Using the MDQ in Primary
The predictive value of the test results
The Briefest Version
Let's say you are using the Mood Disorders Questionnaire in your clinic, perhaps to address the FDA recommendation to screen for bipolar disorder before giving an antidepressant. You hand the patient an MDQ and go on to see your next patient. The result: 7/yes/yes (as detailed on the download-an-MDQ page). What do you tell the patient?
The red ink in the table below offers a quick summary. As you can see, the result to watch out for is a positive test when you were not expecting it: this does not mean the patient actually "has" bipolar disorder. Read the following sections for a detailed, step-by-step procedure for using the MDQ with greater precision. If you need it, here's a quick review of predictive values and how they derive from sensitivity/specificity.
Likelihood of bipolar disorder after MDQ testing
Your prior estimate of likelihood
Intermediate: 50-60 %
This table gives estimates of the likelihood of bipolar disorder, based on the test result and your very rough estimate of the likelihood of bipolar disorder before you gave the test. Of course, you should be familiar with bipolar II variations before you trust your "pre-test" suspicions. Read about bipolar II, using the version for doctors (the patient/family version has more details).
When in doubt, refer the patient for more education about bipolar II (use that same site); and consider interviewing a spouse, parent or friend to add more data before trying to establish a diagnosis (you can send home an MDQ with the patient and ask that they have a significant other complete it as though they were the patient, based on their own experience of her/his behavior -- although interpretation can be tricky if you don't get a resounding positive or negative). If you can't easily refer uncertain cases to a psychiatrist for an opinion, try a good psychotherapist (one who knows about bipolar variations -- and who will send you a report!).
Read on below to see how to interpret the MDQ in 4 different scenarios; learn about bipolar prevalence estimates; and consider how sensitive this whole system is to your guesses about "prior probability"!
A Step-by-Step Approach to MDQ interpretation
Negative predictive value
Let's look at what you're going to do with your patient's test result. If you have time you can continue with the discussion below about how this all works. (Here's that quick review of sensitivity, specificity, and predictive values again)
Here we go: your next patient Ms. Anderson has completed the MDQ and the result is "negative" (5/yes/yes). What do you tell her? Remembering that the "predictive value" of a test depends on the prevalence of the illness, you use the following graph (where did this graph come from?):
You estimate the appropriate "prevalence" for Ms. Anderson by asking yourself: "what did I think the likelihood of her having bipolar disorder was before I did the test?"
Not so tough: you had some impression of that likelihood based on the symptoms you were hearing about that prompted giving her the test. You could say "oh, she had no symptoms at all that I recognize as bipolar" (for a primer on bipolar II, the subtle one that creates so much diagnostic confusion, read the version for doc's). In that case you use the general prevalence rate in patients with depression symptoms only.
What is the prevalence of bipolar disorder in all patients with depressive symptoms? (Note we're not talking about the prevalence of bipolar disorder in the general population, which would be much lower). That's quite debated and depends on the definition of bipolar disorder you use. You can read a brief discussion of bipolar prevalence below.
For now, you choose to use a conservative prevalence estimate, 0.1. Using the graph below, you see that the "negative predictive value" of the MDQ at that prevalence is about 95%. You tell the patient "the test shows you are 95% likely not to have a condition that raises the risks of bad outcomes if we use an antidepressant. In other words, we checked to see if you might have a bipolar-like depression, and we're 95% sure you don't, based on this test. We still have to keep our eyes out for any kind of worsening of your condition when you take an antidepressant, as you read at the bottom of that MDQ page [presuming you used the one from this website], but this test result lowers the likelihood of that kind of problem".
Got it? Want to try a different prevalence estimate? (or you can jump to a similar step-by-step for positive predictive value)
Okay, Round Two, negative predictive value: Another patient, Mr. Brown, also has a negative MDQ. In his case you actually thought it might come out positive, but again it was 5/yes/yes. Because this time your level of suspicion was different, we need to use a different prevalence estimate now.
As always you start by asking yourself this question: "what did I think the likelihood of his having bipolar disorder was before I did the test?" In this case let's say you would have bet there was at least a 50% chance. Now you use that same graph again (see below), this time using a "prevalence" figure of 0.5. Remember, we're looking at the prevalence rate of bipolar disorder in all people who have the symptoms you elicited from this gentleman. If you are correct in your guess about him, you've estimated that such a population has a bipolar prevalence rate of 50%. Admittedly this is making some guesses, but if you're very experienced, they are informed guesses. If you're pretty green, then you'd just have to admit it's a very rough guess. As we'll consider in the brief sensitivity analysis below, the yellow curve in the graph is fairly flat, so your estimate is not crucial.
On the basis of the graph, you determine that the probability of Mr. Brown having a bipolar variant, given his negative test -- despite your high clinical suspicion -- is 20%. (That's 1 minus the negative predictive value, which was 80% in this case). Note that this is not zero: bipolar has not been "ruled out". For comparison, recall that the probability of Ms. Anderson having a bipolar variant was 5% (1 minus the negative predictive value, which was 95% in her case), despite the same test result. The difference comes entirely from the different "pre-test probability", i.e. your clinical suspicion.
Positive predictive value
Let's conduct the same analysis for a positive MDQ result. See the negative predictive value steps above for the logic.
Here we go: your patient Ms. Carson has completed the MDQ and her result is "positive" (7/yes/yes). What do you tell her? Again you recall that the "predictive value" of a test depends on the prevalence of the illness. You estimate the appropriate "prevalence" for Ms. Carson by asking yourself: "what did I think was the likelihood of her having bipolar disorder before I did the test?"
As with Ms. Anderson above, you really didn't think there was reason to suspect bipolar disorder before you gave her the MDQ (you were just going along with the FDA recommendations). So you choose a conservative prevalence estimate, 0.1. In the graph below, you see that the "positive predictive value" of the MDQ at that prevalence is about 50%.
What should you tell Ms. Carson? How about something like this: "The test is telling us that an antidepressant may not be the best approach for you. I'd like you to learn about something we could call 'depression plus'. Technically it's called 'bipolar II', a version of bipolar disorder that does not have manic episodes. One way you can learn about this would be to go to the website shown here on your test, www.PsychEducation.org; read the section on 'mood swings without mania, bipolar II'. If you can't get to that website, let's schedule a longer appointment for you and I'll tell you more about it. This test result does not mean that's what you have, it just means we have to consider such a possibility. All this will make more sense when you know more about this bipolar II thing."
What if you had some suspicions about bipolar disorder before the MDQ result? As a final example, consider Mr. Dodds, who also has a positive MDQ at 7/yes/yes. In his case let's say you would have bet there was at least a 50% chance he has bipolar disorder. So you turn to that graph once more, using a "prevalence" figure of 0.5.
Because of that higher "pre-test probability", Mr. Dodds' likelihood of having bipolar disorder, given his test result, is about 90%. There is still a 10% chance the test is "wrong", and he should know that. Tell him to go learn about bipolar II, just as you instructed Ms. Carson. When he comes back, he's likely to say either "wow, that's me!" or "well, I see what you were concerned about, but you know, I really don't think I have bipolar II and I'd still like to try an antidepressant" -- or something to that effect. Now you can give him an antidepressant knowing you screened for bipolar disorder and educated the patient thoroughly about this issue, including the risks of antidepressants in this context. Just watch him closely, all right? See him back in no more than 2 weeks and tell him to call you if anything bad is happening, or anything "really good" either. (Dr. Gary Sachs, the bipolar specialist from Harvard, says in that case you should say: "that's great, let's celebrate, let's lower the dose of your antidepressant", and get him in to see you right away!).
Positive and Negative Predictive Value: a review of the concept, including the effect of disease prevalence
Where did this graph come from? How did we get here from "sensitivity and specificity?" To answer these questions, we have to review some things which you may know pretty well, or which might be but a dim memory (and perhaps not too fond, either). This all may be simpler than you think, let's see.
When you have a concern about a disease being present in a patient, you might order a test to look for it. Before the test, you have a suspicion, based on history or objective findings. You have established a "prior probability". Note that based on your suspicion, this probability is now higher than the prevalence of the disease in the general population. This is handy, because sometimes we actually have a number for that prevalence (a discussion of bipolar prevalence data follows below).
A test result simply raises or lowers your "prior probability", whether that was as low as the population prevalence, or higher based on some degree of clinical suspicion. After looking at the test result, you will have a "post-test probability" that is lower, or higher, than before the test. You will not have a diagnosis. You will have a different probability of that diagnosis.
The "predictive value" of a positive result refers to how far that result will increase the probability compared to your hunch. You can get a real number for this, and that resulting probability is what you're going to tell the patient.
Similarly, and only slightly more complex, a negative test result will shift your clinical suspicion as well. In this case, you just have to do one more calculation, subtracting from 1, not too bad. You take the "negative predictive value" number, e.g. from that graph I keep showing you, and subtract it from 1. The result is your new "post-test probability": again, your hunch plus a negative test result, a real number indicating the probability that the patient has the illness in question.
If you're like me, all this probability business may have you swirling slightly in a cloud of growing confusion. Hopefully you've already seen the most important point: underlying illness prevalence will change the meaning of a positive, or negative, test result. What you should tell the patient depends critically on what you thought was the likelihood of bipolar disorder before the test -- in other words, the prevalence of bipolar disorder in patients like her/him.
Technically, the graph I keep showing above comes from MDQ sensitivity and specificity data (which?) integrated with some prevalence assumptions, using the old "2 X 2 table". So, to really use that graph to decide what to tell your patient, you'll want to understand more about these prevalence assumptions.
Bipolar disorder prevalence
In 2001, the National Institutes of Mental Health summarized data on 1-year prevalence in 18-and-older U.S. residents as follows [update 2012 - more recent data are available; they make the same point, basically, although bipolar disorder is higher at 2.6%]:
Condition U.S. Prevalence
All Depressive Disorders
(Dysthymia, Major Depression,
9.5% Major Depression 5% Bipolar Disorder 1.2% Schizophrenia 1.1%
However, using "soft" bipolar criteria, Drs. Judd and Akiskal (the former a previous chief of the NIMH, the latter quite famous for very broad interpretation of bipolar disorder) found a lifetime prevalence of 6.4%Judd. Even more dramatic, again using lifetime prevalence data and very soft bipolar criteria, Angst et al found the following in a Swiss dataset: Angst
Condition Swiss Prevalence
All Depressive Disorders
(Dysthymia, Major Depression,
Minor and Recurrent Brief Depression)
24.6% Bipolar Disorder
(soft criteria, all variations)
Are you shocked? Surely you are. Could the Swiss really have that much trouble? The NIMH data suggest a ratio of bipolar-to-unipolar around 1 in 10, which might be pretty close to your old conception of things. But the Swiss data, using soft bipolar criteria, are virtually 1-to-1, one bipolar for every unipolar.
There are multiple estimates in between these two extremes, including a study of bipolar prevalence in the primary care setting, in which patients presenting with depression (or anxiety) were screened for bipolar II by a primary care doctor with training in recognizing bipolar variations.Manning That study found 26% had a bipolar disorder, roughly a midpoint (one in 4 depressed patients had bipolar disorder) between the NIMH and the Swiss studies.
Even for the most strident bipolar "advocates", the Swiss figure is hard to believe. Dr. Guy Goodwin, a fairly sober British psychiatrist and bipolar expert, proclaimed "the apparent size of this so-called bipolar spectrum invites disbelief." Here is his full-text editorial; it's worth a look if you're choking over this. He comments directly on Swiss study in the process of interpreting these kinds of prevalence estimates.
Thus we see a range of prevalence estimates, from roughly 0.1 (1 in every 10 depressed patients has bipolar disorder) to 0.5 (1 of every 2 depressed patients has bipolar disorder). These two estimates are the basis of the graphs above, which cover this range.
A Brief "Sensitivity Analysis"
Suppose you're really struggling with the "prevalence" thing. You don't know what number to use. How much difference does it make?
You can see from the yellow curve on this graph that "negative predictive value" is not very sensitive to prevalence estimates: the curve is fairly flat (at least compared to the blue one...). You could be completely at the wrong extreme of prevalence estimates and only change your interpretation of a negative test result from "5% likely" to "20% likely". Not a horrendous mistake if you were wrong: in both cases you'd probably proceed with an antidepressant medication, if the patient preferred a medication approach.
By comparison, the blue curve is quite steep. Your interpretation of a positive test will shift from "50% likely" to "90% likely" depending on your guess about "prior probability". What is a primary care doctor to do?
One solution is to avoid giving an antidepressant regardless of your prior probability guess. Refer positive MDQ cases to a psychiatrist if you can access one; or to a good therapist if you cannot. The therapist, in addition to trying to treat the depressive symptoms, can monitor for hypomanic symptoms as she/he goes along -- a much more sensitive way to look for such symptoms, at least if you know good therapists and can get the patient in to see them. You also can get more data to resolve puzzling results by asking the patient to have a significant other (spouse, parent, friend) complete an MDQ describing the patient.
If you must think about a medication approach, and cannot get another diagnostic opinion somehow, then you're in a tough spot -- especially if you weren't really thinking your patient had bipolar-like symptoms until the MDQ result emerged. That means the prior probability is at least on the lower side, shifting the likelihood of bipolar downward along the blue curve. At some point you'll have to decide whether you are willing to treat MDQ positives with antidepressant medications before giving a trial of a mood stabilizer; or whether you're willing to use mood stabilizers at all -- perhaps following some of the guidelines laid out in the section on Treatment in the Bipolar II section of this website.
You can stop here, unless you'd like to see citations for the literature on which all these analyses are based. Remember at least this:
Likelihood of bipolar disorder after MDQ testing
Your prior estimate of likelihood
Intermediate: 50-60 %
Validation study data
With two new studies almost reaching publication that provide yet more perspective on the sensitivity and specificity of the MDQ, this section has been revised and moved to its own page. You'll find studies by Miller et al, Ghaemi et al, and a commentary by Zimmerman et al with response from Hirschfeld et al, cited there -- as well as a surprising conclusion, namely that the sensitivity/specificity data one chooses may not affect how the test actually performs in primary care very much.