Error using fitrm function classreg.r​egr.Formul​aProcessor​>parseStr

I have a dataset with 20 individuals (SubjectID 1-20). These individuals were separated in three groups (n=6, n=6, and n=8) and each group was testet(measured) at a specific time point (T1, T5, T10). Each measurement consisted of 4 consecutive time points (Minute 0, 5, 15, 60, time after exposure). So I have a total of 20 subject with 4 individual measurements and divided into three groups. Using fitrm I want to fit a repeated measurement model with my readout parameter x to look at interaction of T1,T2,T3 with time after exposure. I defined the modelspec as:

modelspec = 'X ~ Timepoint * TimeafterExp + (1|SubjectID)';

Data is a 80x4 table with the variables SubjectID, Timepoint, TimeafterExp, X.

When I run the following code:

rm=fitrm(data, modelspec);

I get the following error:

Error using classreg.regr.FormulaProcessor>parseStr

Unable to understand the character vector or string scalar 'X ~ Timepoint * TimeafterExp + (1|SubjectID)'.

Error in classreg.regr.FormulaProcessor (line 374)

[f.str,f.tree] = parseStr(f,modelSpec);

Error in classreg.regr.MultivariateLinearFormula (line 46)

f = [e-mail](varargin{:});

Error in RepeatedMeasuresModel.fit (line 1314)

formula = classreg.regr.MultivariateLinearFormula(model,varNames);

Error in fitrm (line 77)

s = RepeatedMeasuresModel.fit(ds,model,varargin{:});

Error in untitled2 (line 25)

rm=fitrm(data, modelspec);

I appreciate your help!

[the cyclist]

Can you upload the data? You can use the paper clip icon in the INSERT section of the toolbar.

[David Obert]

Thank you for your comment. I added the data file.

Answer 1

[the cyclist]

I agree with @Harald, that a linear mixed effects model is likely a good choice here:

load("data.mat","data")
modelspec = 'X ~ Timepoint * TimeafterExp + (1|SubjectID)';
mdl=fitlme(data, modelspec)

mdl = 
Linear mixed-effects model fit by ML

Model information:
    Number of observations              80
    Fixed effects coefficients           4
    Random effects coefficients         20
    Covariance parameters                2

Formula:
    X ~ 1 + Timepoint*TimeafterExp + (1 | SubjectID)

Model fit statistics:
    AIC       BIC       LogLikelihood    Deviance
    734.32    748.61    -361.16          722.32  

Fixed effects coefficients (95% CIs):
    Name                              Estimate     SE          tStat       DF    pValue        Lower        Upper 
    {'(Intercept)'           }           68.273      5.9383      11.497    76    2.6629e-18       56.446    80.101
    {'Timepoint'             }          0.15558     0.85891     0.18113    76       0.85675      -1.5551    1.8662
    {'TimeafterExp'          }        -0.049118     0.19141    -0.25661    76       0.79817     -0.43034    0.3321
    {'Timepoint:TimeafterExp'}         0.019561    0.027685     0.70654    76       0.48201    -0.035579    0.0747

Random effects covariance parameters (95% CIs):
Group: SubjectID (20 Levels)
    Name1                  Name2                  Type           Estimate    Lower    Upper
    {'(Intercept)'}        {'(Intercept)'}        {'std'}        0           NaN      NaN  

Group: Error
    Name               Estimate    Lower     Upper 
    {'Res Std'}        22.1        18.927    25.803

If you want to use the repeated measures model, I believe you need to "unstack" those measures into their own columns, and use the "y1-y4 ~ ..." form of the model spec.

% Unstack the time after exposure data
data = unstack(data,'X','TimeafterExp');

Warning: Table variable names that were not valid MATLAB identifiers have been modified. Since table variable names must be unique, any table variable names that happened to match the new identifiers also have been modified.
To use the original INDVAR values as table variable names, set 'VariableNamingRule' to 'preserve'.

% Rename the columns
data = renamevars(data,["x0","x5","x15","x60"],["y1","y2","y3","y4"])

data = 20×6 table
    SubjectID    Timepoint    y1     y2    y3    y4
    _________    _________    ___    __    __    __

        1            1         71    36    70    75
        2            1         87    46    72    79
        3            1         90    57    58    41
        4            1         90    39    55    64
        5            1         96    33    78    83
        6            1         91    45    72    46
        7            5         93    35    77    71
        8            5        101    33    53    83
        9            5        101    69    62    82
       10            5         86    59    63    74
       11            5        100    43    79    93
       12            5        110    58    63    90
       13           10         98    28    74    74
       14           10         81    52    63    82
       15           10         92    60    76    80
       16           10        117    31    70    91

% Define the new modelspec
modelspec = 'y1-y4 ~ Timepoint';
rm = fitrm(data, modelspec)

rm = 
  RepeatedMeasuresModel with properties:

   Between Subjects:
         BetweenDesign: [20×6 table]
         ResponseNames: {'y1'  'y2'  'y3'  'y4'}
    BetweenFactorNames: {'Timepoint'}
          BetweenModel: '1 + Timepoint'

   Within Subjects:
          WithinDesign: [4×1 table]
     WithinFactorNames: {'Time'}
           WithinModel: 'separatemeans'

   Estimates:
          Coefficients: [2×4 table]
            Covariance: [4×4 table]

Disclaimer: I got that model to run, but I am not certain it is the model you intend.

Also, I think in the repeated measures model you lose the information that your measurements come at different time spacings. You keep that info in the mixed effects model.

I hope that helps!

[David Obert]

Thank you for your help. I appreciate it! I needed the repeated measure model to test my data for sphericity. The only test matlab provides for this is the mauchly function, which needs a rm structure as input (tbl = mauchly(rm)).

Answer 2

[Harald]

Hi,

I am not an expert here but will give it a try.

In the "More About" section of the documentation of fitrm, there is a section on Wilkinson Notation and it does not go into the use of |. Thus I suppose it is not supported for this type of model.

https://www.mathworks.com/help/stats/wilkinson-notation.html states that the | syntax is supported by random-effects and mixed-effects models. For example, fitlme is such a model and thus in its documentation refers to the | syntax. Also, the following works:

mdl=fitlme(data, modelspec);

Does this provide you the intended type of model?

Best wishes,

Harald