Case-Control Studies

Yes. So. We're back in session. Good. Last time we met, we talked about design of survey studies and about randomness. No, not randomness in that we had a random conversation. We talked about randomness in the context that a branch falling from a tree and hitting you is not, as Dr. Jay would think, just as random as an outbreak of measles or some such. We also talked about outbreaks a little bit. The main gist there was to define an outbreak in terms of person, place, and time.

It's okay if you head over to the other lessons to refresh your memory.

Alright, then... Tonight we're going to talk about case-control studies that are done in response to an outbreak and in order to get an idea of what is causing - or what caused - the outbreak. We'll also touch on why sometimes some outbreaks are just not conclusive in their findings, even with some pretty good theories on what happened. We'll do all this, of course, after the jump...

I DON'T FEEL WELL
Let me tell you a quick story. As always, it is a mash-up of several true stories, and it is intended for educational purposes. Once upon a time, in a town in Texas, several people went to a Greek-style food restaurant to have a birthday party. There were 32 people in the party, and the restaurant was closed off to other customers for the party since the birthday boy was the owner of the restaurant. Five people cooked and served the food, and one person served at the cash bar.

The buffet-style meal had a pasta dish, a meat dish, a potato dish, bread, vegetables, and fruit. I'm over-simplifying, of course. You'll see why in a minute. Everyone had a grand time at the party. Later that night, one of the attendees began to feel sick to his stomach. He became violently ill. His wife took him to the hospital. On the way there, the man began having trouble breathing. He became lethargic. Once in the emergency department, he had to be put on a ventilator. Over the course of the night, twelve other people would develop similar symptoms.

DEPLOY THE EPIDEMIOLOGISTS
The thirteen (13) cases went to different hospitals in the town, but, luckily, the hospitals were networked into the state's health department's system for syndromic surveillance. The system scours the "chief complaints" and "discharge diagnosis" sections of all the medical records in the state. That night and into the next day, the system detected 13 cases of "food poisoning" accompanied by "trouble breathing". One of those cases was later re-coded as "possible botulism", raising the alarms of the epidemiologists at the health department.

Five epidemiologists got in a state-owned car and drove to the town to assist the local health investigators in, well, investigating the 13 cases hospitalized at five facilities. They took with them laptops, notebooks, cellphones, and lots and lots of coffee.


STEP ONE: CONFIRM THE DIAGNOSIS
Outbreak investigations begin by confirming the diagnosis, no matter how vague or specific said diagnosis is. In the first two days of this outbreak, the diagnosis was "food poisoning". By the third day, lab results were in for all 13 cases. They all tested positive for botulism toxin. That many cases of botulism at one time and in one town in Texas constituted an outbreak.

Upon being interviewed, the cases all reported one exposure in common: the party. Not all the cases knew each other, and it was a little tricky to find out if they knew each other because of confidentiality considerations, but they all reported that they had been at the Greek-style restaurant three days previous. As these initial 13 cases were being interviewed, five more were detected. Later that night, the total rose to 24.

TIME WAS OF THE ESSENCE
Seeing that all 13 initial cases had the restaurant in common, the health officer for the district where the restaurant was located decided to issue an emergency order to shut down operations at the restaurant. Sanitarians headed out there to take samples of the food, but it was too late. The restaurant had gotten rid of all the food from days previous after management was tipped off that several customers had become sick. (Remember that some of the party-goers were directly associated with the hospital.)

With the restaurant closed and inspected, it was expected for cases to stabilize. No new cases were reported for the rest of the week.

THE OTHER STEPS
This example is more like what really happens in real life. Rarely do we epidemiologists and other public health workers get to carry out, in sequence, the steps to an outbreak investigation. The steps are as follows:

• Prepare for field work
• Establish the existence of an outbreak
• Verify the diagnosis
• Define and identify cases
• Describe and orient the data in terms of time, place, and person
• Develop hypotheses
• Evaluate hypotheses
• Refine hypotheses and carry out additional studies
• Implement control and prevention measures
• Communicate findings

As you can see, the actions of the health office pretty much skipped over the whole development of the hypothesis of what was going on and jumped straight to implementing control and prevention measures. But that was because virtually all the cases had the restaurant in common.


THEY WERE LUCKY, KIND OF
Several people ended up on a ventilator, and that wasn't so lucky. None of them died, which IS lucky with botulism. The Centers for Disease Control and Prevention had to fly in anti-toxin for all the cases. In the five stories that I'm taking to create this one, the system worked and lives were saved. And a lot of luck went into that, I believe.


SO WHAT WAS IT?
Alright, so let's get to the mystery-solving part of an outbreak investigation. While our brave epidemiologists knew that the restaurant was the most likely source of the outbreak, they had no clue what there was causing the disease. So they put out the word for people who had also been to the restaurant but were not sick to contact the local health department to be part of a study. People who went to the party and were not sick were put at the top of the list of controls. All 24 sick people were the cases.

To get a good statistical measurement of the odds of exposures to different foods, a ratio of two controls per case is recommended. Four controls to a case is optimum. One case to one control is pushing it, especially if the event is rare. (You can't study a handful of people and reach a proper conclusion of what's going on. Are you reading this, Dr. Andrew Wakefield?) Ideally as well, your controls and your cases should be as similar in every demographic and socioeconomic respect, especially if you're looking at an environmental cause for the disease. You need to eliminate as many confounding variables as you can.

THE INTERVIEW
A menu of the items consumed at the restaurant during the party was obtained. A different menu of the food the restaurant served was also obtained. Cases and controls were asked if they ate or did not eat the different foods. Here are the results.

Let me describe to you what you're seeing. Remember how I kept the foods simple? In a real situation, we may have do deal with dozens if not hundreds of food items. This is particularly true if you're dealing with an exposure that lasted for days, like on a cruise ship or at a convention. In this example, we kept it to a bare minimum of food items. "Proportion" represents the proportion of ill folks who consumed the particular food item. As you can see, a lot of the sick folks consumed potatoes. Three quarters of them consumed bread. All of them had tea and water as well. Is this enough to implicate a particular food or drink?

NO!
Remember that we humans tend to look for patterns and can come up with some whacky notions of how the universe works based on our observation. In this example, you might be tempted to say that it was the potatoes that caused the botulism since so many of them were consumed. But what about the tea and the water? Could the toxin had been put intentionally into the water? (That was one of the theories circulating about one of the outbreaks mashed into this example.) If you're going by proportions alone, then how do you explain the one case who did not eat potatoes and still got sick?

The smart ones among you will have recognized that this same line of thinking should have gone into the decision of whether or not to close the restaurant. But the circumstance there was a little different. In that circumstance ALL the sick people had one thing in common: the restaurant. They were from different parts of town, different social networks, and different age groups, but they all had the restaurant in common. That was enough to move the health officer to close down the restaurant. If he was wrong, they would have seen other cases occur. If he was right, no more cases would occur. Dr. Gregory House would have been proud.


SO WHAT DO WE DO?
Let's look at the controls. How did they report any exposures?

Same here as above. The "proportion" is pointing out the proportion of the 48 controls that consumed a particular food item. Unlike with the cases, we are more inclined to look at the lower proportions here. At least that's the intuition. "If you're not sick, then you must not have eaten it," we're inclined to think. Sorry for the lack of row headings, but you can see that the lowest proportion is the potato dish (third row), followed by the pasta (first row), and then the tea (third from the bottom). You're still eyeing that potato dish, aren't you?

NOT THERE YET
Remember that things happen just randomly, we've covered this. So we need to make sure that these observations are not this way by chance alone. So we look at an odds-ratio. The odds-ratio (OR) gives us the ratio of the odds between the two groups (the "ills" and the "wells"). The proportion of people who ate a certain food within a group can give some clues, but not definitive. Hence the need to compare between groups. So let's look at the OR:

As you can see, the potatoes have the highest OR. That is, the odds of being exposed to potatoes is about 7.7 times higher for those who are sick than those who are well (0.96 divided by 0.13). But notice how the OR is higher than 1.0 (even odds of being exposed, or no difference in exposure between wells and ills) for some other foods and the Tea and Water. That's where statistics kick in.

I WON'T BORE YOU
For the sake of expediency, because this lesson is going a little long, I'll tell you that there are statistical analyses that you can do on these numbers to find out if the OR you are looking at are the result of just chance or if there really is something there. Trust me, there have been times when the OR is high and looks impressive, but the results are not statistically significant. It was all due to chance. So, for our example, we can say that the potatoes were the most likely source of the botulism, but, because we don't have any food samples, we'll never know for sure.


ONE MORE TOOL
Another tool that you can use in your investigation is an "epi curve", which is a graph showing the onsets during the outbreak. An epi curve can tell you if you're dealing with a point-source exposure, meaning that only one thing was the source of the outbreak and that it happened in one moment in time (a day or two). A propagated source curve will tell you that you're dealing with something that is out there still causing infections, such as contaminated water. Epi curves can even show you secondary cases or an outbreak that explodes all over again when a case from the original outbreak goes on to infect others. (I've seen this in outbreaks of hepatitis A, where someone from the original outbreak doesn't get treated and goes on to infect others after the 14-to-21-day incubation period.)

Here is the epi curve of our outbreak:

As you can see, we had an explosion of cases the day after the part, then a few more cases after that, and then none after the restaurant is closed. This was a point-source exposure. Here is what a propagated source looks like:

As you can see, the number of new onsets lingers and lingers, and even rises, before it trails off once control measures are in place. In some cases, like in some very bad cholera outbreaks, new cases linger for weeks or months. Just look at what is going on in Haiti.

Finally, this is is a graph showing secondary cases at day 10:

Again, these are very simplistic and meant only to prove a point. Many times, you get epi curves that look like a mishmash of all these graphs.


ONE MORE THING
One quick thing... The CDC does not intervene in each and every outbreak. Their intervention requires an official request for assistance from the State (read the 10th Amendment). And there are times when outbreaks are not detected until long after they're done. Yeah, my job ain't easy.


AND ANOTHER THING
If you want to read about the worst outbreak of botulism (foodborne) in the United States to date, check it out here and here (PDF). Again, I based the above example on some of that outbreak and others.

Next time, I hope to finish off this first "semester" of Epidemiology Night School with a post about cohort studies and why they are the most powerful studies we have... Though they also have a lot of limitations.