- Research
- Open Access
Reliability of environmental sampling culture results using the negative binomial intraclass correlation coefficient
- Sharif S Aly^{1, 2}Email author,
- Jianyang Zhao^{3},
- Ben Li^{3} and
- Jiming Jiang^{3}
Received: 19 November 2013
Accepted: 17 January 2014
Published: 22 January 2014
Abstract
The Intraclass Correlation Coefficient (ICC) is commonly used to estimate the similarity between quantitative measures obtained from different sources. Overdispersed data is traditionally transformed so that linear mixed model (LMM) based ICC can be estimated. A common transformation used is the natural logarithm. The reliability of environmental sampling of fecal slurry on freestall pens has been estimated for Mycobacterium avium subsp. paratuberculosis using the natural logarithm transformed culture results. Recently, the negative binomial ICC was defined based on a generalized linear mixed model for negative binomial distributed data. The current study reports on the negative binomial ICC estimate which includes fixed effects using culture results of environmental samples. Simulations using a wide variety of inputs and negative binomial distribution parameters (r; p) showed better performance of the new negative binomial ICC compared to the ICC based on LMM even when negative binomial data was logarithm, and square root transformed. A second comparison that targeted a wider range of ICC values showed that the mean of estimated ICC closely approximated the true ICC.
Keywords
- Intraclass correlation coefficient
- Generalized linear mixed model
- Negative binomial mixed model
- Variance components
Introduction
In the simple case of estimating the correlation among 2 factors with a set of quantitative observations, an investigator may elect to utilize the Spearman Rank correlation coefficient or Pearson’s correlation coefficient assuming the observations are independent. The measure of agreement κ can be estimated for correlation between binary observations. For more complex data structures that may include either crossed or nested factors of a latent character, the investigator may utilize the Intraclass Correlation Coefficient (ICC). The ICC is related to unexplained variance at the subject level. More specifically, the ICC is defined as the ratio of the covariance of measurements from the factor of interest to the marginal variance of the observations. Ranging between 0 and 1, an ICC close to 1 indicates that the difference in observations due to the factor of interest are ignorable. Hence, using variance estimates attributable to each of a study’s factors, the ICC can be used as a measure of similarity in observations between subjects due to a particular factor. A direct application of the ICC is a measure of the correlation between subjects in a reliability and repeatability gauge study (Aly et al. 2009; Kittawornrat et al. 2012).
Investigators analyze and obtain variance estimates for normally distributed data using linear mixed models (LMM) or non-normally distributed data using generalized linear mixed models (GLMM). Health science researchers more commonly work with count data and while the ICC for the LMM has been extended to the Poisson case (Carrasco and Jover 2005), its equivalence for count data with overdispersion was only recently described (Carrasco 2010). Until the ICC for negative binomial distributed data was developed, researchers transformed such data using different transformations to make their data normally distributed in order to use LMM and their ICC.
An example of count data that may commonly be overdispersed is bacterial culture results. Culture results are commonly reported as colony forming units per specimen mass or culture medium tube. Another example is parasite counts which are commonly reported as parasitic stage count per gram of specimen. Given the nature of such infectious agents, they can exist in very large numbers within their hosts, at the same time not all potential hosts in a population are infected. In fact, more hosts tend to be uninfected leading to the inequality of the mean and variance of the data, hence overdispersion. In the current study, we report on a reliability analysis for environmental sampling to quantify Mycobacterium avium subspecies paratuberculosis (MAP) on California free-stall dairies (Aly et al. 2009). A previous study with these data was unique in that it involved the use of nested and crossed factors and used the natural logarithm to attain normally distributed data for a LMM analysis and ICC estimation. Such transformations may normalize the data provided the number of replicates was large and the variance components were small (Solomon and Taylor 1999). Both sample size and magnitude of variance conditions may be difficult to attain with negative binomial distributed data especially when replicates are limited due to cost or subject use limitations such as in health sciences research. The performance of the negative binomial ICC has not been compared to LMM ICC using previously described data transformations in multilevel models with crossed and nested random effects.
Hence, the objectives of this study were to specify a negative binomial mixed model and estimate and contrast the performance of the resulting ICC to that based on estimates from linear mixed models of several data transformations. In addition to the reliability study on environmental sampling to quantify MAP in dairy pens, a wide variety of negative binomial distributed data was simulated to contrast estimator performance.
Methods
ICC for the negative binomial mixed model
The random variable $2\left(\beta +{a}_{i}+{b}_{\mathit{ij}}+{c}_{k}\right)+{d}_{{l}_{1}}+{d}_{{l}_{2}}$ has the distribution N $\left(2\beta ,4{\sigma}_{a}^{2}+4{\sigma}_{b}^{2}+4{\sigma}_{c}^{2}+2{\sigma}_{d}^{2}\right)$, hence ${e}^{2\left(\beta +{a}_{i}+{b}_{\mathit{ij}}+{c}_{k}\right)+{d}_{{l}_{1}}+{d}_{{l}_{2}}}$ has the distribution log-normal $\left(2\beta ,4{\sigma}_{a}^{2}+4{\sigma}_{b}^{2}+4{\sigma}_{c}^{2}+2{\sigma}_{d}^{2}\right)$.
When the variance of data is much larger than its expectation, the negative binomial distribution is often used as an alternative to the Poisson distribution. The random effects follow the normal distribution and the link function is the logarithm. Based on this formula, the ICC is no longer just based on the random effects, but also the fixed effect intercept and the number of successes. Thus, the negative binomial mixed model may be more reasonable than the LMM or the Poisson GLMM when count data are overdispersed.
Simulations
- 1.
Randomly generate normal random effects a _{ i }, b _{ ij }, c _{ k }, d _{ l }(i = 1, 2,.. n _{ m }; k = 1, 2.. s; l = 1, 2,.. t) with respective scenario’s variances
- 2.
Sum the intercept and random effects as conditional expectation ${\mu}_{\mathit{ijkl}}={e}^{\beta +{a}_{i}+{b}_{\mathit{ij}}+{c}_{k}+{d}_{l}}$, β is estimated intercept from field data
- 3.
Randomly generate negative binomial variable Y _{ ijkl } ~ NB(r, μ _{ ijkl }) r is number of successes
- 4.
Estimate model parameters: intercept β, number of successes r and random effects ${}^{{\sigma}_{a}^{2}+{\sigma}_{b}^{2}+{\sigma}_{c}^{2}+{\sigma}_{d}^{2}}$
- 5.
Calculate the ICC
Parameters of a simulation to compare the true and estimated negative binomial Intraclass Correlation Coefficient (ICC) using an example of culture results for a specific bacterium in pen floor samples (variance 0.5) collected over several days apart and simultaneously by different veterinarians and across different dairies
Scenario | r | β | Variance | E(Y) | True ICC | |||
---|---|---|---|---|---|---|---|---|
Dairy | Pen | Day | Veterinarian | |||||
1 | 1 | 0 | 0.5 | 0.5 | 0.2 | 0.1 | 1.92 | 0.3382 |
2 | 1 | 0 | 1 | 0.5 | 0.2 | 0.1 | 2.46 | 0.3888 |
3 | 1 | 2 | 0.5 | 0.5 | 0.2 | 0.1 | 14.15 | 0.362 |
4 | 1 | 2 | 1 | 0.5 | 0.2 | 0.1 | 18.17 | 0.4011 |
5 | 2 | 0 | 0.5 | 0.5 | 0.2 | 0.1 | 1.92 | 0.4616 |
6 | 2 | 0 | 1 | 0.5 | 0.2 | 0.1 | 2.46 | 0.5275 |
7 | 2 | 2 | 0.5 | 0.5 | 0.2 | 0.1 | 14.15 | 0.5072 |
8 | 2 | 2 | 1 | 0.5 | 0.2 | 0.1 | 18.17 | 0.5503 |
9 | 1 | 0 | 0.5 | 0.5 | 0.2 | 0.5 | 2.34 | 0.2236 |
10 | 1 | 0 | 1 | 0.5 | 0.2 | 0.5 | 3 | 0.2574 |
11 | 1 | 2 | 0.5 | 0.5 | 0.2 | 0.5 | 17.29 | 0.2319 |
12 | 1 | 2 | 1 | 0.5 | 0.2 | 0.5 | 22.2 | 0.2617 |
13 | 2 | 0 | 0.5 | 0.5 | 0.2 | 0.5 | 2.34 | 0.3037 |
14 | 2 | 0 | 1 | 0.5 | 0.2 | 0.5 | 3 | 0.3476 |
15 | 2 | 2 | 0.5 | 0.5 | 0.2 | 0.5 | 17.29 | 0.3192 |
16 | 2 | 2 | 1 | 0.5 | 0.2 | 0.5 | 22.2 | 0.3556 |
One hundred simulated data sets were generated under each scenario. For each simulated data set, the ICC was estimated using four different methods: 1) the negative binomial GLMM, 2) LMM of raw data (untransformed); 3) LMM of square root transformed data; and 4) LMM of logarithm transformed data where taking logarithm of zero was avoided by replacing zeros with 0.5. For LMM, restricted maximum-likelihood estimation (REML) was used, while maximum-likelihood (ML) estimation was used for the GLMM. Relative bias, variance of the ICC, and mean square error (MSE) of the ICC estimate were calculated to evaluate the performance of the ICC. The relative bias was calculated as the difference between the mean of estimated ICC and it’s true value, variance was calculated by unbiased estimation based on the simulation, and MSE was calculated as the sum of squared bias and variance.
A second simulation explored the performance of the ICC estimate over a wider range. The mean estimated ICC was computed using 400 simulations per combination of number of successes (r = 5 and r = 30) and variance estimates for dairy and veterinarian (0 to 1 in increments of 0.2).
Field data analysis
Finally, field data used in the report by Aly et al. (2009) were analyzed using the negative binomial GLMM. Briefly, environmental samples were collected every other day on 3 different occasions from 4 California dairies between November 2006 and June 2007. Samples were cultured using bacterium-specific medium using standard microbiological procedures as reported by Aly et al. (2009). Confidence intervals for model parameters were obtained based on parameter estimates from the field data and using parametric bootstrap similar to that described in Table 1 (Efron and Tibshirani 1993). The resulting negative binomial based ICC was contrasted to that estimated from transformed data and reported previously by Aly et al. (2009). The R package lme4 was used for LMM analysis, and the package glmmADMB for GLMM analysis. All packages were loaded in the R 2.15.1 environment.
Results
Point estimate (PE) relative bias, variance, and mean square error (MSE) of Intraclass Correlation Coefficient (ICC) for culture results of samples collected by 2 veterinarians and based on the negative binomial mixed model, linear mixed model with raw data, square-root transformed data and log-transformed data (bold values are nearest to zero within a row)
Scenario | Parameter | Negative binomial | Transformed data | ||
---|---|---|---|---|---|
Raw | Natural logarithm | Square root | |||
1 | PE relative bias% | -10.35 | -16.14 | -5.41 | -5.32 |
Variance | 0.0098 | 0.0138 | 0.0145 | 0.0137 | |
MSE | 0.011 | 0.0168 | 0.0148 | 0.014 | |
2 | PE relative bias% | -10.8 | -17.21 | -1.13 | -2.55 |
Variance | 0.0108 | 0.0136 | 0.0198 | 0.0183 | |
MSE | 0.0126 | 0.0181 | 0.0198 | 0.0184 | |
3 | PE relative bias% | -5.33 | -7.65 | 8.45 | 12.43 |
Variance | 0.0052 | 0.0118 | 0.0115 | 0.012 | |
MSE | 0.0056 | 0.0126 | 0.0124 | 0.014 | |
4 | PE relative bias% | -9.3 | -16.93 | 9.75 | 9.7 |
Variance | 0.0067 | 0.0107 | 0.0205 | 0.0152 | |
MSE | 0.0081 | 0.0153 | 0.022 | 0.0167 | |
5 | PE relative bias% | -8.28 | -18.37 | -10.46 | -10.92 |
Variance | 0.0114 | 0.0135 | 0.0133 | 0.0136 | |
MSE | 0.0129 | 0.0207 | 0.0156 | 0.0161 | |
6 | PE relative bias% | -19.51 | -30.33 | -21.06 | -22.33 |
Variance | 0.0148 | 0.0138 | 0.0162 | 0.0161 | |
MSE | 0.0254 | 0.0394 | 0.0285 | 0.03 | |
7 | PE relative bias% | -8.02 | -14.27 | 0.41 | 2.54 |
Variance | 0.0095 | 0.012 | 0.0158 | 0.0122 | |
MSE | 0.0112 | 0.0172 | 0.0158 | 0.0124 | |
8 | PE relative bias% | -5.89 | -15.66 | 9.03 | 7.11 |
Variance | 0.009 | 0.0107 | 0.016 | 0.0131 | |
MSE | 0.01 | 0.0181 | 0.0185 | 0.0146 | |
9 | PE relative bias% | 17.53 | 3.26 | 27.01 | 26.74 |
Variance | 0.0129 | 0.0105 | 0.0126 | 0.0118 | |
MSE | 0.0144 | 0.0106 | 0.0162 | 0.0154 | |
10 | PE relative bias% | 8.55 | -0.51 | 31.12 | 27.35 |
Variance | 0.0165 | 0.0145 | 0.025 | 0.0225 | |
MSE | 0.017 | 0.0145 | 0.0314 | 0.0275 | |
11 | PE relative bias% | 30.36 | 27.17 | 62.05 | 66.58 |
Variance | 0.0104 | 0.0134 | 0.0144 | 0.015 | |
MSE | 0.0154 | 0.0174 | 0.0351 | 0.0388 | |
12 | PE relative bias% | 19.56 | 16.28 | 57.01 | 55.29 |
Variance | 0.0118 | 0.0157 | 0.0193 | 0.0183 | |
MSE | 0.0144 | 0.0175 | 0.0416 | 0.0392 | |
13 | PE relative bias% | 13.24 | 7.84 | 18.41 | 18.51 |
Variance | 0.0213 | 0.0185 | 0.0176 | 0.0183 | |
MSE | 0.0229 | 0.0191 | 0.0207 | 0.0215 | |
14 | PE relative bias% | 7.22 | -2.79 | 17.15 | 15.39 |
Variance | 0.0217 | 0.0172 | 0.027 | 0.0255 | |
MSE | 0.0223 | 0.0173 | 0.0306 | 0.0284 | |
15 | PE relative bias% | 28.41 | 23.59 | 45.99 | 48.25 |
Variance | 0.0154 | 0.0182 | 0.0178 | 0.0188 | |
MSE | 0.0236 | 0.0239 | 0.0394 | 0.0425 | |
16 | PE relative bias% | 22.69 | 19.99 | 51.97 | 50 |
Variance | 0.0216 | 0.0205 | 0.0262 | 0.023 | |
MSE | 0.0281 | 0.0256 | 0.0604 | 0.0546 |
Parameter estimates from a negative binomial generalized linear mixed model for culture results from a study on the reliability of an environmental sampling protocol and the Intraclass Correlation Coefficient (ICC) for similarity in samples collected by two veterinarians on the same day and from the same pen
95% Confidence interval | |||
---|---|---|---|
Parameter | Estimate | Lower | Upper |
β | 1.9516 | 1.3745 | 2.6011 |
r | 1.379 | 1.0138 | 2.0225 |
σ_{a} | 0.2691 | 2.07E-09 | 0.8657 |
σ_{b} | 1.352 | 0.5786 | 2.028 |
σ_{c} | 2.11E-09 | 2.06E-09 | 0.0303 |
σ_{d} | 4.72E-04 | 2.06E-09 | 0.0359 |
ICC | 0.5207 | 0.4033 | 0.6091 |
Discussion
The current study updates an earlier report on the reliability of environmental sampling to quantify MAP in freestall dairy pens utilizing the negative binomial ICC for count data. A unique character of the negative binomial ICC is the inclusion of the fixed effect intercept estimate unlike the ICC based on LMM which is based soley on variance components. Fixed effects are similarly included in the formula for the poisson ICC however the negative binomial ICC also includes r, the distribution parameter for number of successes. A performance comparison of the ICC estimates showed that the negative binomial ICC was more suitable for count data that is overdispersed given the smaller MSE and variance estimate than the ICC from LMM. Relative bias tended to the least in more scenarios (7 out of 16) with LMM compared to the GLMM based ICC. The lower relative bias with LMM may be explained by the use of REML estimation. The choice of MLE for GLMM was justified by that REML for GLMM has not been well defined, unlike for LMM (Jiang 2007). Nevertheless, the ICC for the negative binomial data outperformed that based on LMM of logarithm or square root transformed data with respect to MSE and variance. Results of a second simulation with highly overdispersed data showed that the NB ICC tended to overestimate the true ICC with higher variance components and under estimate with lower variance components. This expected behavior was consistent in a higher number of successes (r = 30) which confirms stability of the estimator over a wide variety of negative binomial distributed data.
Aly et al. (2009) estimated the ICC for similarity in HEYM culture results of MAP in samples collected by two different collectors on the same day and from the same pen to be 67.3%. The current study showed that the similarity in culture results estimated using the negative binomial ICC could be as low as 52.07%. Such a difference is expected given that the culture results are overdispersed count data. One reason for overdispersion may relate to the culture of MAP on HEYM protocol itself. Specifically fecal slurry samples undergo a decontamination step to limit bacterial growth on HEYM to mycobacteria. The decontamination step also reduces the number of MAP organisms resulting in samples with low MAP counts which may test negative (zero colony forming units) increasing the variance. For this reason, quantitative real-time PCR (qPCR) may remain the most suitable choice for testing freestall pen environmental samples for MAP.
Declarations
Acknowledgements
Funding for this project was provided by the Dairy Epidemiology Laboratory at the Veterinary Medicine Teaching and Research Center, School of Veterinary Medicine, University of California, Davis, National Science Foundation grant SES-1121794, and the National Institutes of Health grant R01-GM085205A1.
Authors’ Affiliations
References
- Aly SS, Anderson RJ, Whitlock RH, et al.: Reliability of environmental sampling to quantify Mycobacterium avium subspecies paratuberculosis on California free-stall dairies. J Dairy Sci 2009, 92: 3634-3642. 10.3168/jds.2008-1680View ArticleGoogle Scholar
- Carrasco JL: A generalized concordance correlation coefficient based on the variance components generalized linear mixed models for overdispersed count data. Biometrics 2010, 66: 897-904. 10.1111/j.1541-0420.2009.01335.xView ArticleGoogle Scholar
- Carrasco JL, Jover L: Concordance correlation coefficient applied to discrete data. Stat Med 2005, 24: 4021-4034. 10.1002/sim.2397View ArticleGoogle Scholar
- Casella G, Berger RL: Statistical Inference. Pacific Grove, CA: Duxbury/Thomson Learning, Australia; 2002.Google Scholar
- Efron B, Tibshirani R: An introduction to the bootstrap. New York: Chapman & Hall; 1993.View ArticleGoogle Scholar
- Jiang J: Linear and Generalized Linear Mixed Models and Their Applications. London: Springer, New York; 2007.Google Scholar
- Kittawornrat A, Wang C, Anderson G, et al.: Ring test evaluation of the repeatability and reproducibility of a Porcine reproductive and respiratory syndrome virus oral fluid antibody enzyme-linked immunosorbent assay. J Vet Diagn Invest 2012, 24: 1057-1063. 10.1177/1040638712457929View ArticleGoogle Scholar
- Solomon PJ, Taylor JMG: Orthogonality and transformations in variance components models. Biometrika 1999, 86: 289-300. 10.1093/biomet/86.2.289View ArticleGoogle Scholar
Copyright
This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/2.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.