Help with Linear Mixed Models and ESM-data

[Naomi Mueller]

Dear all,

This is my first time ever on a forum, but I really really need your help. I am currently writing my masterthesis, and my supervisors don't respond to me. After 5 years of using and learning how to use SPSS, I was forced to use Stata 13 at my latest internship.

To gain insight into the daily flow of self-esteem, dissociative thoughts, preoccupation, paranoia, visual and auditory hallucinations and a general symptom of psychosis in one's life, a research method is necessary that enables the patient to reflect upon moment-to-moment experiences. Not only because most experiences are internal mental phenomena, but also because some authors suggest that patients suffering from non-affective psychotic disorder are less able to reveal their general experiences and state of mind due to lack of insight (Schwarz, 1998; Dȩbowska, Grzywa & Kucharska-Pietura, 1998). The Experience Sampling Method (ESM; Csikszentmihalyi & Larson, 1987) is the best suitable method to do this. This structured self-assessment diary technique, which the participant will use for six consecutive days, gives a signal (a so-called 'beep') at ten semi-random time-points per day. When answered all the beeps, data-sampling can yield up to 60 beeps per participant. In addition to ease of use, it provides insight into the individual fluctuations of emotion and cognition. The ecological approach of ESM focusses on typical daily life experiences rather than set-up situations. The reliability and validity of ESM have been extensively demonstrated in previous studies with healthy participants (Jacobs et al., 2005), studies that involved schizophrenia patients (Delespaul, 1995; Myin-Germeys, van Os et al., 2001; Myin-Germeys & van Os, 2007), and studies involving other mental disorders (De Vries, 1992; Myin-Germeys et al., 2009).
The main theorem of this research is accompanied by two hypotheses and will be tested using data from the Genetic Risk and OUtcome of Psychosis-study (GROUP-study; Korver et al., 2012). This large multicenter study included patients with a non-affective psychotic disorder, their first-degree relatives and (unrelated) healthy controls without a family history of psychotic disorder:

“Is momentary self-esteem associated with specific psychotic symptoms (auditory hallucinations, visual hallucinations, preoccupation, dissociative thoughts, paranoia and one general symptom) in daily life and between groups (patients, relatives and controls)?”

Hypotheses:

1. Momentary self-esteem will be associated with specific psychotic symptoms in daily life within patients (147 participants), relatives (131 participants), and controls (114 participants).

1. The magnitude of the associations between momentary self-esteem and specific psychotic symptoms in daily life will be higher in patients can in relatives and controls, and more significant in relatives than in controls.

Specific psychotic symptoms will be measured with the following six statements, to be answered on a 7-point Likert-scale.

• 'My thoughts are influenced by others.’ corresponding to dissociative thoughts.
• 'I cannot get these thoughts out of my head.' corresponding to preoccupation, thus (control)delusion.
• 'I feel suspicious.' corresponding to paranoia.
• 'I hear voices.’ corresponding to the auditory hallucination.
• 'I see things that are not really there.' corresponding to visual hallucination.
• 'I am afraid I will lose control.' corresponding to a general symptom of psychosis.

Self-esteem will be measured with the following statements, to be answered on a 7-point Likert-scale.

• 'I like myself.'
• 'I doubt myself.' (reversed for analysis)

Linear mixed models will be fitted in Stata 13.0 (StataCorp, 2014) (with time points nested within participants and participants nested within families) to quantify associations between momentary self-esteem (all measured with the ESM) as the independent variable, and a) dissociative thoughts, delusion, paranoia, visual hallucinations, auditory hallucinations and a general symptom of psychosis (measured with the ESM) as the dependent variables within patients, first-degree relatives, and controls while controlling for potential confounders (age, gender, ethnicity, education level, and marital status) (H1). Then two-way interactions for self-esteem × group (patients, first-degree relatives, and controls) will be added to examine whether the magnitude of associations of momentary self-esteem in daily life with a) dissociative thoughts, delusion, paranoia, visual hallucinations, auditory hallucinations and a general symptom of psychosis is more significant in patients than in controls and first-degree relatives than in controls. A Wald test will be used to examine interaction effects as well as the ‘lincom’ command to compute linear combinations of coefficients for testing H2.


The first table was created with the following Stata command:

*MOMENTARY SELF-ESTEEM BY GROUP (TEST GROUP DIFFERENCES)
xi: mixed se_mean i.mf_status || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

* ^^ MF1 relatives vs controles, MF 2 = patients vs controls. Lincom is patients vs. Relatives.

* Psychotic experiences by group (test group differences)
xi: mixed pe_mean i.mf_status || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

* Psychotic experiences by group (test group differences)
xi: mixed thou_concent i.mf_status || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

* Psychotic experiences by group (test group differences)
xi: mixed pat_th_othrs i.mf_status || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

This was repeated for all 6 symptoms + psychotic experience + self-esteem.

(tabel 2a) VARIABLES BY GROUP - ADJUSTED FOR AGE, GENDER, ETHNICITY, EDUCATION, MARITAL STATUS

xi: mixed se_mean i.mf_status mf_age mf_gender mf_ethn2 i.mf_educ_hml i.mf_marstat || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

xi: mixed pe_mean i.mf_status mf_age mf_gender mf_ethn2 i.mf_educ_hml i.mf_marstat || wb_famid2: || patientid:, reml
*patients vs. relatives
lincom (_Imf_status_2) - (_Imf_status_1)

This was also repeated for all the symptoms + pe + se.

* Tabel 2 left partl:
by mf_status, sort : summarize se_mean
by mf_status, sort : summarize pe_mean

(Also repeated.)

Secondly, I had to run another analysis to test for confounders.

Stata command I used was:

Table 3: Momentary self-esteem and psychotic experiences by groups

*formal test of interaction: momentary self-esteem and psychotic experiences by group - ADJUSTED FOR CONFOUNDERS
xi: mixed pe_mean c.se_mean##i.mf_status mf_age mf_gender mf_ethn2 i.mf_educ_hml i.mf_marstat || wb_famid2: || patientid:, reml
testparm 1.mf_status#c.se_mean 2.mf_status#c.se_mean

*in controls:
lincom c.se_mean

*in relatives
lincom c.se_mean + 1.mf_status#c.se_mean

*in patients
lincom c.se_mean + 2.mf_status#c.se_mean

*patients vs. relatives
lincom (c.se_mean + 2.mf_status#c.se_mean)-(c.se_mean + 1.mf_status#c.se_mean)

*patients vs. controls
lincom (c.se_mean + 2.mf_status#c.se_mean)-c.se_mean

*relatives vs. controls
lincom (c.se_mean + 1.mf_status#c.se_mean)-c.se_mean

The table this generated:

I realise that it is a lot of information, but my question is:
1. Why are my values in the first table almost all positive, while they are almost all negative in the second table.
2. Am I doing something wrong? Because I did not change anything in the model/command, just copied the results.

I really hope there is someone out there who wants to help me because I am at a loss with my very little knowledge of statistics (my supervisor prepared the analysis that I ran).

THANK YOU in advance,

Naomi Müller

[Clyde Schechter]

You didn't change anything in the regressions, but you ran different -lincom- commands for the two tables and you calculated very different things.

In Table 2 you calculated mean values of the self esteem ratings and psychotic symptom ratings in the different subject groups, along with a few differences between them.

In Table 3 you calculated the mean values of how muich each of those psychotic symptom ratings changes in association with a unit change in se_mean in each subject group. and then the differences between the groups. The negative values in Table 3 tell you that we see less psychotic experiences, dissociative thoughts, etc., at times when self_esteem is higher, which is consistent with your first hypothesis. For future reference, these changes in outcome variables associated with a unit change in a predictor variable are called "marginal effects" of the predictor on the outcome.

It does not appear you are doing anything wrong, although I have not scrutinized every line of the code you show.

The switch from SPSS to Stata is not an easy one to make: the programming languages are rather different and the learning curve can be steep at first if you have not used a Stata-like language previously. I sympathize.

[Naomi Mueller]

Dear Clyde,

Thank you already for responding.
If I'm correct, the first table gives me the means of the variables and the interaction between the three groups. So, for momentary self-esteem and the interaction between patients and controls is -0.689, and means that the momentary self-esteem of patients is -0.689 lower than the means of the controls, right?
And for the second table, I should review this apart from the first table, correct?

But I still don't quite understand how to exactly interpret the results from the third table. Would you mind helping me with that by a short example from the results above?

Naomi

[Clyde Schechter]

If I'm correct, the first table gives me the means of the variables and the interaction between the three groups. So, for momentary self-esteem and the interaction between patients and controls is -0.689, and means that the momentary self-esteem of patients is -0.689 lower than the means of the controls, right?

Almost right. The mean self esteem score in patients is 5.22, and in controls it's 5.90. So it's 0.689 lower in patients than in relatives. Not -0.689 lower in patients than controls. -0.689 lower would be the same thing as 0.689 higher.

And for the second table, I should review this apart from the first table, correct?

Yes. It's an analysis of something entirely different and there is no reason to expect the numbers in this table to resemble those in the other table.

But I still don't quite understand how to exactly interpret the results from the third table. Would you mind helping me with that by a short example from the results above?

The table as a whole is about the effect of self esteem on all those symptoms. Let's look at the top left cell in Table 3. This cell is about psychotic experiences in patients. The number in this cell, -.059 means that given two patients whose self esteem scores differ by 1, on average the one with the higher self esteem will have a psychotic experiences score that is 0.059 lower than the other patient. Otherwise put, among patients. a difference of 1 unit in self esteem is associated with a 0.60 difference in the opposite direction (negative sign!), on average, in psychotic experience score. Put still more simply, one can say that the marginal effect (or some people just say effect) of a unit difference in self-esteem score on psychotic experience score is -.059.

Moving a couple of columns over, we see that among healthy controls, given two such people whose self esteem scores differ by 1, on average the one with the higher self esteem score will have a psychotic experiences score that is 0.001 higher than the other healthy control. Looking now at the bottom half of the table, the first column is about differences between patients and healthy controls. And in the first row there, which is, again, about psychotic experiences, it tells us that the effect of a 1 unit difference in self-esteem score among patients is -0.60 (= -.059 - .001) greater (or, in more ordinary English, 0.60 less) than it is among controls.

[Naomi Mueller]

Thank you so much already! I will try to go and finish it, is it okay if I later turn out to have some more questions?

[Clyde Schechter]

This forum is open to the public. As far as I know the only people who have been blocked are people who spammed or posted abusive content. You are always welcome to post questions here. If somebody here finds them of interest, you will get answers. That's how it always works here.

If you do post new questions, give some thought to whether to add them to this thread or start a new topic. It is important to keep threads on topic. While it is easy to imagine these threads as a dialog between a questioner and a responder, they are more than that. There are people who come and search the forum for help on specific topics. And other people who browse regularly may choose which posts to read based on the title of the thread. So to make sure that content is easily accessible to those people, and that they don't waste their time on posts that are irrelevant to them, it is important that thread titles be descriptive of the actual content and that they remain so. So if you have questions about something that is a different statistical issue, even though it is part of your same master's thesis project, do start a new thread. If you have questions that are along the same lines as you have already asked here, add them to this one.