Limitations in Surrogate Measures of Insulin Resistance

“Insulin resistance” is a loosely defined term that implies a reduction below normal in the ability of insulin to mediate one of its physiological functions. The term is usually applied in the context of glucose regulation, where insulin resistance denotes impairment of insulin-mediated stimulation of glucose removal from the circulation and/or suppression of glucose release into the circulation. Insulin resistance has generated considerable interest in the research community because of its association with such common chronic diseases as obesity, type 2 diabetes, cardiovascular disease, and polycystic ovarian syndrome. Insulin resistance can be quantified directly by two general methods: 1) measurement of glucose utilization and/or glucose production rates during steady-state hyperinsulinemia (glucose clamp technique); or 2) measurement of insulin-mediated changes in glucose disappearance rates after iv glucose challenge (minimal model approach). These methods provide precise quantitative measures of insulin resistance. They are also costly, challenging to perform, and moderately invasive. They are best suited for studies involving tens to hundreds of individuals, although they have been used in studies involving more than a thousand individuals. For larger studies, surrogate measures of insulin resistance have been developed using either fasting blood samples or samples obtained during oral glucose tolerance tests (oGTTs). In general, these surrogates rely on the simple principle that, as insulin resistance increases, insulin levels either rise in compensation or fail to do so, resulting in rising glucose levels. Thus, insulin resistance can be estimated using circulating insulin levels with or without circulating glucose levels. In this issue of JCEM, Lorenzo et al. (1) present a comprehensive evaluation of cross-sectional correlations between 25 of these surrogate measures and either glucose clamp-derived measures of insulin resistance or known cardiovascular risk factors. The measurements were made in a cohort of 272 nondiabetic Finnish adult offspring of individuals with type 2 diabetes. Spearman correlation coefficients between clamp and surrogate measures of insulin resistance were in the range of 0.64–0.77 for surrogates that included at least one plasma insulin level. The best correlations were with surrogates that contained multiple measures of glucose and insulin from oral glucose tolerance tests (r-values of ∼0.75). Only slightly worse were surrogates derived from fasting insulin values (r-values of ∼0.70). Surrogates derived from 2-h insulin values performed worst, but r-values were still approximately 0.65. Correlations between measures of insulin resistance and either metabolic syndrome components or cardiovascular risk were modest (i.e. mostly in the range of 0.15–0.35) and similar in strength for clamp-derived and surrogate measures of insulin resistance. The authors conclude that surrogate measures of insulin resistance are valid and that surrogates derived from fasting measures may be sufficient to adequately measure insulin resistance in clinical and epidemiological studies. What can we learn from this extensive analysis? First, we can learn a main point made by the authors—that in the context of simple correlation coefficients, surrogates based on fasting insulin provide nearly all of the information about insulin resistance that is available from more complex surrogates obtained from oral glucose tolerance tests. Thus, one has to weigh very carefully the small additional correlation that can be obtained against the costs required to perform oGTTs. If oGTTs are being done with measures of insulin for other reasons, then the small additional work of calculating an index like the Matsuda index (2), SIISOGTT (3), Belfiore’s ISI(gly)area (4), or Avignon’s SiM (5) might be worthwhile to get the slightly stronger correlations that these measures provide. On the other hand, conducting oGTTs or measuring insulin from oGTTs specifically for the purpose of estimating insulin resistance may not be worth the cost. The second thing we can learn from the Lorenzo et al. paper (1) is that clamp-derived and surrogate measures of insulin resistance were similarly correlated with metabolic and cardiovascular risks. This observation raises the possibility that insulin resistance is not really the important biological variable that accounts for increased risks of things like the metabolic syndrome and its components, low-grade inflammation, or cardiovascular events. Indeed, most surrogates that contain at least one value for circulating insulin performed as well as or better than a clamp-derived measure of insulin resistance in this context. We don’t really know what accounts for these associations, but it looks like it is not just insulin resistance. Beyond the analysis conducted by Lorenzo et al. (1), we need to ask whether simple correlation analysis is a sufficient basis on which to conclude broadly that surrogate measures of insulin resistance are valid and, therefore, applicable to clinical studies. We believe that the answer to this question is “no” for two reasons. First, correlation coefficients are measures of association that tell us whether two variables vary together across a population or over time, but they do not tell us whether two variables are quantitatively equivalent. Therefore, demonstration of good correlation does not permit the drawing of quantitative relationships between variables. For example, a study employing a direct measure of insulin resistance might reveal that a doubling of resistance results in a doubling of the risk of diabetes. The quantitative relationship between diabetes risk and a surrogate measure of insulin resistance could be very different because of the different biological basis for and numerical scale of the surrogate measure. A test of concordance would provide a more valid assessment of quantitative relationships between two measures of insulin resistance. A detailed discussion of such measures of agreement can be found in the literature (6, 7). Unfortunately, no assessment of quantitative agreement was provided in the paper by Lorenzo et al. (1). The second reason that simple correlation analysis does not allow sweeping conclusions about the validity of surrogate measures of insulin resistance is that a strong correlation reflects the contribution of all sources of variation. In the case of cross-sectional clinical or epidemiological studies, this may be the variation of interest, and the results of Lorenzo et al. (1) support the use of surrogate measures in this context. However, strong Spearman or Pearson correlation coefficients don’t necessarily indicate validity when more specific forms of variation are important to the research question. Examples can be found in two settings—genetic association and longitudinal studies. Bergman et al. (8) formally assessed the phenotypic, environmental, and genetic correlations among fasting insulin, homeostasis model of assessment for insulin resistance (HOMA-IR), and minimal model-based insulin sensitivity (SI). They found very strong phenotypic, environmental, and genetic correlations between fasting insulin and HOMA-IR. However, genetic correlations between these two surrogates and SI were much lower—approximately half that observed between fasting insulin and HOMA-IR. The authors concluded that HOMA-IR does not measure the genetic contribution to insulin resistance reflected in SI. The findings of Bergman et al. (8) are supported by a comprehensive study by Ingelsson et al. (9) in which genetic variants known to be associated with fasting glucose were tested for association with other metabolic measures, including a wide range of surrogate measures of insulin resistance. The Ingelsson study (9) showed that tests of association between genetic variants and different surrogate measures of insulin resistance were not consistent. For example, variation in glucokinase was associated with the Gutt index (P = 8.1 × 10−5) but was not associated with the Stumvoll index (P = 0.22), the Matsuda index (P = 0.59), or the Belfiore index (P = 0.54). Clearly, surrogate measures of insulin resistance cannot simply be applied to genetic studies, with the expectation that they will reflect real genetic variation in insulin resistance. Longitudinal studies pose another challenge—quantifying changes in insulin resistance over time. Surrogates don’t perform very well in this context because individual components of the surrogates may reflect changes in physiological factors not related to insulin resistance per se. Hücking et al. (10) first identified this limitation using computer simulation. We demonstrated it using real data from a cohort of 231 Mexican-American adults with fasting glucose less than 7.0 mmol/liter (our unpublished data). Subjects had oral and iv glucose tolerance tests twice, 3.5–5 yr apart. Spearmen correlation coefficients between minimal model SI and either HOMA-IR or the Matsuda index were 0.65–0.70 at baseline and 0.69–0.72 at follow-up. Correlation coefficients for the change in insulin resistance were much lower, between 0.3 and 0.4. Thus, it is not appropriate to apply a surrogate measure of insulin resistance to longitudinal data simply because the surrogate is well correlated with direct measures of insulin resistance in cross-sectional settings. So, what is the take-home message from these considerations? The analysis presented by Lorenzo et al. (1) indicates that surrogate measures of insulin resistance containing a fasting insulin value provide potentially useful information about interindividual variation in insulin resistance at one point in time. A small amount of additional information can be obtained by adding multiple measures of insulin from oGTTs. Thus, simple surrogate measures of insulin resistance may be appropriate for large-scale clinical and epidemiological studies with cross-sectional design. The loss of accuracy compared with the clamp and minimal model approaches can be counterbalanced by large sample sizes, convenience, and cost savings. However, good correlations in cross-sectional settings don’t justify the use of surrogate measures of insulin resistance in other types of studies. Surrogates have already been shown to perform poorly in genetic association studies and in longitudinal metabolic studies. They need to be validated before they can be applied in any setting where cross-sectional variability in insulin resistance from all sources is not the primary characteristic of interest. This work was supported in part by Grants R01DK061628, M01RR00043, and 1UL1RR031986 from the National Institutes of Health. Disclosure Summary: The authors have no relevant conflicts to disclose. For article see page 5082 Homeostasis model of assessment for insulin resistance; oral glucose tolerance test; insulin sensitivity.