Skip to contents

InDetail4-ComparisonFunctions

Source: vignettes/web_only/InDetail4-ComparisonFunctions.Rmd
InDetail4-ComparisonFunctions.Rmd

Introduction

When analysing responses to open-ended questions, you may want to look into whether different groups of survey participants have, in general, responded differently to the prompt. finnsurveytext contains a number of comparison functions which are intended to be used to compare responses between groups. These comparison functions covered are defined in r/04_comparison_functions.R and r/05_comparison_concept_network.R.

One way to split the data is using a different question within the survey such as a categorical question (eg. gender, location, or level of education) or an ordinal variable (such as age or income bracket). In this tutorial, we will look at comparing responses to a question based on gender. Before using the comparison functions, we run the preparation functions which are in r/01_prepare_conll-u.R (and which are covered in detail in ‘Tutorial1-Prepare CoNLL-U’) for each group separately.

The comparison functions are:

  1. fst_summarise_compare()
  2. fst_pos_compare()
  3. fst_length_compare()
  4. fst_comparisoncloud()
  5. fst_get_unique_ngrams()
  6. fst_join_unique()
  7. fst_ngrams_compare_plot()
  8. fst_ngrams_compare_plot2()
  9. fst_plot_multiple()
  10. fst_freq_compare()
  11. fst_ngrams_compare()
  12. fst_cn_get_unique()
  13. fst_cn_compare_plot()
  14. fst_concept_network_compare()

We will look at the following question:

  • q11_2 Jatka lausetta: Kehitysyhteistyö on toimintaa, jossa… (Avokysymys)
    • (In English) q11_2 Continue the sentence: Development cooperation is an activity in which… (Open question)

Installation of package.

Once the package is installed, you can load the finnsurveytext package as below: (Other required packages such as dplyr and stringr will also be installed if they are not currently installed in your environment.)

The Data

These files have been created following the process demonstrated in Tutorial 1.

1. Child Barometer Data
  • data/fst_child.rda
2. Development Cooperation Data
  • data/fst_dev_coop

You can read these in as follows:

df1 <- fst_child
df2 <- fst_dev_coop

Comparison of responses based on gender

All of our comparison functions can compare between groups of respondents based on another field in the raw data which we have included when formatting. The following common arguments appear in multiple functions and are defined the same way in each:

  • field is the field in the data used to split the responses.
  • exclude_nulls (which has a default value of FALSE) is used to indicate whether we want to include a group of repondents which have no value in the field column.
  • rename_nulls (with default value of ‘null_data’) is what to name the group with the missing field data.

First, we will look at some summary functions to compare the responses overall.

The functions are:

  1. fst_summarise_compare()
  2. fst_pos_summary()
  3. fst_length_compare()

Make Comparison Summary

We can run this function to create a summary table for our data: (we use knitr::kable function below to display results in a “prettier” table)

knitr::kable(
  fst_summarise_compare(data = df2,
                        field = 'gender',
                        exclude_nulls = FALSE,
                        rename_nulls = 'no gender provided'
  )
)
Description Respondents No Response Proportion Total Words Unique Words Unique Lemmas
Female 673 13 0.98 2993 823 722
Male 183 8 0.96 795 383 354
no gender provided 89 4 0.96 404 225 208

Remarks:

We can already see that our data is quite unbalanced. There are a lot more female respondents than male or unspecified. The response (to this question) rate is high (97%) but slightly lower for the unspecified respondents. Unsurprisingly, since there are more female responses, the female responses to this question contain a larger variety of words.

POS and Length Comparisons

Next, we will look at part-of-speech tags and lengths of responses.

knitr::kable(
  fst_pos_compare(data = df2,
                  field = 'gender',
                  exclude_nulls = TRUE
  )
)
UPOS Part_of_Speech_Name Female-Count Female-Prop Male-Count Male-Prop
ADJ adjective 276 0.092 69 0.087
ADP adposition 19 0.006 3 0.004
ADV adverb 44 0.015 11 0.014
AUX auxiliary NA NA 3 0.004
CCONJ coordinating conjunction 2 0.001 NA NA
DET determiner 24 0.008 4 0.005
INTJ interjection 1 0.000 NA NA
NOUN noun 2373 0.793 636 0.800
NUM numeral 1 0.000 1 0.001
PART particle 18 0.006 8 0.010
PRON pronoun 8 0.003 3 0.004
PROPN proper noun 19 0.006 12 0.015
PUNCT punctuation NA NA NA NA
SCONJ subordinating conjunction NA NA NA NA
SYM symbol 1 0.000 NA NA
VERB verb 195 0.065 45 0.057
X other 12 0.004 NA NA 

knitr::kable(
  fst_length_compare(data = df2,
                     field = 'gender',
                     incl_sentences = TRUE,
                     exclude_nulls = TRUE
  )
)
Description Respondents Mean Minimum Q1 Median Q3 Maximum
Female- Words 660 5.518 1 4 5 6 28
Female- Sentences 660 1.012 1 1 1 1 3
Male- Words 175 5.417 2 4 5 6 32
Male- Sentences 175 1.029 1 1 1 1 3

Remarks:

In terms of POS tags, the scale differences are likely mostly due to the differences in the number of respondents between genders. We can also see that female responses are generally slightly longer (average of 6 words to 5 words) but that most respondents (across the genders) wrote only a single sentence.

Comparison Cloud

Now that we’ve looked at the overview comparisons, we will create a comparison cloud between the responses, using the versions with stopwords removed so that only more meaningful words remain.

A comparison cloud compares the relative frequency with which a term is used in two or more documents. This cloud shows words that occur more regularly in responses from a specific type of respondent. For more information about comparison clouds, you can read this documentation.

We create our comparison cloud as follows:

# WEIGHTED
fst_comparison_cloud(data = df2,
                     field = 'gender',
                     pos_filter = NULL,
                     norm = NULL,
                     max = 100,
                     use_svydesign_weights = FALSE,
                     use_svydesign_field = FALSE,
                     id = "",
                     svydesign = NULL,
                     use_column_weights = TRUE,
                     exclude_nulls = TRUE
)

To run fst_comparison_cloud(), we provide the following arguments to the function:

  • data, field, exclude_nulls and `rename_nulls`` (as defined above)
  • norm is the method for normalising the data. Valid settings are 'number_words' (the number of words in the responses), 'number_resp' (the number of responses), or NULL (raw count returned, default).
  • pos_filter is an optional list of which POS tags to include such as 'c("NOUN", "VERB", "ADJ", "ADV")'. The default is NULL, in which case all words in the data are considered.
  • max is the maximum number of words to display, the default is 100.
  • use_svydesign_weights is a boolean for whether to get weights for the responses from a svydesign object
  • If weights are coming from a svydesign object, the id field needs to not be empty, as this is used to join the data.
  • Similarly, if weights are coming from a svydesign object this is the named object.
  • use_column_weights is a boolean for if weights have already been included in the formatted data and should be included.

Common Words and N-grams

Now, we will look at common words occurring in the responses.

First, we will consider all the responses for this question. We are not filtering the data based on POS tag, and will leave the default of strict = TRUE which will cut-off the list at 10 words (see the warning note about this). We also use the default for the norm which means we are standardising between groups by dividing count of a word by the total number of words in the responses.

For definition of fst_freq_table() and fst_ngrams_table() functions, see “InDetail2-DataExploration”.

all_top10 <- fst_freq_table(df2,
  number = 10,
  norm = "number_words",
  pos_filter = NULL,
  strict = FALSE
)
#> Note:
#>  Words with equal occurrence are presented in alphabetical order. 
#>  With `strict` = FALSE, words occurring equally often as the `number` cutoff word will be displayed.
all_top10bigrams <- fst_ngrams_table(df2,
  number = 10,
  ngrams = 2,
  norm = "number_words",
  pos_filter = NULL,
  strict = TRUE
)
#> Note:
#>  N-grams with equal occurrence are presented in alphabetical order. 
#>  By default, n-grams are presented in order to the `number` cutoff n-gram. 
#>  This means that equally-occurring later-alphabetically n-grams beyond the cutoff n-gram will not be displayed.

knitr::kable(all_top10)
words occurrence
köyhyys 0.062
nälänhätä 0.057
sota 0.055
ilmastonmuutos 0.034
puute 0.028
ihminen 0.025
vesi 0.023
epätasa-arvo 0.021
ahneus 0.020
nälkä 0.019 
knitr::kable(all_top10bigrams)
words occurrence
puhdas vesi 0.015
vesi puute 0.013
nälänhätä sota 0.010
köyhyys nälänhätä 0.009
sota nälänhätä 0.008
epätasainen jakautuminen 0.007
sota köyhyys 0.007
köyhyys sota 0.005
nälänhätä köyhyys 0.005
ilmastonmuutos köyhyys 0.004

Comparison N-Gram Plots

Now we will look at top words by gender. There are two functions which create the plots in one function simply. These are fst_freq_compare() and fst_ngrams_compare().

fst_freq_compare(
  df2,
  field = 'gender',
  number = 10,
  norm = NULL,
  pos_filter = NULL,
  unique_colour = "indianred",
  strict = TRUE, 
  exclude_nulls = TRUE, 
  use_column_weights = TRUE
)


fst_ngrams_compare(
  df2,
  field = 'gender',
  number = 10,
  ngrams = 2,
  norm = NULL,
  pos_filter = NULL,
  unique_colour = "indianred",
  strict = TRUE, 
  use_column_weights = TRUE, 
  exclude_nulls = TRUE
)

The functions fst_freq_compare() and fst_ngrams_compare() have parameters as defined previously.

Remarks: Note that words that are unique to a gender are highlighted in red. This is only comparing the top 10 words, so be aware this word likely still appears in the other genders, just less frequently! Interestingly, while they have the same top 4 words, it seems that the top 3 words differs across the genders with ‘ilmastonmuutos’ being the 3rd most frequent in male responses but ‘nälänhätä’ is in the top 3 for females. The most frequent bigram in female responses ‘puhdas vesi’ is not commonly mentioned by male respondents.

Comparison Concept Network

Since the top 4 words are ‘köyhyys’, ‘nälänhätä’, ‘sota’, and ‘ilmastonmuutos’, we will create Concept Networks based on these terms.

We run the comparison Concept Network as below. Again, we’re keeping the default norm.

fst_concept_network_compare(data = df2,
                            concepts = 'köyhyys, nälänhätä, sota, ilmastonmuutos',
                            field = 'gender',
                            norm = NULL,
                            threshold = NULL,
                            pos_filter = NULL,
                            exclude_nulls = TRUE,
                            )

The parameters are the same as previously described.

Remarks:

Each of these Concept Networks are created separately, which means that the weights of words are based only on the responses within that gender. Despite this, we have many more words in the female plot (possibly due to there being more responses in this data leading to increased variation).

Conclusion

As demonstrated in this tutorial, finnsurveytext contains a number of functions which can be used to compare responses between groups. This analysis could form the start of further work into the responses to question 11_3 by respondents based on their gender, or as just one of many different splits used to investigate the responses.

Citation

Finnish Children and Youth Foundation: Young People’s Views on Development Cooperation 2012 [dataset]. Version 2.0 (2019-01-22). Finnish Social Science Data Archive [distributor]. http://urn.fi/urn:nbn:fi:fsd:T-FSD2821