UK Biobank Pilot Study: Stability of haematological and clinical chemistry analytes
Department of Epidemiology and Public Health, Faculty of Medicine, St Mary's Campus, Imperial College London, Norfolk Place, London W2 1PG, UK.
* Corresponding author. Department of Epidemiology and Public Health, Faculty of Medicine, St Mary's Campus, Imperial College London, Norfolk Place, London W2 1PG, UK. E-mail: p.elliott{at}imperial.ac.uk
 |
Abstract |
|---|
 Top Abstract IntroductionMethodsResultsDiscussionReferences |
|---|
Background Analytes in blood and urine may vary over time according to conditions of transport and storage.
Methods UK Biobank pilot study to investigate stability through time of 42 haematological and clinical chemistry analytes in blood and four analytes in urine, kept in storage for up to 36 h, for 40 individuals. Random effects linear regressions were used to model the change through time in repeated assay results on a sample, allowing for heterogeneity between individuals and assay variability.
Results Assay results for most analytes tended to show a small negative bias (1–3% per 12 h stored) over time on average. Statistically significant (P < 0.05) heterogeneity in time trends between individuals, found for nearly all analytes, was dominated by differences in the baseline (time 0) assay results with the possible exception of Mean Corpuscular Haemoglobin Concentration (MCHC). Four out of 46 analytes (serum calcium, cholesterol, fibrinogen and HDL cholesterol) had a predicted probability of a negative time trend for a future individual >0.9. Results for freeze-thaw samples were not materially different from those for non-freeze-thaw samples, except that stability of the analyte results was only assessed up to 12 h.
Conclusions The results suggest that any instability in assay results up to 36 h is likely to be small in comparison with between individual differences and assay error, and that a single assay measurement at any time between 0 and 36 h should give a representative value of the analyte concentration at time zero for that individual.
Keywords UK Biobank, clinical chemistry, haematology, epidemiology, cohort studies
Accepted 10 December 2007
|
Introduction |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
UK Biobank is a long-term longitudinal epidemiological investigation into the effect of lifestyle, genes and environment on health. It aims to collect lifestyle, genetic, biological and health data from 500 000 men and women aged 40–69 at entry into the cohort. As part of the baseline assessment, blood and urine samples are being collected, processed and frozen at –80°C and –192°C for long-term storage for future analysis. We report here results of a pilot study to assess the effects of storage conditions on results of clinical chemistry assays of blood and urine, and haematological parameters in plasma. Specifically, this analysis investigates whether there is a systematic trend for assay results from stored samples to increase or decrease within a period of 12–36 h, or whether any observed trends are merely the result of assay variability.
|
Methods |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Blood serum, plasma and urine samples were obtained from 40 individuals according to a defined protocol.1 From each sample, measurements of 46 analytes were taken; 28 of these were measured at 0 and 24 h, and 18 of these were measured at 0, 12, 24 and 36 h. Additionally, 18 clinical chemistry analytes in blood and urine samples were investigated for the effects of freeze-thaw cycling between 0 and 24 h.
The percentage coefficients of variation (CV) of the assay methods for each analyte, calculated from quality control data supplied by the manufacturers (ICON Development Solutions, Ltd), are presented in Table 1. These show that the assay methods were accurate for each analyte (CV <10%), with the possible exception of basophils (intra-assay CV 33.0%, inter-assay CV 23.0%) and eosinophils (intra-assay CV 17.7%).
|
Random effects linear regression2 was used to model the stability through time of assay results from each individual's sample. Each of the 40 individuals was allowed their own regression intercept and slope, assumed to be randomly sampled from a distribution representing a (hypothetical) population of similar individuals, and between-individual variability in intercepts and slopes was explicitly modelled. The fitted model can then be used to predict intercepts and slopes for future individuals drawn from the same hypothetical population as those in the study.
The specific model fitted for a single analyte was as follows. The assay result yijk for individual i (1–40), time tij at occasion j (corresponding to 0, 12, 24, 36 h), replicate k (1–2) was modelled as a linear function of time, plus a normally-distributed random error. All samples and replicates were assumed to be independent within individuals after adjusting for the time trend. The intercept ai and slope bi of the relationship with time for individual i were modelled as correlated, normally-distributed random effects with unknown mean M and covariance V. The models were fitted using the R package nlme.3
|
|
The estimated distribution of the random regression slopes across individuals can be used to assess whether there is a consistent positive or negative trend through time. If the interval between some extreme lower and upper quantiles (for example, 2.5–97.5%) of this distribution does not contain zero, then this suggests that the variability in slopes is not just due to random error, i.e. it provides evidence that the analyte is unstable.
As the analytes were measured on different scales, changes in time were analysed as percentage differences from the baseline, rather than absolute changes. The expected percentage change in concentration per 12 h time interval was calculated from the ratio of the random slope to the random intercept for each individual. That is, as
|
|
The estimated random slope distribution across individuals was then used to estimate the probability that the slope for a future individual is negative. If this probability is high, then there is likely to be a negative trend in assay results over time for that analyte. If this probability is low, then there is likely to be a reverse (i.e. positive) trend.
The random slope model was also compared (using likelihood ratio tests) against a simpler model with only a random intercept term and no random slope:
|
|
|
Results |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
Table 1 and Figure 1 show the expected percentage change from baseline per 12 h stored (and 95% intervals) for predicted future individuals for each of the 46 analytes: clinical chemistry of serum samples (plasma for glucose and whole blood for fibrinogen), urine samples and haematological parameters. While on average, assayed concentrations for most analytes decreased with time, the expected changes were generally small (<1% per 12 h stored for 23 out of 46 analytes, and between 1% and 3% per 12 h stored for 21 analytes). Only insulin (+3.9% bias per 12 h stored) among the clinical chemistry parameters and eosinophils (–12% bias per 12 h stored) among the haematological parameters had larger average bias, but in both cases, there was considerable variability between individuals, (as indicated by the wide 95% prediction intervals) and none of the biases were statistically significant—in the sense that zero change was supported by the 95% prediction intervals for all analytes.
|
An alternative way to summarize the distribution of the expected changes (Ri) is to calculate the predicted probabilities of a negative bias (Ri < 0) for a future individual. These probabilities are listed for each analyte in Table 1 and show the proportion of the distribution of Ri which is less than zero. For four of the clinical chemistry parameters (calcium, cholesterol, fibrinogen, HDL cholesterol), the probability of negative bias over time was >0.9. However, none of these probabilities was above 0.975 (which corresponds to a two-sided ‘significance level’ of 5%), i.e. the average negative bias over time was not statistically significant at conventional significance levels. None of the probabilities for any of the 46 analytes was <0.1 (a probability of <0.1 would indicate a high chance of a positive bias over time).
Heterogeneity in slopes between individuals
The random slope model was compared against a simpler model with only a random intercept term and no random slope (see Methods). Likelihood ratio tests to compare the fit of the random intercept and random slopes model indicated that, in most cases, the random slope model gave a significantly improved fit to the data (P < 0.05, except for serum creatinine, P = 0.25 and phosphate, P = 0.85). This indicates that there was statistically significant heterogeneity between individuals i in the stability (or slope), bi, of their assay results over time, even though the bi were not consistently positive or negative. However, when compared with other sources of variation in assay results (specifically, assay measurement error between samples analysed at the same time and heterogeneity between individuals in the baseline measurements at time zero) the heterogeneity in slopes was generally small. This is illustrated in Figure 2, which shows the percentage of the total variance of the assay results at each time point that is attributable to between-individual heterogeneity, as opposed to assay measurement error. (The total variance at time tij is calculated as follows: 
. The residual or measurement error variance is Var(
ijk) and the remaining terms in the equation give the between-individual variance at time tij.) At time zero, the between-individual heterogeneity was due entirely to differences in their baseline measurements. These differences explained over 90% of the total variation in assay results for 39 of the 46 analytes, and over 80% of the total variation in another four analytes. For the remaining three analytes, between-individual differences in baseline measurements accounted for 72% (haemoglobin A1C), 75% (Mean Corpuscular Haemoglobin Concentration, MCHC) and 76% (plasma sodium) of the total variation, the remaining variation in all cases being due to assay measurement error.
|
At 0 h, the between-individual heterogeneity also includes any differences between individuals in their slopes or stability. Thus, comparison of values after time zero with those at time zero reveals the extent to which between-individual differences in slopes contribute to the total between-individual variation of assay measurements. This contribution was negligible with the possible exceptions of basophils (+6.7% difference in percentage variance attributed to between-individual heterogeneity at 36 h compared with 0 h), eosinophils (3.1%) and MCHC (+15.7%) among the haematological parameters, and bicarbonate (+3.9%), serum sodium (+4.8%) and total protein (+4.4%) among the clinical chemistry parameters. Of these, only MCHC showed more than a 5% difference in percentage variance attributed to between-individual heterogeneity at 36 h compared with 0 h.
Freeze-thaw cycling
In an additional analysis, the effect of freeze-thaw cycling, performed for 18 clinical chemistry analytes in blood and urine samples, was assessed. The linear regression with random intercepts and slopes was fitted to the data consisting of the 0 h results and the 24 h freeze-thaw results. From this fitted model, 95% intervals for the distribution of percentage change from baseline (Ri) are presented in Table 2. These demonstrate that all changes with freeze-thaw cycling were small and not statistically significant (all 95% intervals include the value of zero change). The estimated probabilities of a negative time trend for each analyte are also given in Table 2. None of these are above 0.95 or below 0.2.
|
|
Discussion |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
The results of the pilot study found, on average, small negative biases over time for most analytes. Set against the biological heterogeneity between individuals in their baseline measurements of each analyte, plus inherent variability in measurements due to assay error, these biases appear to be negligible in both statistical and absolute terms. This suggests that a clinical chemistry or haematology assay carried out up to 36 h after collection of the sample is likely to give a representative measurement from an individual. Freeze-thaw cycling up to 24 h did not materially affect the stability of the assay measurements.
The possible exceptions—at least in statistical terms—were serum calcium, cholesterol, fibrinogen and HDL cholesterol, all with moderately high probabilities (between 0.9 and 0.95) of a future individual having a negative bias in assay results over time; however, in all cases the average size of this bias was small (1–3% per 12 h stored). MCHC had a negligible average bias (0.35% per 12 h stored) but showed evidence of moderate heterogeneity in the size of this bias (i.e. in the slopes over time) between individuals—only around 75% of the variability in baseline (time 0) measurements was due to biological differences between individuals (the rest being assay error), whereas this percentage had increased to 90% by 36 h, suggesting increased variability between individuals’ assay results at 36 h compared with baseline. (Note that the percentage variation attributed to between-individual heterogeneity can go down rather than up over time if the intercepts (baseline measurements) and slopes are negatively correlated). In general, assay error was small relative to true biological differences between individuals in the levels of each analyte—although with samples from only one point in time per individual, there was no information from this study on within-individual biological variability.
A few of the analytes such as insulin, CK MB fraction and eosinophils had wide 95% intervals for the distribution of Ri (percentage change from baseline per 12 h stored) across predicted future individuals—which suggests high between-individual variation in stability. However, very little of the total variation in assay results was explained by this heterogeneity. This is because of the large variation in baseline measurements between individuals (of the order of 10–20-fold for these analytes) which dominates the differences between individuals in changes in assay values over time.
We used random effects linear regression to model the stability through time for assay results within each individual's sample, to properly account for sources of variability.2 A fixed effects model, fitting independent regression lines to each individual's data, only models the specific sample of individuals studied in the experiment, and so is not appropriate for drawing conclusions about stability of samples in the wider population. A much simpler method of assessing stability through time would be to compare average assay values at 12 h with average values at t = 0 h, using paired t-tests (and repeating for 24 and 36 h); however, this ignores information in the duplicate measurements on assay variability at each time. Nor does it investigate between-individual heterogeneity in stability of the assay results. The trend in predicted mean assay values over time can also be assessed by linear regression. This assumes that any differences between individuals in their individual trends is just due to chance variation, and so again, does not allow for true heterogeneity in stability between individuals.
In summary, our analysis suggests that for routine clinical chemistry and haematological parameters, any instability in assay results over time is likely to be small in comparison with between-individual differences and assay error, and that a single assay measurement at any time between 0 and 36 h should give a representative value of the analyte concentration at time zero for that individual.
Conflict of interest: None declared.
|
References |
|---|
|
TopAbstractIntroductionMethodsResultsDiscussionReferences |
|---|
1 Peakman TC, Elliott P. The UK Biobank sample handling and storage validation studies. Int J Epidemiol (2008) 37(Suppl 1):i2–i6.
2 Laird NM, Ware JH. Random effects models for longitudinal data. Biometrics (1982) 38:963–74.[CrossRef][Web of Science][Medline]
3 Pinheiro JC, Bates DM. Mixed-Effects Models in S and S-Plus (2001) New York: Springer.
4 Hinkley DV. On the ratio of two correlated random variables. Biometrika (1969) 56:635–39.
CiteULike 
Connotea 
Del.icio.us What's this?
This article has been cited by other articles:
|
|
 |
|
  K. Furuta, K. Yokozawa, T. Takada, and H. Kato Bio-repository of Post-clinical Test Samples at the National Cancer Center Hospital (NCCH) in Tokyo Jpn. J. Clin. Oncol., August 1, 2009; 39(8): 534 - 539. [Abstract] [Full Text] [PDF]
|
|
|
|
 |
|
 T. A Manolio Biorepositories--at the bleeding edge Int. J. Epidemiol., April 1, 2008; 37(2): 231 - 233. [Full Text] [PDF]
|
|
|
|
|
|
T. C Peakman and P. Elliott The UK Biobank sample handling and storage validation studies Int. J. Epidemiol., April 1, 2008; 37(suppl_1): i2 - i6. [Abstract] [Full Text] [PDF]
|
|
| ||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||||