Help!! path analysis with observed variables of diverse types

73 views
Skip to first unread message

Marco Monarca

unread,
Feb 20, 2025, 11:26:09 AM2/20/25
to lav...@googlegroups.com
Dear all,

I'm a nature science student, very new to SEM and lavaan, and I really need your help to evaluate the soundness of my methodology. In my study, I use path analysis to study carnivore-farmers conflicts in Eastern Honduras. My goal is to integrate the path analysis into a Causal Loop Diagram explaining socioeconomic implications in the shaping of stakeholders perceptions towards carnivores. 

The sample size is rather small: 104 observations. In my model, I include variables of different type, and with different levels, and this is the main challenge.

Exogenous Variables:

Residency time: continuous
Education level: ordinal, 4 levels: 1,2,3,4
Age: continuous
Ethnicity: binary (indigenous, non-indigenous)

Endogenous Variables:

a) Moderators
Adaptive Husbandry: binary (yes, no)
Acknowledgement Benefits: ordinal, 3 levels: -1,0,1
Engagement in Conservation: binary (yes,no)
Awareness Ecological Value: binary (yes no
b) Outcome variables
Tolerance: binary (yes,no)
Readiness to Denounce Poaching: binary (yes,no)
Acceptability Lethal Control: ordinal/likert, 5 levels: 1,2,3,4,5

I chose to treat the ordinal variables as numeric, because of the small sample size, using therefore Maximum Likelihood as estimator. I then studied the indirect effects to test the validity of the causal path. I scaled continuous variables to tackle the issue of extreme differences in variances, and used "fiml" to deal with NAs. 

here is the code in lavaan:

df_tolerance$residency_scaled <- scale(df_tolerance$residency)
df_tolerance$economic_loss_scaled <- scale(df_tolerance$economic_loss)
df_tolerance$age_scaled <- scale(df_tolerance$age)
df_tolerance$residency_clean_scaled <- scale(df_tolerance$residency_clean)

model_joint <- 'adaptive_husbandry ~ a1*residency_scaled + a2*ethnicity + a3*age_scaled + a4*education_level
                engagement_in_conservation ~ b2*acknowledgement_benefits
                awareness ~ acknowledgement_benefits
                acknowledgement_benefits ~ residency_scaled + education_level
                tolerance ~ b1*adaptive_husbandry + b4*awareness
                readiness_to_denounce ~ b3*engagement_in_conservation + age_scaled
                acceptability_lethal_control ~ acknowledgement_benefits + engagement_in_conservation
               

#indirect effects
IE_residency_tolerance := a1*b1
IE_ethnicity_tolerance := a2*b1
IE_age_tolerance := a3*b1
IE_residency_denounce := a1 * b2 * b3
IE_acknowledgement_denounce := b2 * b3
IE_acknowledgement_tolerance := b2* b4
IE_residency_tolerance2 := a1*b2*b4
IE_residency_lethal := a1*b2
IE_education_tolerance := a4*b1
IE_education_tolerance2 := a4*b2*b4
IE_education_lethal := a4*b2
'


fit_joint <- sem(model_joint, data = df_tolerance, missing="fiml")
summary(fit_joint, standardized=TRUE, fit.measures=TRUE, rsquare=TRUE)


The fit of the model is good, the regressions make sense and I hope I handled the data well. Although, small r² value are not entirely satisfying, but I hope the reason is that the sample is just very small.

Here are the results:

lavaan 0.6-19 ended normally after 40 iterations

  Estimator                                         ML
  Optimization method                           NLMINB
  Number of model parameters                        31

                                                  Used       Total
  Number of observations                           102         104
  Number of missing patterns                         9            

Model Test User Model:
                                                      
  Test statistic                                39.484
  Degrees of freedom                                32
  P-value (Chi-square)                           0.170

Model Test Baseline Model:

  Test statistic                               173.066
  Degrees of freedom                                49
  P-value                                        0.000

User Model versus Baseline Model:

  Comparative Fit Index (CFI)                    0.940
  Tucker-Lewis Index (TLI)                       0.908
                                                      
  Robust Comparative Fit Index (CFI)             0.940
  Robust Tucker-Lewis Index (TLI)                0.908

Loglikelihood and Information Criteria:

  Loglikelihood user model (H0)               -520.455
  Loglikelihood unrestricted model (H1)       -500.713
                                                      
  Akaike (AIC)                                1102.910
  Bayesian (BIC)                              1184.284
  Sample-size adjusted Bayesian (SABIC)       1086.366

Root Mean Square Error of Approximation:

  RMSEA                                          0.048
  90 Percent confidence interval - lower         0.000
  90 Percent confidence interval - upper         0.092
  P-value H_0: RMSEA <= 0.050                    0.496
  P-value H_0: RMSEA >= 0.080                    0.129
                                                      
  Robust RMSEA                                   0.049
  90 Percent confidence interval - lower         0.000
  90 Percent confidence interval - upper         0.094
  P-value H_0: Robust RMSEA <= 0.050             0.477
  P-value H_0: Robust RMSEA >= 0.080             0.145

Standardized Root Mean Square Residual:

  SRMR                                           0.058

Parameter Estimates:

  Standard errors                             Standard
  Information                                 Observed
  Observed information based on                Hessian

Regressions:
                                 Estimate  Std.Err  z-value  P(>|z|)   Std.lv
  adaptive_husbandry ~                                                       
    rsdncy_sc (a1)                  0.063    0.032    1.951    0.051    0.063
    ethnicity (a2)                  0.239    0.101    2.358    0.018    0.239
    age_scald (a3)                  0.131    0.034    3.876    0.000    0.131
    edctn_lvl (a4)                  0.100    0.038    2.669    0.008    0.100
  engagement_in_conservation ~                                               
    acknwldg_ (b2)                  0.173    0.046    3.735    0.000    0.173
  awareness ~                                                                
    acknwldg_                       0.088    0.035    2.493    0.013    0.088
  acknowledgement_benefits ~                                                 
    rsdncy_sc                       0.190    0.089    2.133    0.033    0.190
    edctn_lvl                       0.207    0.102    2.038    0.042    0.207
  tolerance ~                                                                
    adptv_hsb (b1)                  0.372    0.129    2.874    0.004    0.372
    awareness (b4)                  0.385    0.137    2.802    0.005    0.385
  readiness_to_denounce ~                                                    
    enggmnt__ (b3)                  0.442    0.094    4.676    0.000    0.442
    age_scald                       0.121    0.043    2.838    0.005    0.121
  acceptability_lethal_control ~                                             
    acknwldg_                      -0.361    0.171   -2.113    0.035   -0.361
    enggmnt__                      -1.398    0.330   -4.236    0.000   -1.398
  Std.all
         
    0.176
    0.208
    0.360
    0.244
         
    0.350
         
    0.245
         
    0.204
    0.194
         
    0.265
    0.257
         
    0.423
    0.251
         
   -0.209
   -0.399

Covariances:
                           Estimate  Std.Err  z-value  P(>|z|)   Std.lv
 .tolerance ~~                                                         
   .readnss_t_dnnc            0.021    0.020    1.059    0.290    0.021
   .accptblty_lth_           -0.015    0.067   -0.230    0.818   -0.015
 .readiness_to_denounce ~~                                             
   .accptblty_lth_           -0.071    0.061   -1.162    0.245   -0.071
  Std.all
         
    0.108
   -0.024
         
   -0.125

Intercepts:
                   Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
   .adaptv_hsbndry   -0.002    0.053   -0.039    0.969   -0.002   -0.006
   .enggmnt_n_cnsr    0.743    0.044   16.868    0.000    0.743    1.632
   .awareness         0.895    0.033   26.888    0.000    0.895    2.683
   .acknwldgmnt_bn   -0.502    0.151   -3.327    0.001   -0.502   -0.543
   .tolerance         0.094    0.130    0.721    0.471    0.094    0.187
   .readnss_t_dnnc    0.004    0.080    0.050    0.960    0.004    0.008
   .accptblty_lth_    4.077    0.280   14.539    0.000    4.077    2.554

Variances:
                   Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
   .adaptv_hsbndry    0.089    0.013    7.069    0.000    0.089    0.700
   .enggmnt_n_cnsr    0.182    0.026    7.062    0.000    0.182    0.877
   .awareness         0.105    0.015    7.118    0.000    0.105    0.940
   .acknwldgmnt_bn    0.786    0.111    7.065    0.000    0.786    0.921
   .tolerance         0.215    0.030    7.099    0.000    0.215    0.860
   .readnss_t_dnnc    0.170    0.024    6.990    0.000    0.170    0.756
   .accptblty_lth_    1.884    0.273    6.907    0.000    1.884    0.739

R-Square:
                   Estimate
    adaptv_hsbndry    0.300
    enggmnt_n_cnsr    0.123
    awareness         0.060
    acknwldgmnt_bn    0.079
    tolerance         0.140
    readnss_t_dnnc    0.244
    accptblty_lth_    0.261

Defined Parameters:
                   Estimate  Std.Err  z-value  P(>|z|)   Std.lv  Std.all
    IE_rsdncy_tlrn    0.023    0.015    1.601    0.109    0.023    0.047
    IE_thncty_tlrn    0.089    0.049    1.822    0.068    0.089    0.055
    IE_age_tolernc    0.049    0.021    2.314    0.021    0.049    0.095
    IE_rsdncy_dnnc    0.005    0.003    1.621    0.105    0.005    0.026
    IE_cknwldgmnt_    0.076    0.026    2.914    0.004    0.076    0.148
    IE_cknwldgmnt_    0.066    0.030    2.240    0.025    0.066    0.090
    IE_rsdncy_tlr2    0.004    0.003    1.471    0.141    0.004    0.016
    IE_rsdncy_lthl    0.011    0.006    1.729    0.084    0.011    0.062
    IE_edctn_tlrnc    0.037    0.019    1.957    0.050    0.037    0.065
    IE_dctn_tlrnc2    0.007    0.004    1.716    0.086    0.007    0.022
    IE_eductn_lthl    0.017    0.008    2.172    0.030    0.017    0.085


Questions:
1) (Considering the sample size) Do I need to use DWSL instead of ML? Even if I have a categorical ordinal variable as exogenous? In that case, do I have to set as "ordered" binary variables too?
2) Is there any other thing I have to think about while validating my model that I'm missing?
3) Do You have any suggestions?

Thank You so much in advance for any help you can provide!!!!

path_perceptions.png

Jeremy Miles

unread,
Feb 20, 2025, 3:19:43 PM2/20/25
to lav...@googlegroups.com

First, that's a heck of a complex model for a pretty small sample size. 

Your specific questions:

1) (Considering the sample size) Do I need to use DWSL instead of ML? Even if I have a categorical ordinal variable as exogenous? In that case, do I have to set as "ordered" binary variables too?

No, I don't think your sample is large enough. I would be surprised it it converged / was identified (although there's no harm in trying). It took 40 iterations to converge, which is a lot.

2) Is there any other thing I have to think about while validating my model that I'm missing?

Not that are obvious to me.

3) Do You have any suggestions?

Your sample is too small to draw much in the way of conclusions. You have little power to reject the model if it is wrong (and a lot of parameters with large standard errors).

Your chi-square is not significant, but that's not a huge surprise at this sample size. Your null model is not awful, hence your TLI is not particularly high. The confidence intervals on your RMSEA are pretty wide.

The path diagram is not hugely clear to me (it's also very small, I can't read it). I try to make the diagram 'flow' from left to right - it looks like you have four exgoenous variables. Start with those four on the left, then these all have arrows pointing to the next two variables, etc.

Hope that helps (and wasn't too harsh),

Jeremy





--
You received this message because you are subscribed to the Google Groups "lavaan" group.
To unsubscribe from this group and stop receiving emails from it, send an email to lavaan+un...@googlegroups.com.
To view this discussion visit https://groups.google.com/d/msgid/lavaan/CAFFwN-ytSnw%2BvuB%3D3wAv23VR2Q3qLgozpcsmzjYMQOxFVtiNDQ%40mail.gmail.com.

Stas Kolenikov

unread,
Feb 20, 2025, 4:28:37 PM2/20/25
to lav...@googlegroups.com
It is a triangular system (the arrows flow in a specific direction, so you can write it up with demographics on top, and acceptability on the bottom, and the matrix of the coefficients to be estimated will be lower triangular). I'd say it should be estimated as an econometric simultaneous equations model (series of regressions), and not bother with formulating it as a latent variable model (which is what lavaan really expects). My understanding is that lavaan recognizes demographics as exogenous variables here, but since you asked for FIML, it still tries too hard to fit the model that is too reliant on multivariate normality. The reference I was able to find is https://bookdown.org/ccolonescu/RPoE4/simultaneous-equations-models.html (except of course rm(list=ls()) for which the author's computer should be burned, per Jenny Bryan's guidance).

-- Stas Kolenikov, PhD, PStat (ASA, SSC) 
-- Principal Statistician, NORC @NORCnews
-- Opinions stated in this email are mine only, and do not reflect the position of my employer
-- Social media: @StatStas [ Twitter | mastodon.online ]
-- http://stas.kolenikov.name



Reply all
Reply to author
Forward
0 new messages