Analysis Of Generalized Linear Mixed Models - USDA ARS

xiiMary Christman is currently the lead statistical consultant withMCC Statistical Consulting LLC, which provides statisticalexpertise for environmental and ecological problems. She isalso courtesy professor at the University of Florida. She wason the faculty at University of Florida, University of Maryland,and American University after receiving her Ph.D. in statisticsfrom George Washington University. She is a member of severalorganizations, including the American Statistical Association,the International Environmetrics Society, and the AmericanAssociation for the Advancement of Science. She received the 2004Distinguished Achievement Award from the Section on Statisticsand the Environment of the American Statistical Association.Her current research interests include linear and non-linearmodeling in the presence of correlated error terms, sampling andexperimental design, and statistical methodology for ecological andenvironmental research.Linda J. Young is Professor of Statistics at the University of Florida.She completed her Ph.D. in Statistics at Oklahoma State Universityand has previously served on the faculties of Oklahoma StateUniversity and the University of Nebraska, Lincoln. Linda hasserved the profession in a variety of capacities, including Presidentof the Eastern North American Region of the InternationalBiometric Society, Treasurer of the International Biometric Society,Vice-President of the American Statistical Association, and Chairof the Committee of Presidents of Statistical Societies. She has coauthored two books and has more than 100 refereed publications.She is a fellow of the American Association for the Advancementof Science, a fellow of the American Statistical Association, andan elected member of the International Statistical Institute. Herresearch interests include spatial statistics and statistical modeling.Mark West is a statistician for the USDA-Agricultural ResearchService. He received his Ph.D. in Applied Statistics from theUniversity of Alabama in 1989 and has been a statistical consultantin agriculture research ever since beginning his professional careerat Auburn University in 1989. His interests include experimentaldesign, statistical computing, computer intensive methods, andgeneralized linear mixed models.Matt Kramer is a statistician in the mid-Atlantic area (Beltsville, MD)of the USDA-Agricultural Research Service, where he has workedsince 1999. Prior to that, he spent eight years at the Census Bureauin the Statistical Research Division (time series and small areaestimation). He received a Masters and Ph.D. from the Universityof Tennessee. His interests are in basic biological and ecologicalstatistical applications.GLM.indb 1212/16/2011 10:28:37 AM

C o n v e r si o n F a c t o r sfor SI and Non-SI UnitsTo convertColumn 1 intoColumn 2multiply byColumn 1SI unit0.6211.0943.281.03.94 10 210kilometer, km (103 m)meter, mmeter, mmicrometer, µm (10 6 m)millimeter, mm (10 3 m)nanometer, nm (10 9 m)2.472470.3862.47 10 410.761.55 10 3hectare, hasquare kilometer, km2 (103 m)2square kilometer, km2 (103 m)2square meter, m2square meter, m2square millimeter, mm2(10 3 m)29.73 10 335.36.10 1042.84 10 21.0573.53 10 20.26533.782.11cubic meter, m3cubic meter, m3cubic meter, m3liter, L (10 3 m3)liter, L (10 3 m3)liter, L (10 3 m3)liter, L (10 3 m3)liter, L (10 3 m3)liter, L (10 3 m3)2.20 10 33.52 10 22.2050.01gram, g (10 3 kg)gram, g (10 3 kg)kilogram, kgkilogram, kgpound, lbounce (avdp), ozpound, lbquintal (metric), q45428.40.4541001.10 10 31.1021.102kilogram, kgmegagram, Mg (tonne)tonne, tton (2000 lb), tonton (U.S.), tonton (U.S.), ton9070.9070.9070.8937.77 10 2kilogram per hectare, kg ha 1kilogram per cubic meter,kg m 3kilogram per hectare, kg ha 1kilogram per hectare, kg ha 1Column 2non-SI unitTo convertColumn 2into Column 1multiply byLengthmile, miyard, ydfoot, ftmicron, µinch, inAngstrom, Å1.6090.9140.3041.025.40.1Areaacreacresquare mile, mi2acresquare foot, ft2square inch, in20.4054.05 10 32.5904.05 1039.29 10 2645Volumeacre-inchcubic foot, ft3cubic inch, in3bushel, buquart (liquid), qtcubic foot, ft3gallonounce (fluid), ozpint (fluid), pt102.82.83 10 21.64 10 535.240.94628.33.782.96 10 20.473MassYield and Rate1.49 10 21.59 10 2pound per acre, lb acre 1pound per bushel, lb bu 11.1212.87bushel per acre, 60 lbbushel per acre, 56 lb67.1962.71continuedxiiiGLM.indb 1312/16/2011 10:28:37 AM

xivTo convertColumn 1 intoColumn 2multiply byColumn 1SI unitColumn 2non-SI unitTo convertColumn 2into Column 1multiply by1.86 10 20.1078938930.4462.24kilogram per hectare, kg ha 1liter per hectare, L ha 1tonne per hectare, t ha 1megagram per hectare, Mg ha 1megagram per hectare, Mg ha 1meter per second, m s 1bushel per acre, 48 lbgallon per acrepound per acre, lb acre 1pound per acre, lb acre 1ton (2000 lb) per acre, ton acre 1mile per hour53.759.351.12 10 31.12 10 32.240.44710square meter per kilogram,m2 kg 1square meter per kilogram,m2 kg 1Specific Surface1000square centimeter per gram,cm2 g 1square millimeter per gram,mm2 g 10.10.001Density1.00megagram per cubic meter,Mg m 39.90102.09 10 21.45 10 4megapascal, MPa (106 Pa)megapascal, MPa (106 Pa)pascal, Papascal, Pa1.00 (K 273)(9/5 C) 32kelvin, KCelsius, C9.52 10 40.2391070.7352.387 10 5joule, Jjoule, Jjoule, Jjoule, Jjoule per square meter, J m 21051.43 10 3newton, Nwatt per square meter, W m 23.60 10 2milligram per square metersecond, mg m 2 s 1milligram (H2O) per square metersecond, mg m 2 s 1gram per cubic centimeter, g cm 3 1.00Pressureatmospherebarpound per square foot, lb ft 2pound per square inch, lb in 20.1010.147.96.90 103TemperatureCelsius, CFahrenheit, F1.00 ( C 273)5/9 ( F 32)Energy, Work, Quantity of HeatBritish thermal unit, Btucalorie, calergfoot-poundcalorie per square centimeter(langley)dynecalorie per square centimeterminute (irradiance),cal cm 2 min 11.05 1034.1910 71.364.19 10410 5698Transpiration and Photosynthesis5.56 10 310 435.97milligram per square metersecond, mg m 2 s 1milligram per square metersecond, mg m 2 s 1gram per square decimeter hour,g dm 2 h 1micromole (H2O) per squarecentimeter second,µmol cm 2 s 1milligram per square centimetersecond, mg cm 2 s 1milligram per square decimeterhour, mg dm 2 h 127.81801042.78 10 2continuedGLM.indb 1412/16/2011 10:28:37 AM

xvTo convertColumn 1 intoColumn 2multiply byColumn 1SI unit57.3radian, radColumn 2non-SI unitTo convertColumn 2into Column 1multiply byPlane Angledegrees (angle), 104Electrical Conductivity, Electricity, and Magnetismsiemen per meter, S m 1millimho per centimeter,mmho cm 1tesla, Tgauss, G9.73 10 39.81 10 34.40cubic meter, m3cubic meter per hour, m3 h 1cubic meter per hour, m3 h 18.1197.288.1 10 2hectare meter, ha mhectare meter, ha mhectare centimeter, ha cm10.11centimole per kilogram, cmol kg 1 milliequivalent per 100 grams,meq 100 g 1percent, %gram per kilogram, g kg 1milligram per kilogram, mg kg 1 parts per million, ppm2.7 10 112.7 10 2100100becquerel, Bqbecquerel per kilogram, Bq kg 1gray, Gy (absorbed dose)sievert, Sv (equivalent dose)101.75 10 20.110 4Water Measurementacre-inch, acre-incubic foot per second, ft3 s 1U.S. gallon per minute,gal min 1acre-foot, acre-ftacre-inch, acre-inacre-foot, acre-ft102.8101.90.2270.1231.03 10 212.33Concentration1101Radioactivitycurie, Cipicocurie per gram, pCi g 1rad, rdrem (roentgen equivalent man)3.7 1010370.010.01Plant Nutrient ConversionElemental2.291.201.391.66GLM.indb 6/2011 10:28:37 AM

GLM.indb 1612/16/2011 10:28:37 AM

Chapter 1Introduction1.1 IntroductionOver the past generation, dramatic advances have occurred in statistical methodology, many of which are relevant to research in the agricultural and naturalresources sciences. These include more theoretically sound approaches to theanalysis of spatial data; data taken over time; data involving discrete, categorical,or continuous but non-normal response variables; multi-location and/or multiyear data; complex split-plot and repeated measures data; and genomic data suchas data from microarray and quantitative genetics studies. The development ofgeneralized linear mixed models has brought together these apparently disparateproblems under a coherent, unified theory. The development of increasingly userfriendly statistical software has made the application of this methodology accessible to applied researchers.The accessibility of generalized linear mixed model software has coincidedwith a time of change in the research community. Research budgets have been tightening for several years, and there is every reason to expect this trend to continue forthe foreseeable future. The focus of research in the agricultural sciences has beenshifting as the nation and the world face new problems motivated by the need forclean and renewable energy, management of limited natural resources, environmental stress, the need for crop diversification, the advent of precision agriculture, safetydilemmas, and the need for risk assessment associated with issues such as genetically modified crops. New technologies for obtaining data offer new and importantpossibilities but often are not suited for design and analysis using conventionalapproaches developed decades ago. With this rapid development comes the lack ofaccepted guidelines for how such data should be handled.Researchers need more efficient ways to conduct research to obtain useableinformation with the limited budgets they have. At the same time, they need waysto meaningfully analyze and understand response variables that are very different from those covered in “traditional” statistical methodology. Generalized linearmixed models allow more versatile and informative analysis in these situationsand, in the process, provide the tools to facilitate experimental designs tailored .c1Copyright 2012American Society of Agronomy, Crop Science Society of America, and Soil Science Society of America5585 Guilford Road, Madison, WI 53711-5801, USA.Analysis of Generalized Linear Mixed Models in the Agricultural and Natural Resources SciencesEdward E. Gbur, Walter W. Stroup, Kevin S. McCarter, Susan Durham, Linda J. Young, Mary Christman,Mark West, and Matthew Kramer1GLM.indb 112/16/2011 10:28:37 AM

2Chapter 1the needs of particular studies. Such designs are often quite different from conventional experimental designs. Thus, generalized linear mixed models provide anopportunity for a comprehensive rethinking of statistical practice in agriculturaland natural resources research. This book provides a practical introductory guideto this topic.1.2 Generalized Linear Mixed ModelsIn introductory statistical methods courses taken by nearly every aspiring agricultural scientist in graduate school, statistical analysis is presented in some way,shape, or form as an attempt to make inferences on observations that are the sumof “explanatory” components and “random” components. In designed experiments and quasi-experiments (i.e., studies structured as closely as possible to designed experiments), “explanatory” means treatment effect and “random” meansresidual or random error. Thus, the formulaobserved response explanatory randomexpresses the basic building blocks of statistical methodology. This simple breakdown is necessarily elaborated intoobserved response treatment design effects errorwhere design effects include blocks and covariates. The observed response isinevitably interpreted as having a normal distribution and analysis of variance(ANOVA), regression, and analysis of covariance are presented as the primarymethods of analysis. In contemporary statistics, such models are collectivelyreferred to as linear models. In simple cases, a binomial distribution is considered for the response variable leading to logit analysis and logistic regression.Occasionally probit analysis is considered as well.In contrast, consider what the contemporary researcher actually faces. Table1–1 shows the types of observed response variables and explanatory model components that researchers are likely to encounter. Note that “conventional” statisticalmethodology taught in introductory statistics courses and widely considered as“standard statistical analysis” in agricultural research and journal publication isconfined to the first row and occasionally the second row in the table. Obviously,the range of methods considered “standard” is woefully inadequate given therange of possibilities now faced by contemporary researchers.This inadequacy has a threefold impact on potential advances in agriculturaland applied research. First, it limits the types of analyses that researchers (andjournal editors) will consider, resulting in cases where “standard methods” area mismatch between the observed response and an explanatory model. Second,it limits researchers’ imaginations when planning studies, for example througha lack of awareness of alternative types of response variables that contemporarystatistical methods can handle. Finally, it limits the efficiency of experiments inthat traditional designs, while optimized for normal distribution based ANOVAGLM.indb 212/16/2011 10:28:37 AM

3introductionTable 1 – 1 . Statistical model scenarios corresponding to combinations of types of observedresponses and explanatory model components.Explanatory model componentsType of responsevariableFixed effectsCorrelatederrorsExamples ofdistributionsCategoricalContinuousRandom NOVA†,‡,§,¶regression†,‡,§,¶split omiallogitanalysis§,¶logisticregression §,¶—¶—¶Poisson, negativebinomiallog-linearmodel §,¶Poissonregression mal,gamma, beta—§,¶—§,¶—¶—¶Time to �¶—¶Count† Linear model scenarios are limited to the first two cells in the first row of the table.‡ Linear mixed model scenarios are limited to first row of the table.§ Generalized linear model scenarios are limited to first two columns of the table.¶ Generalized linear mixed model scenarios cover all cells shown in the table.and regression, often are not well suited to the majority of the response variable–explanatory model combinations in Table 1–1.Two major advances in statistical theory and methodology that occurred in thelast half of the 20th century were the development of linear mixed models and generalized linear models. Mixed models incorporate random effects and correlatederrors; that is, they deal with all four columns of explanatory model componentsin Table 1–1. Generalized linear models accommodate a large class of probabilitydistributions of the response; that is, they deal with the response variable columnin the table. The combination of mixed and generalized linear models, namely generalized linear mixed models, addresses the entire range of options for the responsevariable and explanatory model components (i.e., with all 20 combinations in Table1 –1). Generalized linear mixed models represent the primary focus of this book.1.3 Historical DevelopmentSeal (1967) traced the origin of fixed effects models back to the development ofleast squares by Legendre in 1806 and Gauss in 1809, both in the context of problems in astronomy. It is less well known that the origin of random effects modelscan be ascribed to astronomy problems as well. Scheffé (1956) attributed early useGLM.indb 312/16/2011 10:28:38 AM

4Chapter 1of random effects to Airy in an 1861 publication. It was not until nearly 60 yearslater that Fisher (1918) formally introduced the terms variance and analysis of variance and utilized random effects models.Fisher’s 1935 first edition of The Design of Experiments implicitly discussesmixed models (Fisher, 1935). Scheffé (1956) attributed the first explicit expressionof a mixed model equation to Jackson (1939). Yates (1940) developed methodsto recover inter-block information in block designs that are equivalent to mixedmodel analysis with random blocks. Eisenhart (1947) formally identified random,fixed, and mixed models. Henderson (1953) was the first to explicitly use mixedmodel methodology for animal genetics studies. Harville (1976, 1977) publishedthe formal overall theory of mixed models.Although analyses of special cases of non-normally distributed responses suchas probit analysis (Bliss, 1935) and logit analysis (Berkson, 1944) existed in the context of bioassays, standard statistical methods textbooks such as Steel et al. (1997)and Snedecor and Cochran (1989) dealt with the general problem of non-normality through the use of transformations. The ultimate purpose of transformationssuch as the logarithm, arcsine, and square root was to enable the researcher toobtain approximate analyses using the standard normal theory methods. Box andCox (1964) proposed a general class of transformations that include the above asspecial cases. They too have been applied to allow use of normal theory methods.Nelder and Wedderburn (1972) articulated a comprehensive theory of linearmodels with non-normally distributed response variables. They assumed that theresponse distribution belonged to the exponential family. This family of probability distributions contains a diverse set of discrete and continuous distributions,including all of those listed in Table 1–1. The models were referred to as generalized linear models (not to be confused with general linear models which has beenused in reference to normally distributed responses only). Using the concept ofquasi-likelihood, Wedderburn (1974) extended applicability of generalized linearmodels to certain situations where the distribution cannot be specified exactly. Inthese cases, if the observations are independent or uncorrelated and the form ofthe mean/variance ratio can be specified, it is possible to fit the model and obtainresults similar to those which would have been obtained if the distribution hadbeen known. The monograph by McCullagh and Nelder (1989) brought generalized linear models to the attention of the broader statistical community and with it,the beginning of research on the addition of random effects to these models—thedevelopment of generalized linear mixed models.By 1992 the conceptual development of linear models through and includinggeneralized linear mixed models had been accomplished, but the computationalcapabilities lagged. The first usable software for generalized

of software for implementing generalized linear mixed models, we have found researchers increasingly interested in using these models, but it is "easier said than done." Our goal is to help those who have worked with linear mixed models to begin moving toward generalized linear mixed models. The benefits and chal-lenges are discussed from a .