I was trying to fit a multilevel model, but I discovered that my dependent variable is highly skewed and zero-inflated. Individuals report 5 times a day for 7 days their level of paranoia and the perception of the other as dominant in interpersonal exchanges. We also collected pathological traits in the data (es. BPD). In 2000 observations, individuals rated their paranoia "zero" 1500 times.

I tried to run a model taking into account the nature of the data


  • gamma model

    model.1a <- "StatePara1 ~ day + STPD.c + (1 + day|mail)" fit.1a <- glmer(model.1a, data = merged_data_sb, family = Gamma(link = "inverse"), control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 2e5)))

or also


  • mixed logit model

    zero_model <- glmer(StatePara == 0 ~ 1 + BPD.c + day + (1+day|mail), data = merged_data_sb, family = binomial)

  • Filter data for non-zero values

    non_zero_data <- merged_data_sb[merged_data_sb$StatePara > 0, ]

  • Second part: Mixed range model for non-zero values

    gamma_model <- glmmTMB(StatePara ~ 1 + BPD.c + day + (1 + day|mail), data = non_zero_data, family = Gamma(link = "log"))

  • Combine the two models into a two-part model

    two_part_model <- list(zero_model = zero_model, gamma_model = gamma_model)

The models are not working properly (they run, but every model is not significant, which is very unlikely since BPD and other pathological traits have paranoia as a core symptom)

How can I manage this data?

This is the distribution

This is the distribution of the dependent variable

New contributor
miso is a new contributor to this site. Take care in asking for clarification, commenting, and answering. Check out our Code of Conduct.
  • 1
    $\begingroup$ Were4 the ratings of paranoia really continuous? Or were they on a Likert scale or something else? $\endgroup$
    – Peter Flom
    Commented Jun 11 at 11:04
  • $\begingroup$ On a likert scale form 1 to 10. The question was: "For each of the feelings and thoughts described below, please indicate how much you have experienced them with the person you have identified. 0 = not at all - 10 = very much." $\endgroup$
    – miso
    Commented Jun 11 at 12:41
  • $\begingroup$ You indicate it is not count data but continuous data in the title, but in the comment your say that it comes from a Likert scale, which is not really continuous data? $\endgroup$ Commented Jun 11 at 12:59
  • $\begingroup$ Then you might consider ordinal logistic regression. $\endgroup$
    – Peter Flom
    Commented Jun 11 at 17:11


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.