heavily influenced by sponsorships ... Read more

The 16 October 2018 by   Dietdoctor
DIET & WEIGHT, HEALTH, HEALTHY LIVING, MEDICINE
• ## Yahoo Boy On The Run After Chopping off Old Man’s Head and Manhood in Benin...

A Yahoo Boy identified as Bino has disappeared into the thin air after beheading a man for ritual purpose in Ayogwiri community Benin . Read more

The 16 October 2018 by   Naijapary
SOCIETY, A-CLASSER