UBC Faculty Research and Publications

Latent variable mixture models to test for differential item functioning: a population-based analysis Wu, Xiuyun; Sawatzky, Richard; Hopman, Wilma; Mayo, Nancy; Sajobi, Tolulope T; Liu, Juxin; Prior, Jerilynn; Papaioannou, Alexandra; Josse, Robert G; Towheed, Tanveer; Davison, K. S; Lix, Lisa M May 15, 2017

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata


52383-12955_2017_Article_674.pdf [ 1.03MB ]
JSON: 52383-1.0347535.json
JSON-LD: 52383-1.0347535-ld.json
RDF/XML (Pretty): 52383-1.0347535-rdf.xml
RDF/JSON: 52383-1.0347535-rdf.json
Turtle: 52383-1.0347535-turtle.txt
N-Triples: 52383-1.0347535-rdf-ntriples.txt
Original Record: 52383-1.0347535-source.json
Full Text

Full Text

RESEARCH Open AccessLatent variable mixture models to test fordifferential item functioning: a population-based analysisXiuyun Wu1,2, Richard Sawatzky3, Wilma Hopman4, Nancy Mayo5, Tolulope T. Sajobi6, Juxin Liu7, Jerilynn Prior8,Alexandra Papaioannou9, Robert G. Josse10, Tanveer Towheed11, K. Shawn Davison12 and Lisa M. Lix13*AbstractBackground: Comparisons of population health status using self-report measures such as the SF-36 rest on theassumption that the measured items have a common interpretation across sub-groups. However, self-reportmeasures may be sensitive to differential item functioning (DIF), which occurs when sub-groups with the sameunderlying health status have a different probability of item response. This study tested for DIF on the SF-36physical functioning (PF) and mental health (MH) sub-scales in population-based data using latent variable mixturemodels (LVMMs).Methods: Data were from the Canadian Multicentre Osteoporosis Study (CaMos), a prospective national cohortstudy. LVMMs were applied to the ten PF and five MH SF-36 items. A standard two-parameter graded responsemodel with one latent class was compared to multi-class LVMMs. Multivariable logistic regression models withpseudo-class random draws characterized the latent classes on demographic and health variables.Results: The CaMos cohort consisted of 9423 respondents. A three-class LVMM fit the PF sub-scale, with classproportions of 0.59, 0.24, and 0.17. For the MH sub-scale, a two-class model fit the data, with class proportions of 0.69 and 0.31. For PF items, the probabilities of reporting greater limitations were consistently higher in classes 2 and3 than class 1. For MH items, respondents in class 2 reported more health problems than in class 1. Differences initem thresholds and factor loadings between one-class and multi-class models were observed for both sub-scales.Demographic and health variables were associated with class membership.Conclusions: This study revealed DIF in population-based SF-36 data; the results suggest that PF and MH sub-scalescores may not be comparable across sub-groups defined by demographic and health status variables, althougheffects were frequently small to moderate in size. Evaluation of DIF should be a routine step when analysingpopulation-based self-report data to ensure valid comparisons amongst sub-groups.Keywords: Latent class analysis, Item response theory, Mental health, Patient-reported outcome measures, Physicalfunctioning, Population health* Correspondence: lisa.lix@umanitoba.ca13Department of Community Health Sciences, University of Manitoba,S113-750 Bannatyne Ave, Winnipeg, MB R3E 0W3, CanadaFull list of author information is available at the end of the article© The Author(s). 2017 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, andreproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link tothe Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver(http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated.Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 DOI 10.1186/s12955-017-0674-0BackgroundSelf-report health status measures, such as the SF-36,are frequently used to compare the health and well-being of different populations and to establish popula-tion norms for comparative investigations [1]. To makeaccurate comparisons, the scores on these patient-reported outcome measures (PROMs) must be valid andreliable. While many PROMS have been subjected torecommended psychometric evaluations of their con-struct validity, test-retest reliability, and interpretability[2–4], they are less frequently evaluated for differentialitem functioning (DIF), a form of measurement non-equivalence. Specifically, measurement non-equivalence(or non-invariance) exists when a PROM does not meas-ure the same construct (e.g., quality of life) across differ-ent population sub-groups [5, 6]. DIF arises whenindividuals with the same underlying level of health re-spond differently to the items that comprise a healthmeasure; these differences are often associated with re-spondent characteristics such as sex and age [7]. If thereis no evidence of DIF, individuals with the same charac-teristics are expected to produce a similar pattern of re-sponses on the PROM items. However, when there isevidence of DIF, differences on the PROM items may bean artifact of the measurement process.Measurement non-equivalence and DIF have been ex-amined for the SF-36 and SF-12 in population-baseddata [8–12]. DIF has been detected in both instruments[8, 10–12].Item response theory (IRT) models, which were origin-ally developed for application to achievement tests inthe field of education and psychology, are commonlyused to test for DIF [13] amongst manifest (i.e., ob-served) groups, such as males and females. However,comparisons of item responses for manifest groups usingconventional IRT models may lack sensitivity to detectthe true source(s) of DIF [14–17]. Specifically, DIF maybe associated with unobserved (i.e., latent) characteris-tics, as well as observed characteristics. Alternative ana-lytic techniques to conventional IRT models are neededto disentangle the effects of latent and observed charac-teristics [14–17].Latent variable mixture models (LVMMs) have re-cently been proposed to detect DIF-related bias in self-report measures [10, 14–21]. LVMMs combine latentclass techniques with IRT; these models do not assumethat manifest variables fully account for differential pat-terns of item functioning [15, 16]. A LVMM for DIFseeks to identify heterogeneous latent groups (i.e., latentclasses) in a population with distinct patterns of item re-sponses. The LVMM assumes that the IRT measurementmodel holds within each latent class, but allows for dif-ferent sets of item parameters among the latent classes.LVMMs are well suited to the analysis of population-based data, which are likely to exhibit substantial hetero-geneity in patterns of item responses.These models arevaluable because they can be used to generate hypoth-eses about population characteristics that may be associ-ated with DIF [10].The purpose of this study was to test for DIF on self-report measures of physical and mental health in a diversepopulation-based cohort using LVMMs. Specifically, wefocus on the SF-36 physical functioning (PF) and mentalhealth (MH) sub-scales. These sub-scales were selected forinvestigation because previous studies have demonstratedtheir sensitivity to DIF using manifest variable models [8, 9,11, 22, 23]. An investigation of whether LVMMs will detectDIF in these two SF-36 sub-scales can help to elucidate thecauses of DIF and facilitate the interpretation of DIF and itsimpact on the validity of the SF-36.MethodsData sourceStudy data were from the Canadian Multicentre Osteopor-osis Study (CaMos), a population-based prospective cohortstudy that was initiated to provide unbiased national esti-mates of the prevalence and incidence of osteoporosis [24].Individuals were recruited without regard for disease statusbut had to live within a 50-km radius of one of the ninestudy sites, which were primarily in urban centres. Baselinedata, which were the focus of this analysis, were collectedin 1996 to 1997. Details of the data collection methods andsample characteristics have been published elsewhere [24].We included respondents 25 years of age and older.MeasuresThe SF-36 version 1 was used in the CaMos; it encom-passes eight sub-scales: PF, MH, role physical, bodilypain, general health, vitality, social functioning, and roleemotional. The PF sub-scale is comprised of 10 itemswith three response options: limited a lot, limited a lit-tle, and not limited at all. The MH sub-scale is com-prised of five items evaluated on a six-point scale withendpoints of all of the time and none of the time [1, 25].The study cohort was described on the demographicvariables of sex, age, and education level, and on thehealth status variables of self-reported general health sta-tus, measured height and weight from which body massindex (BMI) was computed, and health utilities. The lat-ter were measured via the Health Utilities Index Mark 3(HUI3), a preference-based measure that captures func-tional performance in vision, hearing, speech, ambula-tion, dexterity, emotion, cognition, and pain [26]. HUI3summary scores were estimated using a multi-attributeutility function [27]; a score of zero corresponds todeath and 1.00 corresponds to a complete absence ofimpairment. Negative scores are possible and correspondto a severely impaired health state worse than death.Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 2 of 13Statistical AnalysesDescriptive statistics were used to characterize the studysample on the PF and MH item responses, demographic,and health status measures. Item responses with posi-tively worded formats (i.e., MH3 and MH5) were reversecoded prior to analysis.Analyses were conducted separately for the PF and MHsub-scale items. Unidimensionality, the assumption thatall items measure a single construct, was investigated by adescriptive analysis of the standardized residuals of itemresponses for the one-class model [28]. A χ2 statistic wasused to test for differences between predicted and ob-served responses. We also examined unidimensionalityusing exploratory factor analysis with oblique rotation ap-plied to the polychoric correlations for the items [29]. Theroot mean square error of approximation (RMSEA) andcomparative fit index (CFI) were used to assess model fit.A RMSEA value ≤0.10 [30], and a CFI value >0.90 [31] in-dicate acceptable model fit.Local independence of the items is satisfied if the re-siduals of items are not correlated, conditional on thecommon latent factor (i.e., P[R1, R2 | θ] = P[R1| θ] P[R2|θ] where θ represents underlying health status and R1and R2 are residuals for scale items 1 and 2. In otherwords, the joint probability of correct responses for anitem pair will be a product of the probabilities of correctresponses to the two items, conditional on the latent fac-tor. We evaluated local independence by comparing thepredicted and observed proportions of responses foreach pair of items in a sub-scale in all the models usinga χ2 statistic [32].The LVMM combines a latent class model with anIRT two-parameter graded response model (GRM) [33].This GRM is specified as:Pij Xi≥jjθ;aið Þ ¼exp αi θ−βij h i1þ exp αi θ−βij h i ; ð1Þwhere Pij is the cumulative probability that a person re-ceives a score on the jth or higher category (j = 1, 2,…, J)for item i (i = 1, …, I). In Eq. 1, θ represents underlyinghealth status, βij is the item response parameter (also re-ferred to as the difficulty parameter) for category j or aboverelative to lower categories and αi is the item discriminationparameter (i.e., slope) for item i indicating the relationshipbetween the item and the latent construct [33].We adopted Muthén’s LVMM framework [34], whichestimates factor loadings (λ) and item thresholds (τ);these can be converted to item discrimination and diffi-culty parameters [35]. The item factor loadings andthresholds in the LVMM are allowed to vary across twoor more latent classes (see Figure in Appendix).The conditional cumulative probability of the item re-sponse within the mth latent class is estimated by.Pijm Xi≥jjθ;C¼mð Þ ¼exp λim θ−τijm 1þ exp λim θ−τijm  ð2ÞIn Eq. 2, C denotes the latent class variable (m = 1,…,M), λim is the estimated factor loading for item i withinclass m, and τijm is the estimated threshold for responsecategories at or above category j for item i within classm. The cumulative probability of an item response forcategory j and higher for a respondent is estimated by:Pij ¼XMm¼1 πm  Pijm Xi≥jjθð Þ ; ð3Þwhere πm is the posterior probability for a respondentbelonging to the mth latent class [35]. The posteriorprobability was estimated using Bayes’ theorem [36]. Re-spondents were assigned to the latent class with the lar-gest posterior probability among the classes.To examine DIF and its impact on the underlying la-tent health variable (i.e., PF or MH), we adopted a four-step procedure [10]: (a) fit the standard IRT two-parameter GRM (i.e., one-class model), and multiple-class LVMMs to identify the best-fit model, (b) test fordifferences in the model parameters across the latentclasses, (c) characterize potential sources of DIF by com-paring within-class item response percentages and testassociations of class membership with the covariates,and (d) examine the impact of DIF by comparing pre-dicted factor scores between the one-class GRM and themultiple-class LVMM.A one-class model was compared to two-, three-, andfour-class models to determine the optimal number oflatent classes. The statistics used to select the best-fitmodel included the Bayesian Information Criterion(BIC) [37, 38], Vuong-Lo-Mendell-Rubin likelihood ratiotest (VLMR) [39] and bootstrap likelihood ratio test(BLRT) [37]. A good-fitting multi-class model with esti-mated class-specific item parameters provides evidenceof latent DIF [20]. Quality of class assignment was evalu-ated with the entropy measure, which ranges between 0and 1; larger values indicate a higher proportion of cor-rect classifications [40]. The maximum likelihoodmethod with robust standard errors was used to esti-mate the model parameters [29]. In order to identify themodel, the latent factor mean was set to zero and thevariance was set to one across classes [34], which is inaccordance with conventional IRT parameterization forDIF [34, 41, 42]. The model allowed the item parametersto vary across latent classes.Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 3 of 13The differences in threshold parameters amongstthe latent classes were tested for each item using alikelihood ratio (LR) test [37, 43]. Specifically, amodel with free threshold parameters for an item wascompared with a model in which the threshold pa-rameters were constrained to be equal across classesfor an item. A statistically significant difference be-tween the two models indicates uniform DIF for theitem [7].Multinomial logistic regression with pseudo-class ran-dom draws was used to characterize the latent classes onselected demographic and health status variables. Agewas categorized as younger (25–64 years) and older (65+years). BMI was categorized as overweight and obese(BMI ≥ 25.0) versus underweight and normal weight(BMI < 25.0) [44]. Based on the distribution of theHUI3, scores were categorized as low (<0.8) and high(≥0.8). Odds ratios (ORs) and 95% confidence intervals(95% CIs) were reported.Predicted factor scores and the most likely class mem-bership for each respondent were obtained from thebest-fitting LVMM. The reliability of the predicted factorscores was evaluated by the conditional standard errorof measurement (CSEM), which was calculated as theinverse of the square root of the information function[45]. Skewness was examined for the distribution of thepredicted factor scores [46].SAS 9.3 was used to prepare the data and conduct thedescriptive analyses [47]. The LVMM analysis was con-ducted with Mplus version 7.11 [29].ResultsDescription of cohortThe CaMos cohort was comprised of 9423 respondents.In total, 9337 (99.1%), and 9395 respondents (99.7%)had complete data on the 10 PF items and five MHitems, respectively.For the PF items, close to half (43.9%) of respondentsreported being limited a lot in vigorous activities, whileless than 6.0% of respondents reported being limited alot when climbing one flight of stairs, walking one block,or bathing/dressing (Table 1). For the MH items, lessthan 2.0% of respondents reported that they were ner-vous, so down in the dumps that nothing could cheerthem up, or feeling downhearted and blue all of thetime.Examination of unidimensionality and local independenceFor the PF sub-scale items, the IRT two-parameterGRM fit the data well as evidenced by absolute values ofstandardized residuals less than 1.96 for all items. Thedifference between observed and predicted propor-tions of item responses was not statistically significantTable 1 Distribution of item responses (%) for the SF-36 sub-scale itemsItem Response optionPhysical Functioning (N = 9337) Limited a lot Limited a little Not limited at allPF1: Vigorous activities 43.9 34.2 21.9PF2: Moderate activities 12.7 24.4 62.9PF3: Lifting or carrying groceries 9.1 22.2 68.8PF4: Climbing several flights of stairs 15.2 29.6 55.1PF5: Climbing one flight of stairs 5.6 15.8 78.6PF6: Bending, kneeling or stooping 13.0 33.0 54.0PF7: Walking more than a mile 16.5 20.6 62.9PF8: Walking several blocks 10.8 15.0 74.2PF9: Walking one block 4.0 9.9 86.0PF10: Bathing or dressing self 2.0 6.5 91.5Mental Health (N = 9395) All ofthe timeMostof the timeGood bitof the timeSomeof the timeLittleof the timeNoneof the timeHave you…MH1: Been a very nervous person? 1.3 3.1 5.4 17.2 29.7 43.3MH2: Felt so down in the dumps thatnothing could cheer you up?0.3 1.1 1.8 7.4 17.0 72.4MH3: Felt calm and peaceful? 10.6 46.3 18.2 15.7 5.9 3.2MH4: Felt downhearted and blue? 0.4 1.6 3.4 17.8 34.8 42.1MH5: Been a happy person? 18.5 54.8 12.9 8.7 2.9 2.3Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 4 of 13(χ2 = 9.01, df = 27, p = 0.999), suggesting that the 10items represented a single construct.Confirmatory factor analysis for a single latent con-struct resulted in RMSEA = 0.11 and CFI = 0.98,with the former indicating poor model fit. A con-firmatory factor analysis model with error covariancesbetween five pairs of items (PF1 and PF2, PF2 andPF3, PF4 and PF5, PF7 and PF8, PF8 and PF9) re-sulted in a good fit (RMSEA = 0.057, CFI = 0.99).Thus, the factor structure suggested a single latentfactor solution for the PF scale items. Pairwise com-parisons of predicted and observed proportions ofitem responses for all items revealed a poor fit to thedata (χ2 = 3020.96, df = 351, p < 0.001), suggestingthat the assumption of local independence might notbe tenable and heterogeneous subgroups might existin the sample [16, 48].For the MH sub-scale, the IRT two-parameter GRMfit the data well with respect to a unidimensionalmodel. Absolute values of standardized residuals wereless than 1.96 for all sub-scale items. Univariatemodel fit statistics indicated a non-significant differ-ence between observed and predicted item responses(χ2 = 6.61, df = 29, p = 0.999). Confirmatory factoranalysis revealed that one-factor model was not agood fit as judged by the RMSEA = 0.15, althoughthe CFI (0.97) indicated good fit. These results sup-port a single dominant latent factor for both the PFand MH sub-scale items. However, a statistically sig-nificant difference in the joint distributions of itempairs (χ2 = 1952.12, df = 159, p < 0.001) for MHsub-scale was also observed, suggesting that local de-pendence might exist, and thus a mixture IRT modelwith more than one class might be a better choice toaccount for local dependence.Fitting models with latent classesFor the PF sub-scale, the three-class LVMM had a betterfit than the one- and two-class models; the BIC, VLMRand BLRT all favored the three-class model (Table 2).For the MH sub-scale, the two-class model yielded bet-ter fit indices than the one-class GRM. Therefore, three-and two-class models were chosen as the best-fittingmodels for the PF and MH sub-scales, respectively(Table 2).For the PF sub-scale items, threshold values in class 2were uniformly larger than in class 1. As well, there werelarger threshold values amongst respondents in class 3than in class 2. This indicates that respondents with thesame level of latent ability had higher probabilities of en-dorsing lower categories for items representing limitedPF in class 2 and class 3 relative to class 1. Differencesin factor loadings were found for the items across theclasses, suggesting that items varied in their ability todifferentiate amongst respondents.For the MH sub-scale, the absolute value of differencesin item thresholds between the two classes ranged from0.38 for MH1 to 3.56 for MH5. Factor loadings for class2 were smaller than for class 1.Table 3 shows the LR statistics to test for DIF on PFand MH sub-scale items. Overall, nine of the 10 itemsfor the PF sub-scale and all the five items in the MHsub-scale showed uniform DIF.Sources of DIFFigure 1 compares the response percentages for each ofthe three latent classes on the PF sub-scale items. Theprevalence of PF limitations was significantly higher inclasses 2 and 3 compared to class 1. Overall, 98.2% ofrespondents reported having limitations in class 3 on“vigorous activities”compared to only 67.4% for class 1.Table 2 Model fit statistics and class proportions for latent variable mixture models on the SF-36 physical functioning and mentalhealth sub-scale items#Classes#ParBIC VLMRp-valueBLRTp-valueEntropy Class proportionsaClass 1 Class 2 Class 3 Class 4Physical Functioning1 28 102,298.56 1.002 57 100,569.46 <0.0001 <0.0001 0.55 0.61 (0.68) 0.39 (0.32)3 86 100,083.64 0.0007 <0.0001 0.64 0.59 (0.64) 0.24 (0.22) 0.17 (0.14)4 115 99,806.68 NA NA 0.60 0.51(0.55)0.22 (0.19) 0.17 (0.16) 0.10 (0.10)Mental Health1 20 95,536.81 1.002 41 94,631.65 <0.0001 <0.0001 0.40 0.69 (0.78) 0.31 (0.22)3 62 94,363.51 NA NA 0.44 0.25 (0.20) 0.37 (0.40) 0.38 (0.40)Bold values indicate the model that was selected for further analysis. # Par = number of parameters. BIC Bayesian information criterion. VLMR Vuong-Lo-Mendell-Rubin likelihood ratio test. BLRT Bootstrap likelihood ratio test. NA Not ApplicableaClassification of individuals based on their most likely latent class is shown in parenthesesWu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 5 of 13As well, 99.8% of respondents had limitations on “climb-ing several flights of stairs” in class 3 compared to only26.7% in class 1.Similarly, respondents reported more MH problems inclass 2 than class 1 (Fig. 2). For example, 15.2% of re-spondents in class 2 reported having been a very ner-vous person (MH1) all of the time/most of the time/agood bit of the time, compared to just 8.3% of respon-dents in class 1. Close to half (47.4%) of respondents inclass 2 reported feeling calm and peaceful none of thetime/little of the time/some of the time relative to only18.5% of respondents in class 1.Table 4 contains the results for the logistic regression ana-lysis of the associations between latent class membershipTable 3 Likelihood ratio (LR) test statistics for differential item functioning on the SF-36 physical functioning and mental health sub-scale itemsItem LR Statistic Δdf* p-valuePhysical FunctioningPF1: Vigorous activities 99.63 4 <0.0001PF2: Moderate activities 127.49 4 <0.0001PF3: Lifting or carrying groceries 151.04 4 <0.0001PF4: Climbing several flights of stairs 37.69 4 <0.0001PF5: Climbing one flight of stairs 60.95 4 <0.0001PF6: Bending, kneeling or stooping 141.85 4 <0.0001PF7: Walking more than a mile 87.79 4 <0.0001PF8: Walking several blocks 34.51 4 <0.0001PF9: Walking one block 3.26 2 0.196PF10: Bathing or dressing self 64.76 2 <0.0001Mental HealthMH1: Been a very nervous person? 71.79 3 <0.0001MH2: Felt so down in the dumps thatnothing could cheer you up?21.03 3 0.0001MH3: Felt calm and peaceful? 234.43 3 <0.0001MH4: Felt downhearted and blue? 52.27 3 <0.0001MH5: Been a happy person? 198.28 3 <0.0001LR statistics are for the comparison of nested models: (a) mixture IRT model with free thresholds across classes for each item, and (b) IRT model with constrainedthresholds across classes for each item. The mixture IRT model accounted for local dependence. Note that the latent variable mixture model for the physical functioningsub-scale had 3 classes and the model for the mental health sub-scale had 2 classes. Δdf is the difference in degrees of freedom for models defined in (a) and (b)Fig. 1 Within-class item response percentages for the SF-36 physical functioning (PF) sub-scale items. Legend: The selected best-fit latent variablemixture model had three classes. For items PF9 and PF10, response categories of “limited a lot” and “limited a little” were combined due tolow frequenciesWu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 6 of 13and the demographic and health status variables. For thethree classes for the PF sub-scale items, the odds of classmembership were higher for females, older respondents,those with a lower level of general health status, and lowerHUI3 index scores in both of classes 2 and 3, relative toclass 1. Respondents with lower educational attainment(i.e., less than high school), and who were overweight/obese had a greater odds of being in class 3 than in class1. For the MH sub-scale, the logistic model revealed thatrespondents of older age, with less than high school edu-cation, and good health had a greater odds of being inclass 2 than class 1.Impact of DIF on predicted factor scoresFor the PF sub-scale, factor scores for the standard one-class GRM and the three-class LVMM were stronglycorrelated (Pearson r = 0.89). Despite the high correl-ation, there were differences in the tails of the distribu-tions. The standardized difference was ≤ −0.5 for 20.2%of the sample and ≥0.3 for 26.2% of the sample (Fig. 3).For the MH sub-scale, differences in factor scores be-tween the one-class and two-class models for the MHitems were also observed. Specifically, 2.6% of respon-dents had a standardized difference ≥ 0.3 and 3.9% of re-spondents had a difference ≤ −0.3 (Fig. 4).There were differences in the CSEMs between the one-class and multi-class models. For example, relative to theone-class model for the PF sub-scale items, in the three-class model there was greater measurement precision inthe lower range of the theta distribution (i.e., theta <0) forclass 1; greater measurement precision was observed inthe upper range of the theta distribution (theta >0) forclass 3.DiscussionUsing a LVMM, this study identified three groups of indi-viduals with different response patterns on the SF-36 PFsub-scale items and two groups of individuals with differ-ent response patterns on the MH sub-scale items in thepopulation-based CaMos. For both sub-scale analyses, la-tent class 1 primarily contained respondents with fewerhealth limitations or problems. For the PF sub-scale, class2 and 3 respondents were more likely to be female, older,and in poorer health than respondents in class 1. For theMH sub-scale, older respondents with lower educationand good health were more likely to be in class 2. Thesefindings indicate that DIF is present in unobserved groups,and the source of DIF can be at least partly explained bydemographic and health status variables.The results for the PF sub-scale are consistent withthose from a previous study of the Canadian adult popu-lation [10], in which a three-class LVMM was found tofit the data better than one- or two-class models. Weobserved similar results in terms of item response pat-terns, effect of DIF on the latent factor scores, and therelationship between the latent class membership andselected socio-demographic and health variables. Thatprevious study also found that the first latent class con-tained more people with no limitations in PF than theother two classes, and that sex, age and health statuswere associated with latent class membership. This pro-vides support for our conclusion that DIF can affect re-sponses to the SF-36 items and helps to validate theutility of the mixture model for exploring DIF.For the MH sub-scale, we observed larger differences inboth factor loadings and thresholds between the two la-tent classes for the following items: “Have you felt calmFig. 2 Within-class item response percentages for the SF-36 mental health (MH) sub-scale items. Legend: The selected best-fit latent variable mix-ture model had two classes. Response categories of “all of the time”, “most of the time” and “good bit of the time” were combined due to lowfrequencies. MH3 and MH5 were reverse coded in the analysisWu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 7 of 13Table 4 Distribution (%) and odds ratios (ORs) for demographic and health status characteristics by latent classDistribution (%) OR (95% CI)Class 1 Class 2 Class 3 Class 2 Class 3Physical FunctioningClass Frequencies 5933 2081 1323SexMen 35.5 23.4 20.6 Ref RefWomen 64.5 76.6 79.4 1.66 (1.44, 1.91) 1.76 (1.48, 2.10)Age25–64 years 60.5 40.7 31.0 Ref Ref≥ 65 years 39.5 59.3 69.0 1.69 (1.48, 1.92) 1.81 (1.55, 2.12)EducationPostsecondary 53.7 43.7 33.0 Ref RefHigh school graduate 14.9 14.0 13.8 1.01 (0.84, 1.22) 1.11 (0.88, 1.38)Less than high school 31.4 42.4 53.2 1.13 (0.98, 1.30) 1.32 (1.12, 1.56)General HealthExcellent/Very good 68.1 42.0 30.0 Ref RefGood 26.3 43.2 40.7 1.99 (1.73, 2.29) 1.86 (1.58, 2.18)Fair/Poor 5.6 14.8 29.3 2.42 (1.94, 3.01) 3.53 (2.78, 4.47)Weight StatusUnderweight/Normal 37.1 39.3 31.9 Ref RefOverweight/Obese 62.9 60.7 68.1 0.99 (0.87, 1.12) 1.21 (1.04, 1.41)HUI3High (≥0.8) 81.1 58.9 38.3 Ref RefLow (<0.8) 18.9 41.1 61.8 1.92 (1.67, 2.22) 2.62 (2.22, 3.09)Mental HealthClass Frequencies 7337 2058SexMen 30.9 29.7 RefWomen 69.1 70.3 1.03 (0.90, 1.17)Age25–64 years 53.8 44.8 Ref≥ 65 years 46.3 55.2 1.18 (1.05, 1.33)EducationPostsecondary 50.2 42.4 RefHigh school graduate 14.6 14.1 1.09 (0.91, 1.30)Less than high school 35.2 43.4 1.21 (1.06, 1.38)General HealthExcellent/Very good 58.7 49.8 RefGood 31.1 36.5 1.13 (1.00, 1.28)Fair/Poor 10.3 13.8 1.17 (0.95, 1.43)Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 8 of 13and peaceful?”, and “Have you been a happy person?”. Thelow values of the factor loadings for the two items in thesecond latent class suggests that the two items have lowability to discriminate between respondents having higherand lower levels of mental health. In other words, a unitchange in the factor scores in one class is not associatedwith the same change in the factor scores in the otherclass [5]. Possible reasons may be the different wordingformat for these two items when compared to the otherMH items. Previous studies have reported that combina-tions of negatively and positively worded questions maycompromise internal consistency of measures [49, 50].The LVMM offers a number of advantages for DIF detec-tion relative to traditional IRT models that focus on DIF in-vestigations for manifest variables. The standard IRT modelassumes there is one homogeneous group (i.e. one latentclass) of individuals within a population and DIF is exam-ined for subgroups of the population based on manifestcharacteristics. However, true sources of DIF may not belimited to manifest groups (i.e., sex, age) conditional on theirlatent ability. The LVMM is optimal for DIF examinationwhen the origin(s) of differential response patterns is notknown a priori. Prior studies have revealed that the mixtureIRT model is a good alternative for detecting DIF [15, 17].In implementing the LVMM, a critical step is to identifythe best-fitting model with an appropriate number of latentclasses. A potential problem can arise when the model con-verges on a local solution rather than the expected globalmaximum likelihood solution, resulting in biased model pa-rameters [51]. To test the stability of the parameter esti-mates for the multiple-class models, we performed analysesby fitting models with the same number of classes but withdifferent numbers of random starts [29]. The parameter es-timates from the models with different sets of randomstarts were very similar, suggesting a stable and robustresult. Another critical step is to characterize the latentclasses. To do this, we used a pseudo-class methodology[52, 53] that has been shown to produce unbiased param-eter estimates [52, 54]. Inaccurate numbers of latent classesmay result from misspecification of the model [55], or theTable 4 Distribution (%) and odds ratios (ORs) for demographic and health status characteristics by latent class (Continued)Weight StatusUnderweight/Normal 37.4 34.8 RefOverweight/Obese 62.6 65.2 1.07 (0.96, 1.21)HUI3High (≥0.8) 71.5 65.0 RefLow (<0.8) 28.5 35.0 1.01 (0.89, 1.16)95% CI = 95% confidence interval; HUI3 = Health Utilities Index Mark 3; bold values indicate ORs that are statistically significant at α = .05. Class 1 is the referencegroup for both models. For the physical functioning sub-scale, the latent variable mixture model has three classes, while for the mental health sub-scale, the latentvariable mixture model has two classesFig. 3 Predicted factor scores for the SF-36 physical functioning (PF) sub-scale. Legend: Predicted factor scores were for the one-class gradedresponse model (GRM) and the three-class latent variable mixture model (LVMM)Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 9 of 13distribution of the data [56]. However, simulation and em-pirical studies have demonstrated the sensitivity of theLVMM [14–16, 20, 21, 57–59] for DIF assesment. Use ofLVMMs can aid researchers to identify and characterizeDIF effects in self-reported health outcome measures.In the context of testing for DIF using the LVMM, therehas been discussion in the literature about the optimal ap-proach to scale the latent variable(s) in order to enablecomparisons of item parameters. One approach is to allowlatent factor means and variances to vary in each class,with equality constraints for item thresholds for one ormore class-invariant items (also called anchor items) [60].For this method, anchor items must be identified. Anotherapproach is to set the latent factor means to zero andvariances to one across classes; then the item parameters(i.e., factor loadings or item thresholds) are freely esti-mated across classes. We adopted this latter approach(i.e., with unequal item thresholds across classes). We sub-sequently tested for DIF on each item by fitting a setmodels with constrained item thresholds between latentclasses for each item one at a time. The method has beenused in previous studies for DIF detection using LVMMs[20, 61].The study limitations must also be acknowledged. Weanalysed only the SF-36 PF and MH sub-scale items.Other sub-scales may also exhibit DIF, but they containsmall numbers of items and therefore are more difficult toanalyze using LVMMs. The sample was comprised of asmaller proportion of men and more older adults than inthe general Canadian population; the findings may there-fore not be representative of younger people and men. Al-though we observed comparable findings regarding DIFeffects for the PF sub-scale with previous research inCanada [9], further research is needed to determinewhether our findings are applicable to populations fromother countries. Strengths of the present study include theuse of a large national population-based sample, and theability to consider both demographic and health statusvariables to characterize the latent classes. Other healthstatus variables, such as the presence of selected chronicconditions, might be considered, if supported by research.ConclusionsIn conclusion, this study identified latent groups of respon-dents for whom the SF-36 PF and MH items function dif-ferently in a diverse national adult sample. The LVMM wasa useful tool to define sub-groups of individuals with similaritem response patterns and investigate characteristics po-tentially associated with group member. This analysis pro-vides information that can be useful for generatinghypotheses for future studies about DIF; for example, onemight use these results to conduct DIF analyses for ob-served population sub-groups in independent samples.Testing for DIF should be a routine part of comparativeanalyses of population health status. The comparability ofSF-36 sub-scale scores can be significantly compromised byheterogeneity in item responses, which can affect the inter-pretation of results from clinical and epidemiologic studies.Future study is needed to validate the present findings inother samples to inform generalizability of the results. Fu-ture research might also compare the sensitivity of the PFand MH sub-scales to detect DIF effects using both mani-fest variable and LVMM approaches.Fig. 4 Predicted factor scores for the SF-36 mental health (MH) sub-scale. Legend: Predicted factor scores were for the one-class graded responsemodel (GRM) and the two-class latent variable mixture model (LVMM)Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 10 of 13AppendixAcknowledgementsWe thank all those participants in CaMos whose careful responses andattendance made this analysis possible. We thank the CaMos Research Groupfor access to the data to complete this study:• David Goltzman (co-principal investigator, McGill University), Nancy Kreiger(co-principal investigator, Toronto), Alan Tenenhouse (principal investigatoremeritus, Toronto),• CaMos Coordinating Centre, McGill University, Montreal, Quebec: SuzettePoliquin (national coordinator emeritus), Suzanne Godmaire (researchassistant), Silvia Dumont (administrative assistant), Claudie Berger (studystatistician), Lisa Langsetmo (Fellow).• Memorial University, St. John’s, Newfoundland: Carol Joyce (director),Christopher Kovacs (co-director), Emma Sheppard (coordinator).• Dalhousie University, Halifax, Nova Scotia: Susan Kirkland, Stephanie Kaiser(co-directors), Barbara Stanfield (coordinator).• Laval University, Quebec City, Quebec: Jacques P. Brown (director), LouisBessette (co-director), Marc Gendreau (coordinator).• Queen’s University, Kingston, Ontario: Tassos Anastassiades (director),Tanveer Towheed (co-director), Barbara Matthews (coordinator), WilmaHopman (scientist).• University of Toronto, Toronto, Ontario: Bob Josse (director), Sophie Jamal(co-director), Tim Murray (past director), Barbara Gardner-Bray (coordinator),Angela Cheung (scientist).• McMaster University, Hamilton, Ontario: Jonathan D. Adachi (director),Alexandra Papaioannou (co-director), Laura Pickard (coordinator).• University of Saskatchewan, Saskatoon, Saskatchewan: Wojciech P. Olszynski(director), K. Shawn Davison (co-director), Jola Thingvold (coordinator).• University of Calgary, Calgary, Alberta: David A. Hanley (director), Jane Allan(coordinator).• University of British Columbia, Vancouver, British Columbia: Jerilynn C. Prior(director), Millan Patel (co-director), Nera Andjelic (coordinator), Brian Lentle(radiologist).FundingThe research team was funded by the Canadian Institutes of HealthResearch (Funding Reference #122110). LML is supported by aManitoba Research Chair from Research Manitoba. CaMos is currentlysupported by the Canadian Institutes of Health Research, AmgenCanada Inc., Dairy Farmers of Canada, Merck Canada, Eli Lilly Canada,and Novartis Canada.Availability of data and materialsThe data that support the findings of this study are available from theCaMos principal investigators, Dr. David Goltzman and Dr. Nancy Krieger, butrestricions apply to the availability of these data. The CaMos has developedan Ancillary Project and Data Release Policy that governs ancillary projectapproval and access to the data. To obtain a copy of the Ancillary Projectand Data Release Policy and the requirements for proposal submission, or toobtain further information from the CaMos principal investigators about dataaccess, please send an e-mail to info@camos.org.Authors’ contributionsXW, RS, WH, NM, TTS, JL, and LML conceived the study design. WH, JP, AP,RGJ, TT, and KSD contributed to acquisition of the data and interpretation ofstudy results. XW, RS, and LML analyzed the data. XW, RS, and LML draftedthe manuscript. All authors provided feedback on a draft version of themanuscript and read and approved the final manuscript.Competing interestsThe authors declare that they have no competing interests.Consent for publicationNot applicable.Ethics approval and consent to participateApproval for this research was provided by the University of ManitobaHealth Research Ethics Board. All study particiants provided consent toparticipate in the study.Publisher’s NoteSpringer Nature remains neutral with regard to jurisdictional claims inpublished maps and institutional affiliations.Author details1School of Public Health, University of Alberta, Edmonton, AB, Canada.2School of Public Health and Health Management, Weifang MedicalUniversity, Weifang, Shandong Province, China. 3School of Nursing, TrinityWestern University & Centre for Health Evaluation and Outcomes Sciences,Providence Health Care, Langley, BC, Canada. 4Department of Public HealthSciences, Queen’s University, Kingston, ON, Canada. 5McGill University HealthCentre, Montréal, QC, Canada. 6Department of Community Health Sciences &Fig. 5 The latent variable mixture model (LVMM) with class-specific parametersWu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 11 of 13O’Brien Institute for Public Health, University of Calgary, Calgary, AB, Canada.7Department of Mathematics and Statistics, University of Saskatchewan,Saskatoon, SK, Canada. 8Division of Endocrinology, Department of Medicine,University of British Columbia, Vancouver, BC, Canada. 9Department ofMedicine, McMaster University, Hamilton, ON, Canada. 10Department ofMedicine, Faculty of Medicine, University of Toronto, Toronto, ON, Canada.11Department of Medicine, Queen’s University, Kingston, ON, Canada.12Faculty of Graduate Studies, University of Victoria, Victoria, BC, Canada.13Department of Community Health Sciences, University of Manitoba,S113-750 Bannatyne Ave, Winnipeg, MB R3E 0W3, Canada.Received: 11 July 2016 Accepted: 4 May 2017References1. Hopman WM, Towheed T, Anastassiades T, Tenenhouse A, Poliquin S,Berger C, et al. Canadian normative data for the SF-36 health survey.Canadian Multicentre osteoporosis study research group. Can Med Assoc J.2000;163:265–71.2. Reeve BB, Wyrwich KW, Wu AW, Velikova G, Terwee CB, Snyder CF, et al.ISOQOL recommends minimum standards for patient-reported outcomemeasures used in patient-centered outcomes and comparative effectivenessresearch. Qual Life Res. 2013;22:1889–905.3. Mokkink LB, Terwee CB, Patrick DL, Alonso J, Stratford PW, Knol DL, et al.The COSMIN checklist for assessing the methodological quality of studieson measurement properties of health status measurement instruments: aninternational Delphi study. Qual Life Res. 2010;19:539–49.4. Terwee CB, Bot SD, de Boer MR, van der Windt DA, Knol DL, Dekker J, et al.Quality criteria were proposed for measurement properties of health statusquestionnaires. J Clin Epidemiol. 2007;60:34–42.5. Lix LM, Acan Osman B, Adachi JD, Towheed T, Hopman W, Davison KS,Leslie WD. Measurement equivalence of the SF-36 in the CanadianMulticentre osteoporosis study. Health Qual Life Outcomes. 2012;10:29.6. Vandenberg R, Lance C. A review and synthesis of the measurementinvariance literature: suggestions, practices, and recommendations fororganizational research. Organ Res Methods. 2000;3(1):4–70.7. Teresi JA, Fleishman JA. Differential item functioning and health assessment.Qual Life Res. 2007;16(Suppl 1):33–42.8. Lix LM, Wu X, Hopman W, Mayo N, Sajobi TT, Liu J, et al. Differential itemfunctioning in the SF-36 physical functioning and mental health sub-scales:a population-based investigation in the Canadian Multicentre osteoporosisstudy. PLoS One. 2016;11(3):e0151519.9. Perkins AJ, Stump TE, Monahan PO, McHorney CA. Assessment ofdifferential item functioning for demographic comparisons in the MOS SF-36 health survey. Qual Life Res. 2006;15:331–48.10. Sawatzky R, Ratner PA, Kopec JA, Zumbo BD. Latent variable mixturemodels: a promising approach for the validation of patient reportedoutcomes. Qual Life Res. 2012;21:637–50.11. Bjorner JB, Kreiner S, Ware JE, Damsgaard MT, Bech P. Differential itemfunctioning in the Danish translation of the SF-36. J Clin Epidemiol. 1998;51:1189–202.12. Fleishman JA, Lawrence WF. Demographic variation in SF-12 scores: truedifferences or differential item functioning? Med Care. 2003;41(7 Suppl):III75–86.13. Edelen MO, Reeve BB. Applying item response theory (IRT) modeling toquestionnaire development, evaluation, and refinement. Qual Life Res. 2007;16(Suppl 1):5–18.14. Oliveri ME, Ercikan K, Zumbo B. Analysis of sources of latent class differentialitem functioning in international assessments. Int J Test. 2013;13:272–93.15. Maij-de Meij AM, Kelderman H, van der Flier H. Improvement in detectionof differential item functioning using a mixture item response theorymodel. Multivar Behav Res. 2010;45:975–99.16. Cohen AS, Bolt DM. A mixture model analysis of differential itemfunctioning. J Educ Meas. 2005;42:133–48.17. Samuelsen KM. Examining differential item functioning from a latent classperspective. In G. R. Hancock & K. M. Samuelsen, editors. Advances in latentvariable mixture models, Charlotte NC: Information Age Publishing; 2008, p.177-197.18. Lubke G, Neale M. Distinguishing between latent classes and continuousfactors with categorical outcomes: class invariance of parameters of factormixture models. Multivar Behav Res. 2008;43:592–620.19. De Ayala RJ, Kim S, Stapleton LM, Dayton CM. Differential item functioning:a mixture distribution conceptualization. Int J Test. 2002;2:243–76.20. Chen Y, Jiao H. Exploring the utility of background and cognitive variablesin explaining latent differential item functioning: an example of the PISA2009 reading assessment. Educ Assess. 2014;19:77–96.21. Webb ML, Cohen AS, Schwanenflugel PJ. Latent class analysis of differentialitem functioning on the Peabody picture vocabulary test-III. Educ PsycholMeas. 2008;68:335–51.22. Yu YF, Yu AP, Ahn J. Investigating differential item functioning by chronicdiseases in the SF-36 health survey: a latent trait analysis using MIMICmodels. Med Care. 2007;45:851–9.23. Pollard B, Johnston M, Dixon D. Exploring differential item functioning inthe SF-36 by demographic, clinical, psychological and social factors in anosteoarthritis population. BMC Musculoskelet Disord. 2013;14:346.24. Kreiger N, Tenenhouse A, Joseph L, Mackenzie T, Poliquin S, Brown JP, et al.The Canadian Multicentre osteoporosis study (CaMos): background,rationale, methods. Can J Aging. 1999;18:376–87.25. Ware JE Jr, Sherbourne CD. The MOS 36-item short-form health survey (SF-36). I. Conceptual framework and item selection. Med Care. 1992;30:473–83.26. Horsman J, Furlong W, Feeny D, Torrance G. The health utilities index (HUI):concepts, measurement properties and applications. Health Qual LifeOutcomes. 2003;1:54.27. Feeny D, Furlong W, Torrance GW, Goldsmith CH, Zhu Z, DePauw S, et al.Multi-attribute and single-attribute utility functions for the health utilitiesindex mark 3 system. Med Care. 2002;40:113–28.28. Haberman SJ. The analysis of residuals in cross-classified tables. Biometrics.1973;29:205–20.29. Muthén LK, Muthén BO. Statistical analysis with latent variables. MPlus userguide, 7th ed. 2012. Los Angeles, CA: Muthén & Muthén.30. MacCallum RC, Browne MW, Sugawara HM. Power analysis anddetermination of sample size for covariance structure modeling. PsycholMethods. 1996;1:130–49.31. Hu LT, Bentler PM. Cutoff criteria for fit indices in covariance structureanalysis: conventional criteria versus new alternatives. Struct Equ Model.1999;6:1–55.32. Chen W, Thissen D. Local dependence indexes for item pairs using itemresponse theory. J Educ Behav Stat. 1997;22:265–89.33. Samejima, F. (1969). Estimation of latent ability using a response pattern ofgraded scores. Psychometrika. 1969 34:100.34. Muthén B. Latent variables hybrids: overview of old and new models. In G.R. Hancock & K. M. Samuelsen, editors, Advances in latent variable mixturemodels (pp. 1-24). 2008. Charlotte, NC: Information Age Publishing.35. Kamata A, Bauer DJ. A note on the relation between factor analytic anditem response theory models. Struct Equ Model. 2008;15:136–53.36. Dayton C. Latent class scaling analysis. Thousand Oaks, CA: Sage; 1998.37. Nylund KL, Asparouhov T, Muthén BO. Deciding on the number of classesin latent class analysis and growth mixture modeling: a Monte Carlosimulation study. Struct Equ Model. 2007;14:535–69.38. Li F, Cohen AS, Kim S, Cho S. Model selection methods for mixturedichotomous IRT models. Appl Psychol Meas. 2009;33:353–73.39. Vuong QH. Likelihood ratio tests for model selection and non-nestedhypotheses. Econometrica. 1989;57:307–33.40. Muthén B, Brown CH, Masyn K, Jo B, Khoo ST, Yang CC, et al. Generalgrowth mixture modeling for randomized preventive interventions.Biostatistics. 2002;3:459–75.41. Holland PW. A framework and history for score linking. In: Dorans NJ,Pommerich M, Holland PW, editors. Linking and aligning scores and scales.New York: Springer; 2007. p. 5–30.42. Kolen MJ, Brennan RL. Test equating, scaling, and linking. Methods andpractices. 2nd ed. New York: Springer; 2004.43. Cohen AS, Kim S, Wollack JA. An investigation of the likelihood ratiotest for detection of differential item functioning. Appl Psychol Meas.1996;20:15–26.44. Canada H. Canadian guidelines for body weight classification in adults(publication ID 4645). Ottawa: Health Canada Publications Centre; 2003.45. De Ayala RJ. The theory and practice of item response theory. New York:Guilford Press; 2009.46. Joanes DN, Gill CA. Comparing measures of sample skewness and kurtosis.Statistician. 1998;47:183–9.47. SAS Institute Inc. SAS/STAT 9.3 User's guide. Cary, NC: SAS InstituteInc; 2013.Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 12 of 1348. Reise SP, Gomel JN. Modeling qualitative variation within latent traitdimensions: application of mixed-measurement to personality assessment.Multivar Behav Res. 1995;30:341–58.49. Spector PE, Van Katwyk PT, Brannick MT, Chen PY. When two factors don'treflect two constructs: how item characteristics can produce artifactualfactors. J Manag. 1997;23:659–77.50. Sawatzky R, Ratner PA, Johnson JL, Kopec JA, Zumbo BD. Sampleheterogeneity and the measurement structure of the multidimensionalstudents’ life satisfaction scale. SOCI. 2009;94:273–96.51. Rost J. A logistic mixture distribution model for polychotomous itemresponses. Br J Math Stat Psychol. 1991;44:75–92.52. Wang CP, Hendricks BC, Bandeen-Roche K. Residual diagnostics for growthmixture models: examining the impact of a preventive intervention onmultiple trajectories of aggressive behavior. JAMA. 2005;100:1054–76.53. Muthén B, Muthén L. Wald test of mean equality for potential latent classpredictors in mixture modeling. 2007. http://www.statmodel.com/download/MeanTest1.pdf. Accessed 30 Mar, 2016.54. Clark SL, Muthén B. Relating latent class analysis results to variables notincluded in the analysis. 2009. http://www.statmodel.com/download/relatinglca.pdf. Accessed 30 Mar, 2016.55. Bauer DJ, Curran PJ. The integration of continuous and discrete latentvariable models: potential problems and promising opportunities. PsycholMethods. 2004;9:3–29.56. Bauer DJ, Curran PJ. Distributional assumptions of growth mixture models:implications for over extraction of latent trajectory classes. Psychol Methods.2003;8:338–63.57. Lubke G, Muthén BO. Performance of factor mixture models as a functionof model size, covariate effects, and class-specific parameters. Struct EquModel. 2007;14:26–47.58. Finch WH, French BF. Parameter estimation with mixture item responsetheory models: a Monte Carlo comparison of maximum likelihood andBayesian methods. J Mod App Stat Methods. 2012;11:167–78.59. Jiao H, Macready G, Liu J, Cho Y. A mixture Rasch model–basedcomputerized adaptive test for latent class identification. Appl PsycholMeas. 2012;36:469–93.60. Paek I, Cho S-J. A note on parameter estimate comparability: across latentclasses in mixture IRT modeling. Appl Psychol Meas. 2015;39(2):135–43.61. DeMars CE, Lau A. Differential item functioning detection with latentclasses: how accurately can we detect who is responding differentially? EdPsychol Meas. 2011;4:597–616.•  We accept pre-submission inquiries •  Our selector tool helps you to find the most relevant journal•  We provide round the clock customer support •  Convenient online submission•  Thorough peer review•  Inclusion in PubMed and all major indexing services •  Maximum visibility for your researchSubmit your manuscript atwww.biomedcentral.com/submitSubmit your next manuscript to BioMed Central and we will help you at every step:Wu et al. Health and Quality of Life Outcomes  (2017) 15:102 Page 13 of 13


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items