Monitoring Cholesterol Levels: Measurement Error or True Change?
Published May 6, 2008 · P. Glasziou, L. Irwig, Stephanie R Heritier
Annals of Internal Medicine
125
Citations
2
Influential Citations
Abstract
Context What is the optimal monitoring interval for patients taking cholesterol-lowering medication? Contribution This analysis of data from a trial that compared pravastatin with placebo in patients with coronary disease found that the signalnoise ratio in cholesterol monitoring was weak. Short-term variability of measurement was about 0.80 to 0.80 mmol/L (31 to 31 mg/dL). Calculations suggested that frequent follow-up of patients with values 0.5 mmol/L (19 mg/dL) or more under target detected many more false-positive results than truly elevated cholesterol values. Implication Consider testing adherent patients with well-controlled cholesterol levels every 3 to 5 years rather than every few months or annually. The Editors Cholesterol level monitoring is a common clinical activity. Because indications for treatment have been widening over the past decade, cholesterol-lowering medications have become some of the most widely used and expensive pharmaceutical items, and cholesterol screening, treatment, and monitoring have increased. For example, lipid panels were the third highest contributors to Medicare testing growth between 2000 and 2004, with a 61% increase in volume and a 65% increase in cost (1). Previous studies have suggested that, because of measurement error, frequent monitoring is just as likely to mislead when trying to decide whether changes in treatment are needed (2). Most lipid management guidelines clearly state the number and interpretation of initial measurements but are less specific about subsequent monitoring. The National Cholesterol Education Program in the United States suggests that patients can be monitored for response to therapy every 4 to 6 months, or more often if considered necessary, (3) whereas the Medicare guideline (4) states The LDL [low-density lipoprotein] cholesterol or total cholesterol may be measured three times yearly after treatment goals have been achieved. In the United Kingdom, the PRODIGY (Prescribing Rationally with Decision Support in General Practice Study) guidelines (5) suggest rechecking annually. The Australian National Heart Foundation and the Cardiac Society of Australia and New Zealand guidelines suggest lipid profile measurement every 6 to 12 months (6). However, the basis for recommending these intervals is unclear, and none of the guidelines explicitly describes within-person variability or the likely rates of change in cholesterol levels over time with fixed-dose therapy. We therefore studied the implications of different strategies for monitoring cholesterol level. Our objectives were to estimate, in patients receiving a fixed dose of cholesterol-lowering medication or placebo, 1) the extent to which the initial response to treatment varies among patients; 2) the extent to which the initial response is sustained and long-term change varies within and among patients; and 3) the detectability of these long-term changes in on-treatment cholesterol level (signal), given short-term, within-person variation (noise). Methods We used data from the LIPID (Long-Term Intervention with Pravastatin in Ischaemic Disease) trial (June 1990 to May 1997). The LIPID trial was a randomized trial of 9014 patients with acute coronary syndromes diagnosed 3 to 36 months previously who had been randomly assigned to 40 mg of pravastatin or matching placebo and had been followed for an average of 6.0 years (7). Before randomization, patients entered an 8-week placebo run-in phase. For patients to qualify for the study, their plasma total cholesterol levels 4 weeks before randomization had to be between 4.0 and 7.0 mmol/L (155 and 271 mg/dL), and the fasting triglyceride level had to be less than 5 mmol/L (<443 mg/dL). Lipid concentrations (including concentrations of LDL cholesterol, high-density lipoprotein cholesterol, and triglycerides) were measured at randomization, 6 and 12 months later, and then every 12 months for 5 years. We recorded information on adherence to treatment and uptake of other cholesterol-lowering medications. A single laboratory measured all cholesterol concentrations, eliminating differences among laboratories as a source of variation. Estimation for each of the 3 objectives required different methods. Variation in Initial Response to Treatment Patients receiving placebo will show some variation in apparent response that will be attributable to short-term variation. A greater variation of change in total cholesterol level in the pravastatin group indicates some variation in the true response. Therefore, we used the difference in the variance of change from baseline to 6 months between the pravastatin and placebo groups to estimate the variation in true response to treatment with pravastatin. Variation in Long-Term Change within and among Patients After initial response to therapy, the apparent changes in cholesterol level measurements over time comprise 3 components: 1) the average, true, long-term change in cholesterol level of the whole group, which we estimated from the group average at each time point; 2) short-term variability, which is a combination of analytic variability and week-to-week biological fluctuation around a stable average; and 3) long-term variability, which is a variation in true, long-term change among individuals (as would be seen with the theoretical average of a large number of measurements per individual). To estimate short-term variability, we used 2 methods. First, we used the cholesterol concentrations during the run-in period (excluding the first measurement), measured only 4 weeks apart, to provide a direct estimate of short-term (4-week) variability. Second, we used a linear extrapolation backward from the longer-term measurements (variogram method) (8), to estimate what the apparent variance would be at time 0. We estimated the long-term variability with a linear mixed-effects model and a direct method. Details and equations are provided in the Appendix. The next sections outline the statistical methods, assumptions, and problems. Modeled Method of Estimation We estimated the components of variance by using a mixed longitudinal model, which assumed that each patient had a linear increase over time but that the rate of increase varied between patients. Specifically, if each patient i has a rate of increase i over time and these rates follow a normal distribution N(, 1 2), the model was: with Ti again being the true cholesterol measurement at baseline, and it being the error terms for t= 1, 2, n i, which are independent from each other even for the same patient. The Appendix gives more details and equations. Direct Method of Estimation The direct method to estimate long-term variability uses the variance of the differences between the baseline value and each subsequent time point, calculated as: (cholesterol level at time icholesterol level at baseline) in which the times i are 6 months to 5 years after the stable baseline (which we have taken as 6 months after treatment for the pravastatin group). By subtracting the short-term variability (described in the previous section) from this variability of the change, we estimated the additional long-term variability. In general, the direct method and the modeled method gave similar estimates, but the latter estimate seemed to increase too slowly at first and too rapidly later. Alternative functional relationships may be needed. Censored Values Few patients were lost to follow-up. However, to estimate the change for those who were receiving stable treatment, a key issue was patients who withdrew or started taking a nonstudy cholesterol-lowering treatment. To estimate the change, we used 3 methods. First, when patients discontinued or started taking alternative cholesterol-lowering medication, we censored the data and replaced values thereafter with the last value carried forward for each subsequent measurement. Second, we excluded patients who stopped or began taking the study medication. These 2 methods have small (and opposing) biases, so we examined any discrepancy between the methods. Third, we performed an extrapolation based on a weighted sum of the group trend and the individual patient's own trend. Detectability of Long-Term Changes (Signal) Given Within-Person Variation (Noise) Finally, we aimed to estimate true- and false-positive rates: the number of patients whose true cholesterol level would or would not exceed an acceptable threshold. After a treated patient's cholesterol level has stabilized, 2 elements may lead to a true increase in cholesterol level: the average change of the whole group over time and the real variation around the average change. To estimate these, we used the average change (0.03 mmol/L [1 mg/dL] per year) and the true within-person variability (estimated as described previously), andfor time points of 1, 3, and 5 yearsa normal distribution to estimate the proportion whose true value would or would not have changed beyond that acceptable threshold. For those below or above the target value, we calculated the error rate on the basis of the short-term variability. Role of the Funding Source This work was supported in part by a grant from the Australian National Health and Medical Research Council and a UK National Institute for Health Research program grant. Neither agency had any role in the design, conduct, or interpretation of the study or the decision to submit the manuscript for publication. Results At baseline, 9014 patients with a median age of 62 years (83% male) and an average pretreatment cholesterol level of 5.65 mmol/L (SD, 0.82; range, 3.0 to 9.2 mmol/L) (218 mg/dL [SD, 32; range, 116 to 355 mg/dL]) were randomly assigned to pravastatin or placebo. Table 1 shows the cholesterol levels over time. Only 1 patient was lost to follow-up, but by year 5, about 5% and 6% of patients had died in the pravastatin and placebo groups, respectively. Table 1. Average Total Cholesterol Concentrations in the Placebo and Pravastatin Groups in the L