In StatisMike/stenR: Standardization of Raw Discrete Questionnaire Scores

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Introduction

stenR main focus in on standardization and normalization of raw scores of questionnaire or survey on basis of Classical Test Theorem.

Particularly in psychology and other social studies it is very common to not interpret the raw results of measurement in individual context. In actuality, it would be usually a mistake to do so. Instead, there is a need to evaluate the score of single questionee on the basis of larger sample. It can be done by finding the place of every individual raw score in the distribution of representative sample. One can refer to this process as normalization. Additional step in this phase would be to standardize the data even further: from the quantile to fitting standard scale.

It need to be noted that rarely one answer for one question (or item) is enough to measure a latent variable. Almost always there is a need to construct scale or factor of similar items to gather a behavioral sample. This vital preprocessing phase of transforming the item-level raw scores to scale-level can be also handled by functions available in this package, though this feature is not the main focus.

Factor analysis and actual construction of scales or factors is beyond the scope of this package. There are multiple useful and solid tools available for this. Look upon psych and/or lavaan for these features.

This journey from raw, questionnaire data to normalized and standardized results will be presented in this vignette.

Raw questionnaire data preprocessing

We will work on the dataset available in this package: SLCS. It contains answers of 103 people to the Polish version of Self-Liking Self-Competence Scale.

library(stenR)
str(SLCS)

As can be seen above, it contains some demographical data and each questionee answers to 16 diagnostic items.

Authors of the measure have prepared instructions for calculating the scores for two subscales (Self-Liking and Self-Competence). General Score is, actually, just sum of the subscale scores.

Self-Liking: 1R, 3, 5, 6R, 7R, 9, 11, 15R
Self-Competence: 2, 4, 8R, 10R, 12, 13R, 14, 16

Items numbers suffixed with R means, that this particular item need to be reversed before summarizing with the rest of them to calculate the raw score for a subscale. That's because during the measure construction, the answers to these items were negatively correlated with the whole scale.

All of this steps can be achieved using the item-preprocessing functions from stenR.

Firstly, you need to create scale specification objects that refer to the items in the available data by their name. It need to also list the items that need reversing (if any) and declare NA insertion strategies (by default: no insertion).

Absolute minimum and maximum score for each item need to be also provided on this step. It allows correct computation even if the absolute values are not actually available in the data that will be summed into factor. This situation should not happen during first computation of the score table on full representative sample, but it is very likely to happen when summarizing scores for only few observations.

# create ScaleSpec objects for sub-scales
SL_spec <- ScaleSpec(
  name = "Self-Liking",
  item_names = c("SLCS_1", "SLCS_3", "SLCS_5", "SLCS_6", "SLCS_7", 
                 "SLCS_9", "SLCS_11", "SLCS_15"),
  min = 1,
  max = 5,
  reverse = c("SLCS_1", "SLCS_6", "SLCS_7", "SLCS_15")
)

SC_spec <- ScaleSpec(
  name = "Self-Competence",
  item_names = c("SLCS_2", "SLCS_4", "SLCS_8", "SLCS_10", "SLCS_12",
                 "SLCS_13", "SLCS_14", "SLCS_16"),
  min = 1,
  max = 5,
  reverse = c("SLCS_8", "SLCS_10", "SLCS_13")
)

# create CombScaleSpec object for general scale using single-scale 
# specification
GS_spec <- CombScaleSpec(
  name = "General Score",
  SL_spec,
  SC_spec
)

print(SL_spec)
print(SC_spec)
print(GS_spec)

After the scale specification objects have been created, we can finally transform our item-level raw scores to scale-level ones using sum_items_to_scale() function.

Each ScaleSpec or CombScaleSpec object provided during its call will be used to create one variable, taking into account items that need reversing (or sub-scales in case of CombScaleSpec), as well as NA imputation strategies chosen for each of the scales.

By default only these columns will be available in the resulting data.frame, but by specifying the retain argument we can control that.

summed_data <- sum_items_to_scale(
  data = SLCS,
  SL_spec,
  SC_spec,
  GS_spec,
  retain = c("user_id", "sex")
)

str(summed_data)

At this point we successfully prepared our data: it now describes the latent variables that we actually wanted to measure, not individual items. All is in place for next step: results normalization and standardization.

Both ScaleSpec and CombScaleSpec objects have their specific print and summary methods defined.

Normalize and standardize the results

We will take a brief look at the procedural workflow of normalization and standardization. It should be noted, that it is more verbose and have less features than the object-oriented workflow. Nevertheless, it is recommended for useRs that don't have much experience utilizing R6 classes. For more information about both, read Procedural and Object-oriented workflows of stenR vignette.

To process the data, stenR need to compute the object of class ScoreTable. It is very similar to the regular score tables that can be seen in many measures documentations, though it is computed directly on the basis of available raw scores from representative sample. After that first, initial construction it can be reused on new observations.

This is a two step process. Firstly, we need to compute a FrequencyTable object that is void of any standard score scale for every sub-scale and scale.

# Create the FrequencyTables
SL_ft <- FrequencyTable(summed_data$`Self-Liking`)
SC_ft <- FrequencyTable(summed_data$`Self-Competence`)
GS_ft <- FrequencyTable(summed_data$`General Score`)

There were some warnings printed out there: they are generated if there were any raw score values that were missing in-between actual minimal and maximal values of raw scores. By the rule of the thumb - the wider the raw score range and the smaller and less-representative the sample is, the bigger possibility for this to happen. It is recommended to try and gather bigger sample if this happens - unless you are sure that it is representative enough.

After they are defined, they can be transformed into ScoreTable objects by providing them some StandardScale object. Objects for some of more popular scales in psychology are already defined - we will use very commonly utilized Standard Ten Scale: STEN

# Check what is the STEN *StandardScale* definition
print(STEN)

# Calculate the ScoreTables
SL_st <- ScoreTable(SL_ft, STEN)
SC_st <- ScoreTable(SC_ft, STEN)
GS_st <- ScoreTable(GS_ft, STEN)

At this point, the last thing that remains is to normalize the scores. It can be done using normalize_score() or normalize_scores_df() functions.

# normalize each of the scores in one call
normalized_at_once <- normalize_scores_df(
  summed_data,
  vars = c("Self-Liking", "Self-Competence", "General Score"),
  SL_st,
  SC_st,
  GS_st,
  what = "sten",
  retain = c("user_id", "sex")
)

str(normalized_at_once)

# or normalize scores individually
SL_sten <- 
  normalize_score(summed_data$`Self-Liking`,
                  table = SL_st,
                  what = "sten")

SC_sten <- 
  normalize_score(summed_data$`Self-Competence`,
                  table = SC_st,
                  what = "sten")

GC_sten <- 
  normalize_score(summed_data$`General Score`,
                  table = GS_st,
                  what = "sten")

# check the structure
str(list(SL_sten, SC_sten, GC_sten))