使用lavaan拟合三变量两时间点LCSM遇梯度警告求助

我在研究中尝试用lavaan对三个随时间变化变量（PM_event、PM_time、SP_total）的两时间点纵向数据拟合潜变化分数模型（LCSM），数据集包含6列（如PM_event_base、PM_event_fu等）。编写代码后，lavaan持续抛出警告：优化器NLMINB声称模型收敛，但并非所有梯度元素趋近于零，可能未找到局部解，同时参数估计完全不符合预期。我的目标是用三个变量的基线值共同预测每个变量从基线（时间0）到随访（时间1）的变化量，作为lavaan新手，参考教程调整语法后仍未解决问题，附上代码寻求帮助。

原始代码 #

#define phantom endogeneous variables
#variance of phantom variables is 0 as all residual is accounted for in #exogeneous factor
PM_event_base_p =~ 1 * PM_event_base 
PM_event_base ~ 0
PM_event_base ~~ PM_event_base
PM_event_base_p ~~ 0* PM_event_base_p

PM_event_fu_p =~ 1 * PM_event_fu 
PM_event_fu ~ 0
PM_event_fu ~~ PM_event_fu
PM_event_fu_p ~~ 0*PM_event_fu_p

PM_time_base_p =~ 1 * PM_time_base
PM_time_base ~ 0
PM_time_base ~~ PM_time_base
PM_time_base_p ~~ 0*PM_time_base_p

PM_time_fu_p =~ 1 * PM_time_fu
PM_time_fu ~ 0
PM_time_fu ~~ PM_time_fu
PM_time_fu_p ~~ 0*PM_time_fu_p

SP_total_base_p =~ 1 * SP_total_base 
SP_total_base ~ 0 
SP_total_base ~~ SP_total_base
SP_total_base_p ~~ 0*SP_total_base_p

SP_total_fu_p =~ 1 * SP_total_fu
SP_total_fu ~ 0
SP_total_fu ~~ SP_total_fu;
SP_total_fu_p ~~ 0*SP_total_fu_p

#covariances of baseline scores
PM_time_base_p ~~ PM_event_base_p
PM_time_base_p ~~ SP_total_base_p
PM_event_base_p ~~ SP_total_base_p

#set regressions of adjacent timepoints to 1
PM_time_base_p ~ 1*PM_time_fu_p
PM_event_base_p ~ 1*PM_event_fu_p
PM_time_base_p ~ 1*SP_total_fu_p

#define latent change variables
delta_PM_event =~ 1*PM_event_base_p
delta_PM_time =~ 1*PM_time_base_p
delta_SP =~ 1*SP_total_p

#set variance of change scores to 0
#delta_PM_event ~~ 0*delta_PM_event
#delta_PM_time ~~ 0*delta_PM_time
#delta_SP ~~ 0*delta_SP

#regress latent change variables on baseline scores
delta_PM_event ~ PM_event_base_p + PM_time_base_p + SP_total_base_p
delta_PM_time ~ PM_event_base_p + PM_time_base_p + SP_total_base_p
delta_SP ~ PM_event_base_p + PM_time_base_p + SP_total_base_p

#covariances between change scores
delta_PM_event ~~ delta_PM_time
delta_PM_event ~~ delta_SP
delta_PM_time ~~ delta_SP

#define paths through time
#follow-up = baseline + delta
#already fully determined by the change score and baseline, so no regression coefficients included
PM_event_fu_p ~ 1* delta_PM_event + 1 * PM_event_base_p
PM_time_fu_p ~ 1* delta_PM_time + 1* PM_time_base_p
SP_total_fu_p ~ 1* delta_SP + 1* SP_total_base_p

#remove unwanted covariances
#these should not be needed as follow-up data is already fully determined by change scores
PM_event_fu_p ~~ 0*PM_time_fu_p
PM_event_fu_p ~~ 0* SP_total_fu_p
PM_time_fu_p ~~ 0*SP_total_fu_p

问题分析与修正方案 #

原始代码存在多个逻辑与语法问题，导致模型收敛异常：

• 变量名错误：定义delta_SP时使用了未声明的SP_total_p，应为SP_total_base_p
• 冗余的Phantom变量：两时间点LCSM无需额外定义这类潜变量，直接用观测变量构建模型即可，冗余变量会增加模型复杂度与识别风险
• 反向回归路径：PM_time_base_p ~ 1*PM_time_fu_p将基线变量回归到随访变量，完全违背LCSM“随访=基线+变化量”的逻辑
• 不必要的约束：给观测变量固定截距为0、给phantom变量固定方差为0，这类约束会干扰模型识别，导致参数估计异常
• 路径冲突：同时定义了时间点间的反向回归与“随访=基线+变化量”的路径，造成模型过度约束

修正后的代码

# 两时间点LCSM标准语法
model <- '
  # 定义潜变化分数：变化量 = 随访得分 - 基线得分
  delta_PM_event =~ 1*PM_event_fu + (-1)*PM_event_base
  delta_PM_time =~ 1*PM_time_fu + (-1)*PM_time_base
  delta_SP =~ 1*SP_total_fu + (-1)*SP_total_base

  # 用三个基线变量预测每个变化分数
  delta_PM_event ~ PM_event_base + PM_time_base + SP_total_base
  delta_PM_time ~ PM_event_base + PM_time_base + SP_total_base
  delta_SP ~ PM_event_base + PM_time_base + SP_total_base

  # 允许变化分数之间存在协方差
  delta_PM_event ~~ delta_PM_time
  delta_PM_event ~~ delta_SP
  delta_PM_time ~~ delta_SP

  # 基线变量间的协方差由模型自由估计（默认行为，可省略）
  PM_event_base ~~ PM_time_base
  PM_event_base ~~ SP_total_base
  PM_time_base ~~ SP_total_base
'

# 拟合模型（若数据存在非正态/缺失，可改用MLR估计器）
fit <- cfa(model, data = your_dataset, estimator = "ML")
# 输出结果，含拟合指标与标准化系数
summary(fit, fit.measures = TRUE, standardized = TRUE)

额外调试建议

• 检查数据质量：排查缺失值、极端值，确保变量分布合理，基线与随访得分的相关性符合预期
• 更换优化器：若仍有收敛警告，尝试使用estimator = "MLR"（稳健最大似然）或optim.method = "BFGS"
• 逐步调试：先拟合简化模型（比如先只让每个变化分数由自身基线预测），再逐步增加预测变量，排查冲突点

内容的提问来源于stack exchange，提问作者JSK