Trajectory Analysis Of Longitudinal Cohort Based On Latent Class Mixed Model

Backgrounds:A Longitudinal cohort study collects repeated measurements of variables observed at different periods of time(age)in the life course.It captures the dynamic changing characteristics of risk factors,explores their relationships with diseases and focuses more on the cumulative effects or sensitive period.Trajectory analysis is one of the most frequently applied methods for longitudinal cohort data.Recently,latent class mixed model(LCMM)was frequently implemented for trajectory analysis,with aims of identifying potential trajectory groups and testing the relationship between trajectory groups and diseases.However,the validity of trajectory-assignment and trajectory parameter estimation of LCMM was unknown in different scenarios due to the unknowns of real trajectory groups.To date,studies on the reliability of LCMM in real population data(missing data,censoring data,missing-censoring data and data with different minimum follow-up times)are limited.This study intends to verify the accuracy of the trajectory-assignment and parameter estimation of LCMM in different scenarios by Monte Carlo simulations,identify potential BMI dynamic changing trajectories during young adulthood,examine the association of BMI trajectories with incident hypertension and determine the potentially critical period of the development of hypertension.Methods:Monte Carlo simulations were conducted to test the goodness of LCMM’s performance under the following scenarios:(1)different sample sizes in data without missing;(2)data with different missing rates under different missing patterns(missing completely at random,missing at random and missing not at random);(3)data with different censoring rates under three censoring patterns(left-truncated,right censored,both-sides-censored);(4)missing-censoring data(data with both MCAR and both-sides-censored);(5)missing-censoring data with different minimum follow-up times.Every simulation was repeated 1000 times and the following indexes were summarized to evaluate the performance of LCMM:probabilities of correct classification(PCC),bias,standard error of mean(SE),and mean squared error(MSE).In the cohort study,3271 participants(1712 males and 1559 females)who had BMI and blood pressure(BP)repeatedly measured 4 to 11 times during 2004-2015 were included.LCMM was applied for BMI trajectory analysis.The optimal model was determined according to the following criteria:(1)Bayesian information criterion decreased at least 20;(2)high mean posterior class membership probabilities(>0.65);(3)high mean posterior probabilities(>0.7).Cox proportional hazard models were fitted to investigate the association between the trajectory group membership and incident hypertension.The model-estimated BMI levels and linear slopes were calculated at each age point in 1-year intervals according to the model parameters and their first derivatives,respectively.Logistic regression analyses were used to examine the association of model-estimated levels and linear slopes of BMI at each age point with incident hypertension.Results:The simulation study showed:(1)The PCCs were all>0.95 in data without missing.The biases,SEs and MSEs became smaller with the increasing of sample size.The performance of LCMM was stable while the sample size>500.(2)In missing data,the PCCs became smaller while the biases,SEs and MSEs became bigger with the increasing of missing rate.Besides,the performances of LCMM were almost the same in the MCAR and MAR scenarios.LCMM had huge biases,SEs and MSEs in MNAR scenarios.(3)In censoring data,when the censoring rate increased,the PCCs became lower while the biases,SEs and MSEs got bigger.The performance of LCMM in left-truncated data was the best in censoring data when compared with the other two scenarios on condition that the censoring rate was fixed.(4)In missing-censoring data,the PCCs became lower while the biases,SEs and MSEs got bigger when the total missing rate increased.In addition,the performance of LCMM was worse if the data had a higher proportion of censoring.(5)In cohorts with different minimum follow-up times,as the increment of minimum follow-up time,the PCCs increased from 0.79 to 0.92 while the sample sizes decreased from 8663 to 1000.Besides,the biases were hardly changed while the SEs and MSEs increased slightly.Four distinct trajectory groups were identified using latent class growth mixture model in the cohort study:low-stable(n=1497),medium-increasing(n=1421),high-increasing(n=291),sharp-increasing(n=62).Model-estimated levels and linear slopes of BMI at each age point between ages 20 and 40 were calculated in 1-year intervals using the latent class growth mixture model parameters and their first derivatives,respectively.Compared with the low-stable group,the hazard ratios and 95%CI were 2.42(1.88,3.11),4.25(3.08,5.87),11.17(7.60,16.41)for the 3 increasing groups,respectively.After adjusting for covariates,the standardized odds ratios and 95%CI of model-estimated BMI level for incident hypertension increased in 20 to 35 years,ranging from 0.80(0.72-0.90)to 1.59(1.44-1.75);then decreased gradually to 1.54(1.42-1.68).The standardized odds ratios of level-adjusted linear slopes increased from 1.22(1.09-1.37)to 1.79(1.59-2.01)at 20 to 24 years;then decreased rapidly to 1.12(0.95-1.32).Conclusions:(1)As the increasing of missing rate,the probabilities of correct classification and the accuracy of LCMM decreased.(2)In missing data,censoring data and missing-censoring data,the performance of LCMM was good except MNAR data.(3)LCMM should not be applied in MNAR data for the biases,SEs and MSEs were too large.(4)The performance of LCMM was related to the dynamic separation of the repeated measured variable in censoring data.(5)Including subjects with only 1 or limited follow-ups would lead to the notable decreasing of PCCs other than the improvement of parameter estimation.(6)The results from cohort data analysis indicate that the level-independent BMI trajectories during young adulthood have a significant impact on hypertension risk.(7)Age between 20 and 30 years is a crucial period for incident hypertension,which has implications for early prevention.