Journal of Psychosomatic Research 73 (2012) 401–407 Contents lists available at SciVerse ScienceDirect Journal of Psychosomatic Research Correspondence Are we witnessing the decline effect in the Type D personality literature? What can be learned? James C

Subject:WritingPrice:9.99 Bought8

Share With

Journal of Psychosomatic Research 73 (2012) 401–407 Contents lists available at SciVerse ScienceDirect Journal of Psychosomatic Research Correspondence Are we witnessing the decline effect in the Type D personality literature? What can be learned? James C. Coyne a, b,?, Jacob N. de Voogd a, c a b c Health Psychology Section, Department of Health Sciences, University Medical Center Groningen, University of Groningen, The Netherlands Department of Psychiatry, Perelman School of Medicine of the University of Pennsylvania, USA Department of Pulmonary Rehabilitation, Center for Rehabilitation, University Medical Center Groningen, The Netherlands a r t i c l e i n f o Article history: Received 2 July 2012 received in revised form 18 September 2012 accepted 21 September 2012 a b s t r a c t After an unbroken series of positive, but underpowered studies seemed to demonstrate Type D personality predicting mortality in cardiovascular disease patients, initial claims now appear at least exaggerated and probably false. Larger studies with consistently null ?ndings are accumulating. Conceptual, methodological, and statistical issues can be raised concerning the construction of Type D personality as a categorical variable, whether Type D is suf?ciently distinct from other negative affect variables, and if it could be plausibly assumed to predict mortality independent of depressive symptoms and known biomedical factors, including disease severity. The existing literature concerning negative affect and health suggests a low likelihood of discovering a new negative affect variable that independently predicts mortality better than its many rivals. The apparent decline effect in the Type D literature is discussed in terms of the need to reduce the persistence of false positive ?ndings in the psychosomatic medicine literature, even while preserving a context allowing risk-taking and discovery. Recommendations include greater transparency concerning research design and analytic strategy; insistence on replication with larger samples before accepting “discoveries” from small samples; reduced con?rmatory bias; and availability of all relevant data. Such changes would take time to implement, face practical dif?culties, and run counter to established practices. An interim solution is for readers to maintain a sense of pre-discovery probabilities, to be sensitized to the pervasiveness of the decline effect, and to be skeptical of claims based on ?ndings reaching signi?cance in small-scale studies that have not been independently replicated. © 2012 Elsevier Inc. All rights reserved. Introduction Many “discoveries” turn out to be exaggerated or simply false across diverse research areas and disciplines. John Ioannidis [1] provoked concern with his demonstration that of the 49 most cited clinical research studies in major journals, replications of 34 had been attempted, and 41% of key ?ndings had been refuted or shown to be substantially diminished. In a subsequent paper [2] he offered evidence that “most claimed research ?ndings are false” and he and others [3–6] have since identi?ed some mechanisms by which discoveries appear in the literature and then undergo a decline or refutation. Claims of breakthrough discoveries frequently arise in small positive studies that would seem to be too underpowered to detect an effect, and the inadequate sample size makes ?ndings all the more attention grabbing. Yet, portrayals in the literature of such “discoveries” ignore the larger context of a strong con?rmatory bias in published research papers, ? Corresponding author at: Department of Psychiatry, Perelman School of Medicine of the University of Pennsylvania, 3535 Market St. Rm 676, Philadelphia, PA 19104, USA. Tel.: +1 215 662 7035; fax: +1 215 349 5067. E-mail address: jcoyne@mail.med.upenn.edu (J.C. Coyne). 0022-3999/$ – see front matter © 2012 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.jpsychores.2012.09.016 with unknown numbers of negative ?ndings not being published and ?ndings being declared discoveries simply because of having achieved an arbitrary level of signi?cance. Moreover, statistically signi?cant ?ndings in small studies are necessarily always large because larger effect sizes are needed to cross the higher threshold for signi?cance. Apparent discoveries are thus created and perpetuated by a combination of con?rmatory bias, ?exible rules of design, data analysis and reporting [5,6], and signi?cance chasing [3]. There is a recurring decline effect [4] in diverse literatures, usually occurring when discoveries declared on the basis of small studies attract more resources or the attention of new investigator groups with less of an investment in the discovery, and attempted replications or extension of the original ?ndings fail. The rise and apparent decline of Type D personality Are we now witnessing a decline effect in the literature concerning Type D personality? The concept, which has been de?ned as the tendency to experience negative emotions and to inhibit self-expression in social interaction, was riding high as a promising prognostic indicator for mortality in cardiovascular disease (CVD) patients. A preliminary 402 J.C. Coyne, J.N. de Voogd / Journal of Psychosomatic Research 73 (2012) 401–407 study of post myocardial infarction (MI) patients had employed a different distressed personality type constructed by cross tabulating dichotomized continuous measures of trait anxiety and social inhibition, selecting the high/high quadrant and compared it to the other three [7]. Low exercise tolerance, a previous MI, anterior site of the MI, smoking, and age were entered into a logistic regression predicting 15 deaths and it was found that addition of the distressed personality type added to the prediction. A subsequent Lancet study constructed a Type D personality with different, but similar personality measures and claimed that patients with Type D personality were four times more likely be among the 21 dead during the observation period [8]. A later study claimed that patients with Type D personality were ten times more likely to die after a cardiac transplant [9]. Claims were also made that its predictive value was independent of disease severity and other biological variables. An unbroken series of ?ve published studies [8–12] found a signi?cant relationship between Type D personality and subsequent death. However, all studies were conducted by the same investigator group at Tilburg University, with very small numbers of deaths being explained in each study: 15 in the preliminary study [7], 21 in the Lancet study [8], followed by 6 [10], 4 [11], 6 [9], and 12 [12] respectively (See Table 1). These studies consistently found Type D to be associated with mortality, but none of the studies had enough deaths to reliably detect a signi?cant predictor of mortality with a strength similar to other predictors of mortality in heart disease, even if the effect was present. Observed magnitudes of effect were excessive relative to what could be reasonably expected and, in most cases, far stronger than traditional risk factors for mortality and recurrence, unless an extraordinary and unprecedented psychological risk for mortality had been uncovered. After the initial “discovery” of an effect on mortality in a sample with as few deaths to explain as was the case in Type D studies, the continued reliance on small studies implicitly committed investigators to the assumption that they were pursuing a replication of an effect that is larger than many traditional biomedical risk factors. In hindsight, this commitment should have been made explicit and should have required elaborate justi?cation beyond the initial ?nding with a small sample. Moreover, the statistical signi?cance of results in these earlier studies was typically vulnerable to reclassi?cation or addition or subtraction of a few or even a single death. Observed results might well be due to capitalization on chance or con?rmatory bias allowed by such factors as ?exible extension or contraction of follow-up periods; selective exclusion of some deaths; selective de?nition of endpoints (cardiac-speci?c mortality versus all cause mortality versus mortality included in a composite endpoint); selective inclusion of possible covariates and exclusion of others, depending on the impact on Type D signi?cance and magnitude; or reporting of signi?cance levels only for adjusted, rather than unadjusted analyses. For example, reports of these studies offered no indication that follow-up periods and their stopping points were predetermined or whether follow-up was stopped or extended, depending on whether a signi?cant prediction of mortality had been obtained. Subsequently, the Tilburg group produced three studies that had more deaths to explain and therefore a greater likelihood of reliably detecting the value of Type D as a prognostic indicator for mortality. The ?rst of the investigator groups' larger studies attempted to predict 47 deaths, but the prognostic value of Type D personality did not hold, once appropriate controls were introduced [13]. The second study attempted to predict 123 deaths among 641 heart failure patients, but Type D personality failed to be a signi?cant prognostic indicator in either bivariate or multivariate analyses with statistical controls [14]. The third study attempted to predict 187 deaths among 1234 consecutive CVD patients receiving a percutaneous coronary intervention and found no association between Type D personality and mortality [15]. Recently, there have been several studies from outside the original investigator group. The ?rst one failed to ?nd a prognostic value for Type D, but was underpowered to do so, with only 11 deaths being explained [16]. Then, another two more studies from outside the original investigator group were published, with larger numbers of Table 1 Studies assessing Type D and mortality in patients with cardiovascular diseases Publication Sample size Number of deaths Type D measure Type D prevalence Unadjusted effect size with con?dence interval Adjusted effect size with con?dence interval Number of covariates 27% nr nr 5 29% nr OR 4.1 CI 1.9–8.8 5 31% OR 11.65 CI 1.34–101.07 OR 4.84 CI 1.42–16.52 OR 2.51 CI 1.09–5.82 OR 16.5 CI 1.72–158.22 OR 2.16 CI 1.05–4.43 OR 1.16 CI 0.72–1.87 OR 1.40 CI 0.38–5.14 OR 0.84 CI 0.57–1.25 HR 0.89 CI 0.58–1.37 nr OR 8.9 CI 3.2–24.7 nr 5 HR 2.61 CI 1.12–6.09 HR 11.33 CI 1.24–103.3 HR 1.40 CI 0.93–4.29 HR 1.09 CI 0.67–1.77 HR 0.91 CI 0.25–3.32 HR 0.99 CI 0.61–1.59 HR 0.78 CI 0.49–1.24 HR 1.19 CI 0.76–1.85 9 Denollet 1995 [7] 105 15 Denollet 1996 [8] 210 21 Denollet 2000 [10] 319 6 SI: heart patients psychological questionnaire ; NA: trait STAI median split SI ≥12, NA≥ 40 SI: heart patients psychological questionnaire; NA: trait STAI; median split SI ≥12, NA≥ 43 DS16; median split NA ≥9, SI ≥15 Denollet 2006 [11] 337 4 DS16; median split A≥ 9, SI ≥15 29% Pedersen 2007 [12] 358 12 DS14; NA/SI ≥ 10 30% Denollet 2007 [9] 51 6 DS14: NA/SI ≥ 10 29% Schiffer 2010 [13] 232 47 DS14; NA/SI ≥ 10 21% Pelle 2010 [14] 641 123 DS14; NA/SI ≥ 10 20% Volz 2011 [16] 111 11 DS14; NA/SI ≥ 10 30% Grande 2011 [17] 977 172 DS14; NA/SI ≥ 10 25% Coyne 2011 [18] 706 192 DS14; NA/SI ≥ 10 13% Damen 2012 [15] 1234 187 DS14; NA/SI ≥ 10 29% nr 2 3 21 2 19 2 22 SI: social inhibition, NA: negative affectivity, STAI: State-Trait Anxiety Inventory, DS16: Type D scale 16-item version, DS14: Type D scale 14-item version, CI: con?dence interval, nr: not reported in original paper. N.B.: length of follow-up is not reported because descriptive statistics were inconsistent. J.C. Coyne, J.N. de Voogd / Journal of Psychosomatic Research 73 (2012) 401–407 deaths to be explained than most of the previous studies, but neither found evidence for a prognostic value of Type D personality in unadjusted or adjusted models, despite one having 172 deaths among 977 patients with CVD [17] and the other study having 192 deaths at 18 months follow-up of 706 HF patients [18]. Both of the studies relied on all cause, rather than cardiac-speci?c mortality as the primary outcome. The authors could justify the reliance on all cause mortality since cause of death is not reliably reported in death certi?cates and because of the unreliability of judgments that death is cardiac speci?c in an elderly population with numerous comorbidities who do not die at a research hospital. One of these studies [18] had a shorter follow-up period (18 months) than earlier studies and a lower prevalence of Type D personality. However, there was a nonsigni?cant trend in the wrong direction in terms of prediction of mortality from Type D personality and a large enough number of deaths being explained so that the possibility can be dismissed of the trend likely being reversed with continued follow-up. More importantly, this study [18] provided a critique of the methods used to identify Type D patients in previous studies and the over?tted regression models used in testing multivariate prediction controlling for potential confounders [19]. These factors would be expected to lead to spurious ?ndings of a prognostic value for Type D. These points were echoed and ampli?ed in an accompanying editorial [20]. Is a personality type tenable conceptually and statistically? The Tilburg group identi?ed patients as having a Type D personality if they were above the median [8,10] or, in later studies [9,13,14], if they scored above 10 on continuous measures of both negative affectivity (NA) and social inhibition (SI), and then compared the outcomes of these patients in the high/high quadrant to the other three quadrants. Such an analytic strategy for combining continuous predictor variables has been widely rejected for decades in the personality, education, and basic statistics literatures [21–27] because it likely leads to spurious results. There are two objections: splitting continuous variables into arbitrarily dichotomized categories, but also the isolation of the high/high category as a personality type. Whereas dichotomizing a single continuous variable loses information and statistical power, cross tabulation of a pair of dichotomized variables is likely to lead to in?ated estimates of statistical signi?cance and dramatically increased risk of spurious ?ndings [24]. Humphreys [27] declared the construction of a 2 × 2 matrix from continuous variables to be “unnecessary, crude, and misleading” (p. 874) and Cohen and Cohen [26] called it “an abuse of data” (p. 310). In labeling a person as an introvert or as neurotic, theorists adopted a typological language as “a verbal convenience rather than a meaningful mode of categorization” (p. 1159) [28] without any evidence that a speci?c demarcation point exists, at which persons above this point not only resemble each other, but differ from persons below it in crucial ways [21]. While the success of typological thinking in biology can be demonstrated in the classi?cation of plants and animals, examples in psychology are harder to identify. Thus, biological sex as male or female approaches a sharply de?ned category, but the representation of the psychological dimensions of masculinity and femininity does not [28]. Applied to the assessment of Type D personality, the implications are for a strong preference for preserving NA and SI as continuous variables and examining their interaction, rather than postulating a sharp distinction between high/high quadrant and the other three quadrants in a dichotomization of these variables. Among the problems in isolating the high NA/high SI quadrant to construct Type D personality is that NA and SI are moderately correlated and so patients are selected for greater distress than if either of the measures were considered separately. Scoring high on two correlated measures of distress is a more reliable indication of distress than being high on only one. Long-standing concerns about statistical and conceptual dif?culties of categorical/typological constructs have been substantiated in 403 recent work using sophisticated simulated comparison data techniques [29,30] that consistently favors dimensional conceptualizations. One systematic review [29] compared 377 articles with 311 distinct ?ndings and concluded: The domains of normal personality, mood disorders, anxiety disorders, eating disorders, externalizing disorders, and personality disorders (PDs) other than schizotypal yielded little persuasive evidence of taxa [categories]…. This review indicates that mostly variables of interest to psychiatrists and personality and clinical psychologists are dimensional, and that many in?uential taxonic ?ndings of earlier taxometric research are likely to be spurious (p. 903). Ferguson and colleagues [31] applied two taxometric procedures MAMBAC and MAXCOV to scores from the DS14, the standard measure of Type D personality and found clear evidence for a representation of it in dimensional rather categorical terms. The strong suggestion of these ?ndings is that Type D research should refocus on the additive and multiplicative in?uence of NA and SI in the context of other risk factors. While it is possible that in?ection points for the continuous variables could be identi?ed that maximize in?uence of these variables on particular CVD outcomes, the task remains of determining whether such in?ection points need to be identi?ed for NA and SI separately or whether there should be single or sliding cutpoints speci?ed for the interaction term with different outcomes. Rescuing Type D personality as an interaction term in a regression analysis of continuous variables? There are multiple reasons for doubting the validity of earlier claims from small studies that categorical Type D personality de?ned by the high/high quadrant predicted mortality. The two larger studies conducted outside the Tilburg group [17,18] had analyzed Type D personality data using both the scoring procedure of the original investigator group, but also by preserving NA and SI as continuous variables and examining their interaction effect, and in neither instance found a signi?cant bivariate or multivariate prediction of mortality. Could claims about Type D personality's independent prognostic value nonetheless be revived by reanalysis of the data from the earlier studies that preserved the continuous nature of component variables? Probably not. It is unlikely that a signi?cant interaction effect would support categorical over dimensional conceptions of Type D personality or that the optimal weighting of NA and SI would correspond to the arbitrarily chosen scoring of Type D personality in past studies. More basically, given the small number of deaths being explained in the positive studies, it is highly unlikely that the interaction term would prove signi?cant. Smith [20] pointed out that even if NA or SI alone or some additive combination predicted mortality, there was the risk of what Block [32] has termed the “jangle fallacy” of assuming that simply assigning a new name to previously studied traits constitute a discovery. However, Smith [20] provided a more profound criticism of relying on the one cell versus three (high NA and high SI versus the three other combinations of these variables) to test the speci?c prediction of a synergistic effect of NA and SI. The three versus one contrast could occur for numerous other reasons than a synergy of these two variables. Namely, a signi?cant contrast could be produced by an additive effect of either NA or SI alone; two additive ?rst order ef...