In CANHEART very high HDL cholesterol was actually associated with higher non-cardiovascular mortality.
Especially levels over 90 mg/dl (2.33 mmol/l), but also over 70 mg.dl in men (1.81 mmol/l).
These are very high HDL levels, I don't remember seeing levels this high in non-drinkers on LCHF diets, no matter how much coconut oil they eat.
The most obvious question is, what about alcohol? Alcohol elevates HDL but at high intakes promotes secretion of useless and atherogenic HDL subtypes. Ko at al claimed to have adjusted for excess alcohol intake, which was highest in those with highest HDL;
"Heavy alcohol consumption, as defined by the use of 5 or more drinks on 12 or more occasions per year was also included in the model for non-cardiovascular non-cancer death."
Newsflash - drinking 6 drinks 13 times per year will not raise your HDL. You really need to be a chronic alcoholic. In 2012, approximately 5 million Canadians (or 18 % of the population) aged 15 years and older met the criteria for alcohol abuse or dependence at some point in their lifetime, but how many at any one time qualify as chronically alcoholic is unknown.
Even so, this adjustment was far from perfect.
"Since the use of smoking and alcohol was not available in entire CANHEART cohort, we imputed smoking status and heavy alcohol use for those with missing data based on the characteristics of the respondents to the Canadian Community Health Survey. Multiple imputation using complete observations and 10 imputation datasets was conducted. Smoking status was available for 5,093 individuals and alcohol use was available for 5,077 individuals who completed the survey."
This was a tiny fraction of the 631,762 individuals in the study - less than 1% - and presumably was either restricted to a single geographical area, or a few especially obliging subjects.
Alcohol intake is known to be misreported in dietary surveys by a factor of 2-3. Alcoholism is probably under-reported to health professionals to a much greater extent, especially in countries where health insurance is a major factor in access to care.
Another confounder is the effect of genetic hyperalphalipoproteinemia. One genetic cause of very high HDL is a CETP defect.
"...the in vitro evidence showed large CE-rich HDL particles in CETP deficiency are defective in cholesterol efflux. Similarly, scavenger receptor BI (SR-BI) knockout mice show a marked increase in HDL-cholesterol but accelerated atherosclerosis in atherosclerosis-susceptible mice. Recent epidemiological studies in Japanese-Americans and in Omagari area where HALP subjects with the intron 14 splicing defect of CETP gene are markedly frequent, have demonstrated an increased incidence of coronary atherosclerosis in CETP-deficient patients. Thus, CETP deficiency is a state of impaired reverse cholesterol transport which may possibly lead to the development of atherosclerosis."
Ko et al do not mention the likelihood of such conditions affecting their analysis. Even if we assume that both chronic alcoholism and hyperalphalipoproteinemia are rare conditions, men with HDL over 90mg/l were less than 0.3% of the study population, and of these few men, only a few dozen died during the study. The exact number isn't clear because the only mortality data given is for adjusted age-standardized rates per 1,000, but from total deaths and these rates I estimate it to be (at the very most) 70-80 deaths, of which 30-35 were non-cardiovascular and non-cancer deaths, out of about 2240 men. The majority of alcohol-related such deaths in Canada are due to alcoholic liver disease, motor vehicle accidents and alcohol-related suicides. Had Ko et al given a breakdown of non-cardiovascular causes of death for the highest HDL categories, it would have been relatively easy to tell how many of these were due to alcoholism.
Overall, people in the high HDL categories exercised more, had lower triglycerides, less diabetes, lower LDL, more ideal BMI, and ate more fruit and vege than people in the middle and lower ranges.
Did these things cause them to die at a higher rate?
Here's an alternative explanation - the baseline characteristics represent only the vast majority of people in each category. The vast majority of people in each HDL category, even the highest, didn't die. The people who died in the high HDL categories tended to be the people with alcoholism and poorly-managed genetic hyperalphalipoproteinemia, and their baseline characteristics, had they been isolated, would have been quite different. These are the people for whom high HDL is not protective, and, as their numbers increased in categories of increasing HDL, the usual dose-response relationship between HDL and cardiovascular disease and cancer, seen in better-controlled populations, was lost.
A criticism is that Ko et al have misrepresented the lipid lowering trial data to support their thesis.
They say "Several contemporary studies have shown a lack of significant association of HDL-C levels and outcomes for patients on higher-intensity statins, with coronary artery disease, or who had undergone coronary artery bypass graft surgery (12,13,15)."
However, reference 12 states
"In 8901 (50%) patients given placebo (who had a median on-treatment LDL-cholesterol concentration of 2.80 mmol/L [IQR 2.43-3.24]), HDL-cholesterol concentrations were inversely related to vascular risk both at baseline (top quartile vs bottom quartile hazard ratio [HR] 0.54, 95% CI 0.35-0.83, p=0.0039) and on-treatment (0.55, 0.35-0.87, p=0.0047). By contrast, among the 8900 (50%) patients given rosuvastatin 20 mg (who had a median on-treatment LDL-cholesterol concentration of 1.42 mmol/L [IQR 1.14-1.86]), no significant relationships were noted between quartiles of HDL-cholesterol concentration and vascular risk either at baseline (1.12, 0.62-2.03, p=0.82) or on-treatment (1.03, 0.57-1.87, p=0.97). Our analyses for apolipoprotein A1 showed an equivalent strong relation to frequency of primary outcomes in the placebo group but little association in the rosuvastatin group."
In other words, people in the top quartile for HDL and ApoA1 on placebo had the lowest vascular risk, and these people got no extra benefit from LDL lowering with a statin. And because we are looking at quartiles, not isolating a small number of people who have freakishly high HDL for some reason, there is a true dose-response effect of HDL between quartiles in the placebo arm.
This effect has been seen in multiple trials. Drug trials are likely to exclude alcoholics and binge drinkers.
All these 3 references tell us is that the predictive value of HDL is excellent, but is lost when people are undergoing intensive treatment for coronary artery disease, a classic case of Goodhart's law, "When a measure becomes a target, it ceases to be a good measure." We see this again and again with intensive drug treatment of metabolic markers.
Thankfully, it doesn't seem to apply to diet and lifestyle interventions.
 Ko DT, Alter DA, Guo H, et al. High-Density Lipoprotein Cholesterol and Cause-Specific Mortality in Individuals Without Previous Cardiovascular Conditions: The CANHEART Study. J Am Coll Cardiol. 2016;68(19):2073-2083. doi:10.1016/j.jacc.2016.08.038.
 Yamashita S, Maruyama T, Hirano K, Sakai N, Nakajima N, Matsuzawa Y.
Molecular mechanisms, lipoprotein abnormalities and atherogenicity of hyperalphalipoproteinemia.
Atherosclerosis. 2000 Oct;152(2):271-85.
 Ridker P.M., Genest J., Boekholdt S.M., et al; for the JUPITER Trial Study Group. HDL cholesterol and residual risk of first cardiovascular events after treatment with potent statin therapy: an analysis from the JUPITER trial. Lancet. 2010;376:333-339.