Discussion: View Thread

AUC vs Mixed Modeling

  • 1.  AUC vs Mixed Modeling

    Posted 10-06-2011 07:50

    Hi Everyone,
    I too am appreciative of your thoughful exchange of views.  I have a question for you to consider -- in critquing statistical methodology supporting the comparison of longitudinal patient reported outcome by intervention, I frequently encounter the use of the AUC as a summary measure, with group comparisons based on a two-sample t-test.  I consistently recommend the use of mixed modeling to take into account full information and increased power.  I understand that missing data in this scenario is frequent and should be considered in the choice of appropriate methodology.  Can you comment on various situations when you might recommend one approach over the other?  Thanks so much.

    -------------------------------------------
    Alexandra L. Hanlon
    Associate Research Professor
    University of Pennsylvania
    -------------------------------------------


  • 2.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 08:15
    I agree with you about the mixed model.  I think the mixed model is the best approach and it does handle missing data as long as the missing at random assumption is valid.  If you have non-ignorable missingness I think any method that does not model the mechanism for missingness will have problems.  So I don't understand why the AUC approach would be appropriate.

    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 3.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 09:21

    Completely agree with Michael & Alex, nice work & keep going with these recommendations. There are of course a ton of books out there on the subject, but LDA references I usually go with are Diggle, Heagerty, Liang & Zeger (2002) and the Verbeke & Molenberghs duo on linear & discrete longitudinal models (2001 & 2005). For this particular question though (paired ttest vs mixed model), Robert Weiss has a nice chapter on "Critques of Simple Analyses" in his book: Modeling Longitudinal Data (2005). 

    -------------------------------------------
    Michael Griswold
    Executive Director
    Univ MS Medical Biostatistics
    -------------------------------------------








  • 4.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 09:42
    I am familiar with many of the recommended references.  They are all good recommendations.  Places wher use of AUC show up and is appropriate are bioequivalence testing of two treatments and comparing Operating Characteristic curves.  For longitudinal data, mixed linear models with repeated measures is usually the choice.

    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 5.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 12:36
    I'm afraid I don't understand the arguments against using AUC and feel that a blanket statement like "mixed models are better than AUC" is unwarranted.  A mixed model will not be more powerful if it is not appropriate to model the data at hand.  AUC is one of the most commonly used measures of exposure in pharmacokinetics and has an important interpretation.  It can be estimated after fitting a mixed model or in a non-compartmental model (it is just the area under the curve after all) and both methods can handle missing data.  For pharmacokinetic data, modeling concentrations over time can be complex and the non-parametric model (the non-compartmental model) may be the best model. I think the answer to what is the best model depends on the data and the question of interest: are you more interested in trends over time or exposure?

    -------------------------------------------
    Colleen Kelly
    Principal Consultant
    Kelly Statistical Consulting
    -------------------------------------------








  • 6.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 13:00

    How do you explain or interpret, an AUC of quality of life?  Its a patient reported outcome, so to speak, how the patient feels. An example, what would the AUC of the  SF-36 domain,  Mental Health, mean?

    -------------------------------------------
    Chris Barker, Ph.D.
    President - San Francisco Bay Area Chapter of the American Statistical Association
    www,barkerstats.com
    -------------------------------------------








  • 7.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 13:22
    I'm not saying that AUC can be applied to every problem.  However, I think there would be examples of patient reported outcomes that, similar to the situation in PK, might be best interpreted with an AUC.  For example, if a treatment aims to change a behavior like eating or exercise, there could be a lot of variation and no clear trend in how the treatment affects the behavior.  In this case, you might be more interested in the total exposure to calories or exercise during the treatment period (which I think would correlate with weight loss) rather than the trend.

    -------------------------------------------
    Colleen Kelly
    Principal Consultant
    Kelly Statistical Consulting
    -------------------------------------------








  • 8.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 13:33

    I appreciate Colleen's point about AUC.  We should not reject it unequivocally.  In my message I pointed out to applications wher comparing two AUCs is important.  My position was based on the assumption that the data was appropriate for mixed effects linear models or else Alexandria would not be comparing the two approaches.  In such cases especially if AUC is not the primary endpoint (or quantity of interest) the mixed model is probably better for handling missing data.  In calculating AUC for cases with missing time points for the concentrations (linear interpolation or some other smoothing/interpolation method)?  If several time points are missing I think this could be a very crude and possibly inaccurate way to estimate AUC.  Of course any missing data method will have problems when there is a lot of missing data.
    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 9.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 13:53

    I agree as well and also did not intend a blanket statement against AUC. I adhere to the general statistical consulting rule that the answer is usually "well, it depends". I think both Colleen and Michael raise good points for when and where AUCs can be of interest (bioequivalencepharmacokinetic, total exposures vs trends, etc.), as well as potential issues to consider (missing data, interpretability, etc.) I myself find that it is usually fairly simple these days to fit a variety of models and examine the results and interpretations across them to form more interesting conclusions. I generally try to adhere to an extension of Box's famous quote: "All models are wrong, but some models are useful"; and many useful models can be illuminating.

    -------------------------------------------
    Michael Griswold
    Executive Director
    Univ MS Medical Center Biostatistics
    -------------------------------------------








  • 10.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 14:23

    As a point of reference for the discussion of AUC"s for Patient Reported Outcomes.  For the stat. analysis of  PRO's in a -regulatory- setting, where a sponsor wants to get a "label claim", below is a link to the FDA guidance on PRO's.  Among the topics, It discusses missing data, etc.
     
    Pertinent to the AUC - see  page  3, "evaluation of an instrument" and the reference to the conceptual model and later on "endpoint model" and Figure 1 and 2

    As to the AUC, what is the conceptual model? These are fundamentally psychometric or "measurement"  issues.
    Page 18, refers to the content validity, reliability etc. One would have to define/measure these for the AUC of a PRO domain.

    http://www.fda.gov/downloads/Drugs/GuidanceComplianceRegulatoryInformation/Guidances/UCM193282.pdf


    -------------------------------------------
    Chris Barker, Ph.D.
    President - San Francisco Bay Area Chapter of the American Statistical Association
    www,barkerstats.com
    -------------------------------------------








  • 11.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 14:24

    Thanks for the discussion. The specific scenario of interest here is to determine whether an intervention is effective in reducing pain.  Pain is measured repeatedly over treatment using one item having a five point Likert scale. The proposed primary analysis relies on a two-sample t-test comparing mean AUC pain score.
    -------------------------------------------
    Alexandra L. Hanlon
    Associate Research Professor
    University of Pennsylvania
    -------------------------------------------








  • 12.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 14:51
    Alexandra, Sometimes semantics gets in the way of this.  Of course, depends on what AUC and mixed models are being used.  We have used "AUC" like measures in cases like this (integrated scores).  For example, suppose you had pain readings at times t=4,6,8, and 10. There may not be a clear horizon that is the most important--or more often the team doesnt know where it is the most effective.  You could construct an AUC like measure, that is the average score at each time period, AUC = Y4 + Y6 + Y8 + Y10 (can divide by 4 or not).  Then compare the mean of this AUC variable using a means test.  This quantity has a pretty clear meaning--correlations go away and get added over to inflate the variance of the AUC.  The change from one treatment to another has a pretty clear meaning. If two time periods are better for a treatment and two are the same -- you can still have a difference between the groups.  You wouldnt understand the granularity of the effect from this analysis -- but it can "work".  It avoids a priori selecting one time period -- say t6 -- and missing with no effect there.  If there is a consistent effect over time it can become stronger with this type of model.  Many measures are already AUC liek measures -- averaging pain over 24 hours or a 7 day period.  The visits you have are probably 7-day (14?) diaries.

    You could also use a model that has subject and each visit as a factor.  There may be additional correlation after the subject specific effects or not.  This is a better representation of the error/correlation at each visit, especially if you have any prediction or confidence intervals to construct.  Usually this model will have a "constant effect" -- additive -- of the treatment.  This is not "too dissimilar" from the AUC model.  Once you start allowing interactions -- or treatment effects that can vary at each visit things get very hard--and VERY difficult to interpret--from a testing viewpoint.  You could conclude treatment 1 is better at t=4,8 and treatment 2 is better at t=6,10.  Of course you couldn't even do much testing in the AUC.

    So, do I have a point?  I think most statisticians would prefer the latter of these in most cases, but I suspect you wont get much difference in results.  Missing data is important here for sure.  How you construct the AUC with missing data.  What you do in the model approach is probably easier, and the latter provides nice alternatives (imputation) to LOCF. It would be neat to see some simulations or work to see which is more powerful under different assumptions of effects at the time periods.  

    We have used both models in similar cases.

    -------------------------------------------
    Scott Berry
    Berry Consultants
    -------------------------------------------








  • 13.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 15:14
    Wow.  Such a simple question.  It is amazing that it started such an interesting discussion.  Thanks Scott.

    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 14.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 15:34

    Regarding AUC, I like its use in evaluting operating characteristic curves such as in the case of the sensitivity and specificity of a diagnostic tool.  There the AUC can be compared to chance as a chance OC curve would be a 45 degree line with AUC=0.50  whereas a good diagnostic tool would have an AUC much greater than 0.50.
    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 15.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 16:16

    I mostly use it in ROC diagnostic applications as well. Even then I recommend folks look at the whole ROC as well and not just the AUC, and have discussions with substantive collaborators on whether Sensitivity and Specificity are equally important or whether one is more valued than the other (such as in screening populations). I commonly give the same caution in areas such as the repeated pain score application that Alexandra brings up (nice discussion by Scott below). Looking a the full profile data through Exploratory Data Analysis techniques (Tukey, etc.) informs appropriate analyses.

    Consider 3 individuals on different treatments measured 5 times on the 5 point likert pain scale with trajectories

     Patient 1 (trt A): 5, 4, 3, 2, 1 (decreasing pain)
     Patient 2 (trt B): 3, 3, 3, 3, 3 (constant pain)
     Patient 3 (trt C): 1, 2, 3, 4, 5 (increasing pain)

    All AUCs are the same, if trajectories like these are common within each treatment, then blindly calculating the AUCs and running tests to examine significance would likely lead to no perceived differences across treatments, even though the differences are clearly there. In this case, mixed models examining trends would likely be preferred to AUCs (even if the apriori specified analysis plans specified AUC approaches). If instead the trajectories were more constant and additive (as Scott mentions)

     Patient 1 (trt A): 1, 1, 1, 1, 1 (lower pain)
     Patient 2 (trt B): 3, 3, 3, 3, 3 (middle pain)
     Patient 3 (trt C): 5, 5, 5, 5, 5 (higher pain)


    Then I'd expect the mixed model and AUC overall analyses to give comparable testing results (if missing data were not an issue), but would potentially prefer the mixed model interpretations. Given that pain scores generally have high degrees of informative missingness, I'd likely go with a mixed model (MAR) approach with some patter-mixture or alternative informative missingness sensitivity approaches.

    Regardless, I'd always start with the EDA and show my audience as many visual results as possible to support my analytic framework.

    -------------------------------------------
    Michael Griswold
    Executive Director
    Univ MS Medical Center Biostatistics
    -------------------------------------------








  • 16.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 16:32

    I think it is very nice to see examples like this demonstrating what good consulting is all about.  Our young statisticians should take note.  Classification problems with 2 classes are essentially the same as diagnostic problems except for the terminology describing the two types of errors.  Tradeoff discussion with regard to the two type of misclassification are very important too.  A problem I encountered a lot when I worked in space defense was the identification of reentry vehicles from decoy balloons based on their spectral signatures.  This was in regard to the Reagan strategic defense initiative's (STAR WARS) proposed missile defense system.  The classification errors are (1) classifying an RV as a decoy and (2) classifying a decoy as an RV.  (1) is a more serious error than (2) because letting an RV through the system results in a city being destroyed while (2) just wastes a shot.  Of course (2) cannot be ignored because you do not want to exhaust all your shots before all the missiles have come through.
    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 17.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 16:40


    -------------------------------------------
    J. Dobbins
    Delmarva Foundation
    -------------------------------------------

    While I maintain my original comment about all these consulting group emails being very interesting, it seems like they are landing more frequently than flights at O'Hare field or Atlanta and they light up on the screen as they come in.  Does anyone know how to keep the comment but turn off "the show the comment coming in feature" on days where you are really busy?






  • 18.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 16:48

    There are two ways you can control the messages from comming:

    You can change the settings in the ASA site: under My Profile>My Preferences
    or
    You can change turn off the desktop notifications of whichever email server you are using.


    Hope this helps.

    -------------------------------------------
    Shahidul Islam, MPH
    Biostatistician
    Winthrop University Hospital
    Mineola, NY 11590
    -------------------------------------------








  • 19.  RE:AUC vs Mixed Modeling

    Posted 10-06-2011 08:29

    I often work with PRO's. An AUC makes some kind of sense for a blood level measurement as a type of measurement of total exposure.  There is no such "exposure " interpretation  or other psychometric interpretation that I know of for  the AUC.   I always recommend against, but I don't always win the argument. :)


    I think the AUC interpretation arises out of a lack of familiarity with PRO's and their interpretation.  As it can be analyzed with a t-test is a "simple" analysis. 

    The AUC misses a lot of important information. One can have two patients with identical AUC's but completely different trajectories over time.

    As if not more important, is whether the study uses the proper PRO instrument, and collects that data at the appropriate times.

    Despite that, there may be some use for an estimate of the AUC in some versions of a Health Economics Cost Effectiveness, or QALY (quality adjust life year). However, in the CE, the AUC is only a component of a calculation.

    An FYI, a great text for you on PRO's and their analysis using a mixed model is by Diane Fairclough "Design and analysis of quality of life studies"

    -------------------------------------------
    Chris Barker, Ph.D.
    President - San Francisco Bay Area Chapter of the American Statistical Association
    www,barkerstats.com
    -------------------------------------------








  • 20.  RE:AUC vs Mixed Modeling

    Posted 10-07-2011 08:32

    It's safe to say that we all agree that the mixed model approach is superior methodology for such data.  But the real issue may be acceptability by the research community or more importantly by the regulatory agency (FDA) if the ultimate goal is drug/device approval.  I have come across several situations where I would rather have done other analyses, but the reviewer at the FDA was unwilling to accept some other than what he was used to reviewing...and yes, he was the statistical reviewer.
    -------------------------------------------
    Susan Spruill
    Statistical Consultant
    -------------------------------------------








  • 21.  RE:AUC vs Mixed Modeling

    Posted 10-07-2011 08:56
    Thanks, Susan, as well as the rest of you weighing in on this question.  As I suspected, you contributed quite thoughtfully to validating my initial thoughts on the AUC approach.  I was interested in hearing your views on both sides.  I have wanted to recommend that the analysis rely on a comprehensive approach to provide the entire picture in a meaningfuil way, and this can include the AUC with assumptions, visualization, and certainly mixed modeling to take advantage of full data, as well as careful consideration and attention to missing data along with sensitivity analyses.  As you suggest, the problem here is that the higher order is interested in definitive findings on the basis of one primary outcome measure, and one analytic approach.  I think I will still push for the full picture.  Thanks again.

    -------------------------------------------
    Alexandra L. Hanlon
    Associate Research Professor
    University of Pennsylvania
    -------------------------------------------








  • 22.  RE:AUC vs Mixed Modeling

    Posted 10-07-2011 09:37
    There is no law that says you cannot argue with the FDA.  It may not always work but sometimes in meetings the supervisor might overrule.  I just think if you have a strong case you should argue.  There is nothing to lose.  If you don't win the day you go with their direction as long as it is not crazy.

    -------------------------------------------
    Michael Chernick
    Director of Biostatistical Services
    Lankenau Institute for Medical Research
    -------------------------------------------








  • 23.  RE:AUC vs Mixed Modeling

    Posted 10-11-2011 01:00
    I recently encountered a related problem. I was trying to compute the required sample size for an equivalence study where the goal is to exhibit equivalence at every visit, with respect to a clinical variable. (1) I could  convert the problem to a univariate problem by comparison of average AUC; this is a straight-forward sample size calculation, assuming that I have a value for the variance, but the analysis of AUC won't answer the question. (2) Alternatively, the repeated measures analysis could be used to answer the question. Construction of the confidence interval for the treatment effect could be used for sample size determination and analysis. (3) The purist in me considered a confidence region for the vector of differences of means using Hotelling's t2. This might  be good for analysis, but not useful for sample size. I used the confidence intervals based on Hotelling's t2, which results in a v-dimensional "box" instead of an ellipsoid, where v is the number of visits, and is independent of the correlations. However, I continue to think that the repeated measures approach could be appropriately used (assuming no time-by-treatment interaction). Has anybody used this approach to this problem?

    -------------------------------------------
    David Bristol
    Statistical Consulting Services
    -------------------------------------------








  • 24.  RE:AUC vs Mixed Modeling

    Posted 10-07-2011 09:35
    A nonparametric approach is perhaps another option for longitudinal data with an ordinal response.  The reference is Brunner, Domhof & Langer.  Nonparametric Analysis of Longitudinal Data in Factorial Experiments.  The analyses can be done with SAS Proc Mixed, or with R scripts available from Brunner's website.

    Whether or not it would be acceptable to FDA or other agencies is, as Susan indicates, is another issue.
    But I do like Edgar Brunner's approach.  Larry Madden and I wrote a layman's paper (for agricultural scientists) illustrating the methodology with agricultural datasets (in Phytopathology, 2004). 

    BTW, I'm one of those silent followers, but do enjoy and value the discussions very much.

    -------------------------------------------
    Denis Shah
    Independent Consultant, Agricultural Sciences
    -------------------------------------------