No question in psychiatric nosology has attracted such consistent controversyover thepast several generationsas the setting of the boundary between schizophrenia andmajormood disorders. Indeed, Kraepelin,1 the originator of this diagnostic distinction, was well aware of the difficulties of defining precisely this boundary. A closely related question has been whether there exists, between these 2 pivotal diagnostic entities, a third valid disorder that shares some clinical features with each: schizoaffective disorder (SAD). In this issue, Kotov and colleagues2 address these questionsusing thewell-knownSuffolkCountyMentalHealthProject cohort.A total of 526participantswere followedclosely for 4 years after admission, with their symptom course charted and 10-year outcome data obtained for 413 of these individuals. The percentages of their course that theywere psychotic, manic, and depressed were assessed, as was their nonaffectivepsychosis (NAP) ratio,definedas thepercentof their course when they were psychotic but not in a mood episode. The key parts of the article on which I will focus here are their results with the NAP ratio. This concept has a long and distinguished genealogy. The first operationalized diagnosis for SAD from the Research Diagnostic Criteria3 contained a “mainly schizophrenic” subtype, one criterion of which was thepresence of core psychotic symptoms for at least 1 week in the absence of major mood symptoms. When the authors of DSM-III-R4 struggled to provide operationalized criteria for SAD, which had been lacking in DSM-III,5 that would define a boundary with psychotic mood disorders, they adopted this as criterion B, increasing theminimal length to 2 weeks. This wasmaintained in theSADcriteria forDSM-IV6 andDSM-5.7We thereforehave a tradition dating back more than 35 years that the presence of psychotic symptoms for sustained periods outside of mood episodes is important in setting a boundary between psychotic mood disorders and more schizophrenia-like syndromes. Turning to theirmodel,Kotovandcolleagues sought touse the course of illness during the first 4 years to predict 10-year outcome.Adapting themodelof regressiondiscontinuity from Kendell andBrockington8 to theNAP ratio, they sought todiscriminatebetween3hypotheses that they termed linear,kraepelinian, and DSM-IV. That is, would they see (1) a smooth monotonic relationship between an increasing NAP ratio and poor outcome, (2) 2 flat regions of the regression line separatedby a sharp “step function” representingpsychoticmood disorders and schizophrenia, or (3) 3 flat regions separated by 2 step functions representingpsychoticmooddisorders, SAD, and schizophrenia? Before we turn to the empirical evaluation of these hypotheses by Kotov and colleagues, we should look at the distribution of their data for theNAP ratio (Supplement [eFigure 2, bottom right graph]). It is very U-shaped, which is a potential problem: 51% of their sample was psychotic only while in moodepisodes and21%hadonlynonaffectivepsychosis. This leaves just28%of theparticipants toprovide informationabout nonlinearity or discontinuities other than at the boundaries. Even this large study may be underpowered to evaluate the subtle but important distinctions between their hypotheses that they are seeking to evaluate. Using a statistical method of locally weighted scatterplot smoothing andpredicting 10-yearGlobalAssessmentof FunctionalPerformance levels, their key result is seen in thebottom rightgraphofFigure1.Whenexaminedclosely, thiscurveshows a gradually decliningGlobal Assessment of Functional Performance score at the 10-year follow-upas theNAP ratio assessed over the first 4years of illness rises fromzero to approximately 40%.The line thenflattens,meaningthatNAPratiosofapproximately40%havenopredictiverelationshipwithoutcome.These findings, which fit significantly better than a standard linear model, effectively ruleoutmodel 1, (whichbycontrast is clearly seen for the “%timepsychotic” factor [top left graphofFigure 1]). There is no single smooth continuum between NAP ratios andoutcome.The figure alsopresents 95%CIs asdotted lines. Noticehowwide theyare forNAPratiosof 10%to30%, indicating the sparsity of data in this critical region. Discriminating hypotheses 2 and 3 for the NAP ratio from these data is a subtler problem for which Kotov and colleagues use spline regression models (Table 3). With 4 fit indexes and4outcomemeasures,wehave 16 results for each of 8 models. For all 16, the 2 best-fitting models reflect hypotheses 2 and 3 (“2 flat kraepelinian” and “3 flat [DSM-IV]”), with 13 of them, often by verymodest amounts, favoringmodel 2, and 3 of them favoring model 3. Althoughmodels 2 and 3 are clearly better than the alternatives, andalthoughmodel 2 performs somewhatbetter thanmodel 3, the results donot, tomy eye, permit as confident a distinction between them as argued by the authors. My concerns are increased when examining the results from their best-fitting splinemodels (Figure 2). The figures show the predicted single-step function but at a very lowNAP ratio of 1.5%. This does not agreewellwith the results from the locallyweighted scatterplot smoothingmodels that show a curve declining rather steeply from NAP ratios of zero to 20%. Related article page 1276 Opinion