Two-part random effects models (Olsen and Schafer,1 Tooze et al.2) have been applied to repeated measures of semi-continuous data, characterized by a mixture of a substantial proportion of zero values and a skewed distribution of positive values. In the original formulation of this model, the natural logarithm of the positive values is assumed to follow a normal distribution with a constant variance parameter. In this article, we review and consider three extensions of this model, allowing the positive values to follow (a) a generalized gamma distribution, (b) a log-skew-normal distribution, and (c) a normal distribution after the Box-Cox transformation. We allow for the possibility of heteroscedasticity. Maximum likelihood estimation is shown to be conveniently implemented in SAS Proc NLMIXED. The performance of the methods is compared through applications to daily drinking records in a secondary data analysis from a randomized controlled trial of topiramate for alcohol dependence treatment. We find that all three models provide a significantly better fit than the log-normal model, and there exists strong evidence for heteroscedasticity. We also compare the three models by the likelihood ratio tests for non-nested hypotheses (Vuong3). The results suggest that the generalized gamma distribution provides the best fit, though no statistically significant differences are found in pairwise model comparisons.
OlsenMSchaferJ. A two-part random effects model for semicontinuous longitudinal data. J Am Stat Assoc2001; 96: 730–745.
2.
ToozeJGrunwaldGJonesR. Analysis of repeated measures data with clumping at zero. Stat Meth Med Res2002; 11: 341–355.
3.
VuongQ. Likelihood ratio tests for model selection and non-nested hypotheses. Econometrica1989; 57: 307–333.
4.
ManningWMorrisCNewhouseJA two-part model of the demand for medical care: preliminary results from the health insurance study. In: van der GaagJPerlmanM (eds). Health, Economics, and Health Economics, Amsterdam: North-Holland, 1981, pp. 103–123.
5.
XieHMcHugoGSenguptaA. A method for analyzing long longitudinal outcomes with many zeros. Mental Health Serv Res2004; 6: 239–246.
6.
HyndmanRGrunwaldG. Generalized additive modelling of mixed distribution Markov models with application to Melbourne's rainfall. Aus NZ J Stat2000; 42: 145–158.
7.
JohnsonBARosenthalNCapeceJA. Topiramate for treating alcohol dependence - A randomized controlled trial. J Am Med Assoc2007; 298: 1641–1651.
8.
TobinJ. Estimation of relationships for limited dependent variables. Econometrics1958; 26: 24–36.
9.
AmemiyaT. Introduction to statistics and econometrics, Boston: Harvard University Press, 1994.
10.
MoultonLHalseyN. A mixture model with detection limits for regression analyses of antibody response to vaccine. Biometrics1995; 51: 1570–1578.
11.
ZhangMStrawdermanRCowenM. Bayesian inference for a two-part hierarchical model: an application to profiling providers in managed health care. J Am Stat Assoc2006; 101: 934–945.
12.
ManningWBasuAMullahyJ. Generalized modeling approaches to risk adjustment of skewed outcomes data. J Health Econom2005; 24(3): 465–488.
13.
LiuLStrawdermanRCowenM. A flexible two-part random effects model for correlated medical costs. J Health Econom2010; 29: 110–123.
14.
AzzaliniA. A class of distributions which includes the normal ones. Scand J Stat1985; 12: 171–178.
15.
ChaiHBaileyK. Use of log-skew-normal distribution in analysis of continuous data with a discrete component at zero. Stat Med2008; 27: 3643–3655.
16.
MahmudSLouWJohnstonN. A probit-log-skew-normal mixture model for repeated measures data with excess zeros, with application to a cohort study of paediatric respiratory symptoms. BMC Med Res Methodol2010; 10: 55.
17.
BoxGEPCoxDR. An analysis of transformations. J Roy Stat Soc Series B1964; 26: 211252.
18.
ToozeJAMidthuneDDoddKW. A new statistical method for estimating the usual intake of episodically consumed foods with application to their distribution. J Am Diet Assoc2006; 106: 1575–1587.
19.
RaudenbushSYangMYosefM. Maximum likelihood for generalized linear models with nested random effects via high-order, multivariate Laplace approximation. J Comput Graph Stat2000; 9: 141–157.
20.
LiuLConawayMKnausW. A random effects four-part model, with application to correlated medical costs. Comput Stat Data Anal2008; 52: 4458–4473.
21.
LittellRMillikenGStroupW. SAS for mixed model, 2nd ed. Cary, NC: SAS Institute Inc, 2006.
22.
SobellLSobellMTimeline follow-back: a technique for assessing self-reported alcohol consumption. In: LittenRAllenJ (eds). Measuring Alcohol Consumption: Psychosocial and Biochemical Methods, Totowa, NJ: Humana Press Inc, 1992, pp. 41–72.
23.
SobellMSobellLKlajnerF. The reliability of a timeline method for assessing normal drinker collegestudents recent drinking history - utility for alcohol research. Addict Behav1986; 11: 149–161.
24.
SobellLSobellMLeoG. Reliability of a timeline method - assessing normal drinkers reports of recent drinking and a comparative-evaluation across several populations. Brit J Addict1988; 83: 393–402.
25.
CareyKBCareyMPMaistoSA. Temporal stability of the timeline followback interview for alcohol and drug use with psychiatric outpatients. J Stud Alcohol2004; 65: 774–781.
26.
StramDLeeJ. Variance components testing in the longitudinal mixed effects model. Biometrics1994; 51: 1196.
27.
LuSELinYShihW. Analyzing excessive no changes in clinical trials with clustered data. Biometrics2004; 60: 257–267.
28.
AlbertPS. Letter to the editor. Biometrics2005; 47: 879–881.
29.
McDonaldJXuY. A generalization of the beta-distribution with applications. J Econom1995; 66(1–2): 133–152.
30.
SunJFreesERosenbergM. Heavy-tailed longitudinal data modeling using copulas. Insur Math Econom2008; 42: 817–830.
31.
ChiouJMullerH. Quasi-likelihood regression with unknown link and variance functions. J Am Stat Assoc1998; 93: 1376–1387.
32.
ChiouJMullerH. Estimated estimating equations: semiparametric inference for clustered and longitudinal data. J Roy Stat Soc Series B2005; 67: 531–553.
33.
BasuARathouzPJ. Estimating marginal and incremental effects on health outcomes using flexible link and variance function models. Biostatistics2005; 6: 93–109.
34.
LengCZhangWPanJ. Semiparametric mean-covariance regression analysis for longitudinal data. J Am Stat Assoc2010; 105: 181–193.
35.
ChenJLiuLJohnsonB. New estimation method for semiparametric mixed models, with application to alcohol treatment research. Stat Med2012. (Under revision.
36.
RigbyRDMS. Generalized additive models for location, scale and shape. Appl Stat2005; 54: 507–554.
37.
NelsonKLipsitzSFitzmauriceG. Use of the probability integral transformation to fit nonlinear mixed-effects models with nonnormal random effects. J Comput Graph Stat2006; 15: 39–57.
38.
LiuLYuZ. A likelihood reformulation method in non-normal random effects models. Stat Med2008; 27: 3105–3124.
39.
LiuLMaJJohnsonB. A multi-level two-part random effects model, with application to an alcohol dependence study. Stat Med2008; 27: 3528–3539.
40.
BerkKNLachenbruchPA. Repeated measures with zeros. Stat Meth Med Res2002; 11: 303–316.
41.
EpsteinMLinXBoehnkeM. A tobit variance-component method for linkage analysis of censored trait data. Am J Human Genet2003; 72: 611–620.
42.
BjerreBMarquesPSelenJ. Swedish alcohol ignition interlock programme for drink-drivers: effects on hospital care utilization and sick leave. Addiction2007; 102: 560–570.
43.
LiPSchneiderJWardM. Effect of critical access hospital conversion on patient safety. Health Serv Res2007; 42: 2089–2108.
44.
TwiskJRijmenF. Longitudinal tobit regression: a new approach to analyze outcome variables with floor or ceiling effects. J Clin Epidemiol2009; 62: 953–958.
45.
AbramowitzMStegunI. Handbook of mathematical functions, New York: Dover, 1972.
46.
GolubGHWelschJH. Calculation of Gaussian quadrature rules. Math Comput1969; 23: 221–230.
