CDC Drug Overdose Data: Patterns and Data Quality Issues

I went digging around in the CDC’s mortality data at the individual record level and it pretty much confirmed my suspicions about the recent surge in overdose deaths. I had done several data-pulls from CDC’s Wonder database, but that only allows you to pull grouped data with various different cuts and summaries. The CDC actually has the individual details on every single death that occurred for every year since 1999 (data files here and data layout here). Unfortunately, the data files are big (~1.5 GB or so) and have an alien file extension that nothing on my computer would open. Fortunately, I have a generous friend who is more tech-savvy than myself. I obtained the full list of the ~47,000 “drug poisoning” deaths, including every listed cause of death (up to 20 are possible, and most list a few), age, gender, marital status, autopsy indicator (was one done, Y or N?).

I was correct to suspect that most of these deaths are drug interactions, not single drug overdoses. I slightly suspect that part of the increase in recent years (tripling from 17,400 in 2000 to just over 47,000 in 2014) is not real, but is an artifact of more thorough toxicology screenings. I can’t prove that based on these data sets (I’d have to know what they *would have* looked like in a counterfactual world where this reporting bias didn’t exist), but it’s hard to explain some of the data trends without assuming there’s some kind of trend in reporting bias. Anyway, I won’t say much about the reporting bias in this post, because I’m only looking at one year’s records. I was probably right to suspect that many of these deaths were wrongly attributed to a drug that happened to be present; I think I have some good evidence that the medical examiner didn’t really know what they were doing (based on the lack of an autopsy, or the assignment to a clumsy “catch-all” cause of death code T50.9, “Other and unspecified drugs, medicaments and biological substances”).

The ~47, 000 figure often given for the total number of drug poisonings is a collection of very different causes of death. It includes ICD-10 codes X40-X44, X60-X64, X85, and Y10-Y14. Codes X40-X44 are unintentional drug poisonings; assuming the individual death certificates were coded correctly, they unambiguously belong in the total. Codes X60-X64 are intentional poisonings, suicides. These are very different classes of social problems. It’s not clear that adding together intentional suicides and accidental overdoses gives you a meaningful sum, just because the chemical mechanism behind the deaths was the same. Codes Y10-Y14 are “undetermined intent”, so many of them are likely suicides and many are likely accidental, but it’s impossible to determine the mix. It seems deceptive to include these in the sum without some sort of caveat. Code X85 is murder; there are only a few of these, but it clearly doesn’t belong in the sum. Alcohol poisonings are coded separately; the 47,000 figure doesn’t include alcohol poisonings unless it’s a combination of alcohol and other substances (which is actually fairly common). Of the 47,196 drug poisonings, 5,447 were suicides, 2,862 were “undetermined intent”, and 82 were murder. 38,841 were unintentional drug poisonings, codes X40-X44. So it’s not as though the ~47,000 figure is off by an order of magnitude. If your intent is to quantify the drug problem by counting overdose deaths, it’s off by about 20% or so. The vast majority are still “unintentional drug poisonings.” But it reeks of bad faith when someone overstates the magnitude of a problem in this way. To be fair to the CDC, some of its reports clearly specify the breakout (like this one). But it’s not always so forthcoming. I have often seen articles or blog-posts use the 47,000 figure as a headline or attention-grabbing block-quote.

I can’t quantify the following consideration, but it’s worth noting that some suicides could be miscoded as “unintentional poisonings.” The family might have religious, insurance, or other reasons not to mark the death as a suicide. Unfortunately, nothing in the data file allows me to quantify this kind of miscoding.

Multi-Drug Interactions

The codes mentioned in the above paragraph are broad cause of death categories; there are also “multiple cause-of-death” codes (mostly in the T40 through T50 range) that list specific causes of death. A few of these (T40.1 for Heroin, T40.5 for cocaine, T40.3 for methadone) are codes for a specific drug; others (T40.2 for “Other Opioids”, T42.4 for “Benzodiazepines”, T40.4 for “Other synthetic narcotics”) are for broad categories of drugs. It might be useful to know specifically if a death coded T40.4 was, say, a fentanyl poisoning from spiked street heroin or a synthetic prescription opioid painkiller that was carelessly mixed with alcohol; the cause of death codes unfortunately are not granular enough to reveal the true cause. A drug poisoning death is coded by one of the codes in the above paragraph (X40-44, 60-64, 85, and Y10-14); such a death will also typically include one or more of the “multiple cause of death” codes (the T40-T50 codes, including others related to organ damage, alcohol, etc.). This way, you can look at a death record and answer the two questions “Was it a drug poisoning?” and (with the multiple cause of death codes) “What drug or class of drug or combination of drugs was involved?” I found that most drug overdose deaths involve multiple substances. See the following table, and notice how rarely a single substance is implicated:

This doesn’t even include alcohol. I was a little unclear about which cause of death codes were alcohol, and apparently so are the medical examiners and nosologists who assign the cause of death codes. Sometimes a death involving alcohol is listed as T51.0 (“Ethanol”), sometimes X45 (“Accidental poisoning by and exposure to alcohol”), sometimes T51.9 (“Alcohol, unspecified”). Sometimes it’s a combination of these codes, as if the coder didn’t know which one to write down. I chalked this inconsistency up as another data quality issue. Anyway, when we include alcohol, even fewer of these deaths are single-substance overdoses:

Take heroin as an example, an archetype for a drug that’s easy to overdose on (and I’m not at all claiming that it isn’t). Only 29.5% of heroin deaths involved *just* heroin and no other drug; 70.5% of heroin poisoning deaths have other substances on the death certificate. Only 1.4% of benzodiazepine deaths have no other substances on the death certificate. This is not an inherently dangerous category of drugs, but people who take them need to be starkly warned not to take them with alcohol, opioids, and certain antidepressants. I thought the single-drug overdoses from methadone would be higher; the stuff stays in the system so long it can build up to dangerous levels for a naïve user. Still, it appears that the culprit is often a deadly combination, not methadone by itself.

I hesitate slightly in saying all the above. It’s possible that the medical examiners are simply writing down every goddamn thing they find on the toxicology screening, in the digestive track, or mentioned in interviews with family. I’m sure if you scoured the records you’d find death certificates with combinations of substances that aren’t really dangerous in combination, which would cast doubt on the notion that every drug mentioned on the death certificate contributed to the death. But this cuts the other way, too. If medical examiners are promiscuously listing every substance in a person’s body whether it contributed to the death or not, then no doubt some of these deaths are inappropriately categorized as drug poisonings. Unfortunately, this is a suspicion that I can’t quantify with the given data. If the death is miscoded, nothing in the CDC data file allows me to correct that error.

Why do I make such a big deal out of the multi-drug interactions? Because the solution to this problem is different from the solution to the problem that the public *believes* exists. The public thinks that people become enslaved by drugs, and then compulsively pop pill after pill (or snort after snort, or injection after injection) until they drop dead. This doesn’t really describe how most drug poisonings happen, or even how most single-drug overdoses happen. If a typical drug poisoning involves a bad combination of substances, then the solution is to make the users avoid *some* of those substances, not necessarily *all* of them. The solution is education and good PSAs: “Don’t take X with Y. It will kill you!” Don’t tell people “Don’t take X at all. It will kill you!” The moment they survive an encounter with X, they’ll stop believing everything else you say. People should be warned about the dangers of drug use, but the warnings should be accurate and useful. The warnings should be crafted with the level-headed realization that some people will imbibe anyway. We might as well warn them so they do it safely. You can’t look at the above tables and say, “The solution is to prohibit drug use, beat up and imprison dealers, and force users into counseling.” Prohibition is an extremely ham-fisted (and ineffective) solution to the problem that actually exists.  Most of these deaths are naïve users who don’t understand the interactions between the substances they are using. Society should accommodate this kind of recreational use with proper guidance on potentially dangerous dosage and combinations.

If you want to know what kinds of drugs tend to found in combination, I’ve built a cross-table with the  14 most common drug poisoning codes, plus alcohol. In the top table, you can see that the first row and the first column are identical, as they should be. 4,791 deaths, for example, involved “Other Opioids” and “Other and unspecified drugs, medicaments and biological substances”, as you can see in row 1 column 2, and row2 column 1. In the bottom table, I am dividing by the largest number in the row. So read this as saying “25.5% of ‘Other and unspecified drugs, medicaments and biological substances’ deaths involve ‘other opioids’” and “47.7% of ‘other opioid’ deaths also involved ‘Other and unspecified drugs, medicaments and biological substances’. Clearly there’s 100% across the diagonal; 100% of deaths involving “cocaine” also involve “cocaine”, get it? 34.2% of “other opioid” deaths also involved Benzodiazepines, which confirms my suspicion that this is a big deal. 52.3% of benzodiazepine deaths also involve “other opioids”. (The 34.3% and 52.3% figures have the same numerator, 3431, but different denominators.)

Acute vs Chronic Causes of Death

Often these overdose death records include a cause of death that is associated with chronic alcohol or drug use, rather than acute poisoning (K700, “Alcoholic fatty liver”, accompanies several of the above mentioned alcohol poisoning codes). This suggests there is some confusion and uncertainty about the actual cause of death. The causes of death related to chronic drug use are F11-F16 and F18-F19; they are separate from the codes for acute poisoning and are not included in the “47,000” figure so often cited. A death that includes both chronic and acute cause of death codes is cast into doubt. (See page 74 of this document; the city of New York had the “chronic” and “acute” drug deaths confused for many years in one of its official documents.) I don’t have the chronic drug use cause of death totals handy, but they are quite a bit smaller than the number of acute poisonings (they are roughly in the ~2,000 range rather than the ~40,000 range). It seems like a meaningful measure of the drug problem would add chronic and acute accidental drug poisonings, not accidental and intentional drug poisonings (including murder). That number would be right around 41,000 in 2014.

Absent Autopsies and Catch-all Categories

There are some patterns in the data that are cause for concern, and I’ll spell out two of them in this section. The CDC data file has an autopsy indicator, with “Y” for “Yes, an autopsy was done”, “N” for “No” and “U” for “Unknown.” Autopsies are much more common for drug overdoses than for other categories of deaths (75 – 80% of drug overdoses involve an autopsy; I don’t have the same figure for total deaths but it’s definitely smaller). Of course that still calls into question 20-25% of poisoning deaths. There is just no way to identify a drug poisoning death without an autopsy. In his textbook “Pathology of Drug Abuse,” Dr. Steven Karch discusses all the various ways that, even *with* an autopsy, an unwary medical examiner might be misled about the cause of death. He repeatedly cautions that a toxicology screening alone can be extremely misleading. Methadone patients, for example, often walk around with levels of the drug in their bodies that would be toxic to most people. If such a person drops dead of a sudden heart arrhythmia, it’s unlikely that the true cause of death will be discovered (autopsy or not). Of the 38,841 accidental drug poisonings in 2014, 8,077 (20.8%) had no autopsy and 8,847 (22.8%) were marked with either an “N” or a “U”. This seems like a huge caveat that should be added to every CDC report that simply adds up the totals by cause of death code. Such a sum overlooks these underlying problems with data quality.

Another data quality issue I noticed was the liberal use of the cause of death code T50.9, “Other and unspecified drugs, medicaments and biological substances.” This is a very generic-sounding category that suggests the medical examiner didn’t really know what was going on. A total of 18,782 of the accidental overdose deaths were marked with this cause of death code; 7,389 were marked with this cause of death code *alone* and no specific category of drug (see above tables; the number falls to 7,024 if I include alcohol as a substance). I suspect that if the medical examiner really knew what killed the person, they’d mark down one of the other drug categories. Some of those other categories are pretty broad, so it shouldn’t be hard to find at least *one* that’s appropriate. The number of asterisks marking the drug poisoning deaths figure is growing here; 19% (=7,389 / 38,841) of these deaths can’t even be attributed to any specific class of drugs.

I don’t want to double-count in my last two paragraphs; there are some deaths for which there wasn’t an autopsy *and* it was marked with code T50.9 alone (with no other substances). When I take out the double-counting, I get 36.1% of drug poisoning deaths fall into one or the other category. So just over a third of the 38,841 accidental drug poisonings are in serious question. (Details: 5160 deaths had an autopsy and had T50.9 as the only drug-related cause of death, 8847 deaths had “N” or “U” under autopsy, (5160 + 8847)/38841 = 36.1%. Please, please, please, download the data file and check my work.)

There is another issue that I’ll mention without quantifying. Many of these deaths include cause of death codes for various types of organ failure. I don’t know if it matters what order the causes of death are listed in (remember there are up to 20, and most deaths include several), but sometimes the organ failure code is listed prior to the drug poisoning code on the death record. Here are some of the organ failure causes of death listed on some records that I’m essentially pulling at random: “Hypertensive heart disease without (congestive) heart failure”, “Atherosclerotic cardiovascular disease, so described”, “Cardiomegaly”, “Influenza due to identified avian influenza virus”, “Sleep apnoea”, “Chronic obstructive pulmonary disease, unspecified”. Sometimes multiple organ-related causes like those listed above were pulled from the same record, and sometimes the only category of drug poisoning on the record was the catch-all “Other and unspecified drugs, medicaments and biological substances” mentioned above. I’m really not sure what’s happening here. Did the medical examiner infer with certainty that a drug caused the death, then pedantically list the effect the drug had on the decedent’s organs? Or did the medical examiner write down everything that was wrong with the body and someone at the CDC (speculatively) coded it as a drug poisoning?

I’m sure that some of the increase in drug poisonings since 1999 is real. I don’t want to imply that the entire trend is spurious. The sheer tonnage of opioids prescribed has increased dramatically, and these drugs are bumping into other drugs that people are already taking (alcohol, benzodiazepines, antidepressants, etc.) But this is not the moral panic it’s being sold as. As I’ve said in previous posts, the illicit use of painkillers, sedatives, tranquilizers, and other prescription drugs is flat, according to the SAMSHA and Monitoring the Future surveys. It’s simply not the case that we’ve made recreational drugs more available, and drug-crazed addicts are popping them until they drop dead. This is not an addiction crisis. Until we properly understand the mechanism behind the increase in drug poisonings, we can't really claim an informed opinion about the solution.