Inaccurate Disparities Estimates: Systematic Measurement Error’s Biasing Influence on Disparities in National Rates of Children with Special Health Care Needs across Spanish and English Speaking Households Adam C. Carle, M.A., Ph.D. adam.carle@cchmc.org Division of Health Policy and Clinical Effectiveness Cincinnati Children’s Hospital Medical Center University of Cincinnati School of Medicine AcademyHealth June 26, 2010: Boston, MA Introduction • To validly compare treatments and policies, the field needs sound measurement. – Observation and evaluation depend on self-report. – Sound measurement = reliable and valid. • To accurately examine outcomes across subgroups, field requires measures demonstrating similar reliability and validity across subgroups. Introduction • Little research has empirically evaluated child health measures. • (Drotar, 1998; Stein, 2004). • Ensues despite fundamental importance of reliable and valid measurement in research. • Consider a measure to identify children with special health care needs. Introduction • Children with special health care needs (CSHCN): • Children with or at “increased risk for a chronic physical, developmental, behavioral, or emotional condition and who also require health and related services of a type beyond that required for children generally.” • (McPherson et al., 1998). Introduction • CSHCN account for 12.8% of all US children. • (van Dyck, et al., 2004). • Yet, CSHCN’s medical expenses account for over 80% of all children’s medical expenses. • (Institute of Medicine, 1998). Introduction • Researchers have observed substantial differences in the prevalence of CSHCN across: – Race and ethnicity. • White, Black, Hispanic. – Main difference across NH and Hispanic children. – Primary household language. • English vs. Other. • Largest difference across Spanish speaking Hispanics and all others. • 4.6% vs. 14%! – (Read, et al., 2007). Introduction • These rates specifically inform allocation of resources to different subpopulations of children. • Rates also guide how we understand the effectiveness of health policies, services, and treatments for different subgroups of children. • Given a limited set of resources, we should: – Direct resources to the groups most in need. – Direct resources to the most successful policies. – Direct resources to the most successful treatments. Introduction • Before making comparisons across subgroups, we must consider cross-group measurement equivalence. – If responses affected by translation problems, impossible to compare groups on observed scores. • We must consider whether observed differences reflect true differences. • Do we allocate resources and deliver treatments based on systematic measurement error? Measurement Bias Measurement Bias • A possibility… • Individuals with identical health statuses who receive different translations may respond differently to questions about their health. – Should respond similarly, but don’t. • Systematic measurement error. – Measurement bias. – Differential item functioning (DIF). Why Study Bias? • Generally decreases reliability and validity. • (Knight & Hill, 1998). • Attenuate or accentuate group differences. • (Carle, 2008). • Lead to inaccurate diagnoses and classifications. • (Carle, 2009). • Can render cross group comparisons impossible. • (Prelow, et al., 2002). Bias and Translation • No studies empirically examine equivalence of CSHCN measures across translations. – Does not imply failure to follow stringent translation process methods. – Despite best methods, differences still occur. • Leaves unclear to what extent bias influences disparities research. Evaluating Bias • Latent variable models potently investigate bias. • (Millsap & Kwok, 2004, Muthén, 1989). • Equations describe the relations among questions. • Examine the cross-group equivalence of the measurement parameters in the equations. • (Millsap & Yun-Tien, 2004). • Differences in these parameters across groups reflect bias. Evaluating Bias • Multiple group (MG) confirmatory factor analyses for ordered-categorical measures (CFA-OCM). • A popular latent variable method. Evaluating Bias: MG-CFA-OCM • Let X ij equal the ith individual’s score on the jth ordered-categorical item. – Let the number of items be p (j = 1, 2, .., p). – Scores, m, range {0, 1, …, s}. • We assume a continuous latent response variate, X ij , determines observed responses. Evaluating Bias: MG-CFA-OCM X • A threshold value on ij determines responses: – If X ij less than the threshold, respond in one category. – If X ij greater than threshold, respond in at least next highest category. X ij m if jm X j m1 ij • j 0 , j1 ,..., j s 1) represent threshold parameters. Evaluating Bias: MG-CFA-OCM • Suppose some factor or set of factors, , is responsible for the observed scores. • X ij relates to the factor(s) as follows: X j ji ij * ij Evaluating Bias: MG-CFA-OCM • Subscript the parameters to allow for group differences. • To the extent that cross-group constraints lead to problematic fit, measurement bias presents. ? Evaluating Bias: MG-CFA-OCM Evaluating Bias: MG-CFA-OCM • Essentially, we ask if we “see” the same picture for each group. • If not, bias exists. – Not good! • If yes, we have equivalent measurement. – A good thing! Evaluating Bias • Methodological and substantive issues can limit MG-CFA-OCM. • Difficult to incorporate multiple grouping variables simultaneously. • Why does this matter? Observational Research • Survey respondents not randomly assigned. • Differences in translation may result from translation problems, or…. • Differences may result from other variables that covary with language. Observational Research • Individuals who speak only Spanish tend to: – Come from poorer homes. – Have lower educational achievement. • (Yu, et al., 2002). • Both of these predict CSHCN status. • Apparent translation differences may reflect systematic differences in the distribution of SES, not bias. Observational Research • Difficult to simultaneously include multiple variables in “traditional” latent variable approaches. • Failure to include available information in model estimation may lead to erroneous conclusions. • What do we do? MG-MIMIC Models • Multiple group (MG) multiple indicator, multiple cause (MIMIC) models. – Build on developments in structural equation modeling, IRT, and CFA-OCM. • (Jones, 2003; Jones, 2006; Muthén, 1989) • Control for “extra” (background) variables by incorporating them as covariates. • More fully address heterogeneity within and across groups. MG-MIMIC Models Review • Children in Spanish speaking households are much less likely to be identified as CSHCN. – 4.6 vs. 14%. • Measurement bias may partially or fully account for this difference. • No studies have examined this possibility. Why should we care? • Beyond “just” achieving accurate estimates, why should we care? • The Maternal and Child Health Bureau allocates approximately one BILLION dollars each YEAR to the child health population. • States must spend 30% of this money on CSHCN. • Distribution of funds depends in part on distribution of CSHCN. Why should we care? • If CSHCN measures systematically fail to identify minority CSHCN…. • Minority CSHCN fail to receive services as a function of measurement bias. • Measurement bias may influence hundreds of millions of dollars in federal spending. • Measurement bias may cause minority children to suffer needlessly. Current Study • Utilized MG-MIMIC. – Probed for bias across English to Spanish translation of the CSHCN Screener in the 2005-2006 National Survey (NS) of CSHCN. • Principal source of information on US CSHCN. • MG-MIMIC models simultaneously included the family’s education and income level in analyses. • Compared MG-MIMIC to MG-CFA-OCM. – MG-CFA-OCM “only” examined language. Methods • Participants (n = 350,345) came from the National Survey of CSHCN (NS-CSHCN). – English n = 326,050. – Spanish n = 24,295. – Sponsored by Maternal and Child Health Bureau. – Fielded by National Center for Health Statistics (NCHS). • (Blumberg, et al., 2008). • Provides state and national data on CSHCN. – Represents noninstitutionalized children. – All analyses accounted for complex design. Results: MG-CFA-OCM • Similar general model fit across language versions. – – – – – RMSEA = 0.025 CFI = 0.99 TLI = 0.99 McDonald’s NCI > 0.99 Gamma Hat = 0.998 • Items should measure a similar construct and have comparable meaning regardless of translation. Results : MG-CFA-OCM • Model constraining the loadings demonstrated measurement bias. – – – – – – RMSEA = 0.019 CFI > 0.99 TLI > 0.99 McDonald’s NCI > 0.99 Gamma Hat = 0.999 Δ2 = 11.727 (3, n = 350,345, p < 0.01) Results : MG-CFA-OCM • Items would not relate to underlying construct similarly across groups. • Estimates wouldn’t deliver similar reliability across language. Results : MG-CFA-OCM • Model constraining the thresholds demonstrated statistically significant measurement bias. – – – – – – RMSEA = 0.018 CFI > 0.99 TLI > 0.99 McDonald’s NCI > 0.99 Gamma Hat = 0.999 Δ2 = 76.434 (3, n = 350,345, p < 0.01) • Reponses would not correspond to similar health levels. • Translation would not provide equivalent measurement. Discussion • MG-MIMIC showed that measurement bias did not present across translation. – After accounting for systematic differences in SES across groups, I found NO bias. – Revealed that the CSHCN Screener did provide equivalent measurement across translations. – Indicates we can trust estimates across translations. • When comparing within similar sociodemographic groups. Discussion • MG-CFA-OCM indicated measurement bias across translations. – Suggested that the CSHCN Screener did not provide equivalent measurement across translations. – This would mean that we could not trust estimates across English and Spanish versions of the Screener. Discussion • Traditional methods unable to model the systematic influence of multiple variables. • Led to erroneous conclusion. – Suggested that bias presented across translation. • This would incorrectly indicate that differences in prevalences of CSHCN across reflect systematic measurement error due to language. – When disregarding sociodemographic groupings. Discussion • MG-MIMIC provided more accurate picture. – Modeled systematic influence of multiple variables. • Differences in prevalences of CSHCN across languages likely do not reflect measurement bias. – When considering sociodemographic differences. • Likely reflect true differences. Limitations • Self report data. – Responses may not reflect children’s actual health. • Only address these five consequences. – Consequences not included on the Screener may have more relevance among Spanish speaking Hispanics. Limitations • MG-MIMIC can only detect biased thresholds for background variables. – Leaves loadings unaddressed. – Problem for background variables only. – Missing data approach possibilities. • If bias permeates the entire Screener, analyses cannot detect this. – No statistical approach can. • (Millsap, 2006). Summary • Findings strongly support CSHCN Screener responses across Spanish and English versions. – Only when examining within similar sociodemographic groups. • Increases confidence in research showing significant differences in CSHCN prevalences across English and Spanish speaking children. – When similar sociodemographic groups. Summary • MG-MIMIC and MG-CFA-OCM differences: – Highlight importance of simultaneously considering multiple influences on measurement. – Some cross-cultural measurement differences result from systematic education and income differences. – Using model-based estimates can mitigate systematic measurement error and lead to more accurate disparities estimates and research. Conclusion • When conducting research across distinct racial, ethnic, and/or cultural groups, investigators should establish measurement equivalence to insure validity of results before developing policy and allocating resources based on the results. • In this way, we can better advance positive outcomes in a diverse world. Inaccurate Disparities Estimates: Systematic Measurement Error’s Biasing Influence on Disparities in National Rates of Children with Special Health Care Needs across Spanish and English Speaking Households Adam C. Carle, M.A., Ph.D. adam.carle@cchmc.org Division of Health Policy and Clinical Effectiveness Cincinnati Children’s Hospital Medical Center University of Cincinnati School of Medicine AcademyHealth June 26, 2010: Boston, MA