README.md
In malsch/occupationCoding: Supervised Learning for Occupation Coding
Overview
This is an R package for occupation coding. It accompanies the following publication:
Malte Schierholz, Matthias Schonlau (2021): Machine Learning for Occupation Coding — a Comparison Study. Journal of Survey Statistics and Methodology 9(5), https://doi.org/10.1093/jssam/smaa023
Two groups of users will profit from this package:
- Applied users, who have German survey answers and want to run machine learning for automated occupation coding
- Programmers, who want to compare various machine learning algorithms for occupation coding and set up a coding pipeline that works best for them
Installation
The easiest way to install this package.
install.packages("remotes")
remotes::install_github("malsch/occupationCoding")
Applied German users will also want to download the file Gesamtberufsliste_der_BA.xlsx from https://download-portal.arbeitsagentur.de/ (requires free registration).
Usage
For Applied Users (German only, coding according to the 2010 German Classification of Occupation)
We will demonstrate how to code answers using the Alphabetical Dictionary (algorithm 1 in the paper) and using Similarity-based Reasoning (not included in the paper, due to space restrictions).
To prepare, load the coding index Gesamtberufsliste_der_BA.xlsx. Enter the path were you saved Gesamtberufsliste_der_BA.xlsx.
path_to_file <- "./Gesamtberufsliste_der_BA.xlsx" # change path
coding_index_w_codes <- prepare_German_coding_index_Gesamtberufsliste_der_BA(path_to_file,
count.categories = FALSE)
Imagine five persons from your survey answered as follows. We wish to code these answers according to the 2010 German Classification of Occupation.
text_input <- c("Bürokauffrau", "Stadtjugendpfleger", "Erzieherhelfer im Lehrlingswohnheim.",
"Mitarbeit bei einer Filmproduktion", "Abschleifer")
Coding with the Alphabetical Dictionary
Consult the coding index. For typical survey data, there is an entry in the coding index for about ~45% of all answers. In rare cases, especially when identical job titles are used in various industries, there can be more than one entry in the processed coding index.
(res <- predictWithCodingIndex(text_input, coding_index = coding_index_w_codes))
## id ans pred.code
## 1: 1 Bürokauffrau 71402
## 2: 2 Stadtjugendpfleger 83124
## 3: 3 Erzieherhelfer im Lehrlingswohnheim. <NA>
## 4: 4 Mitarbeit bei einer Filmproduktion <NA>
## 5: 5 Abschleifer 24222,21212
Similarity-based Reasoning
This was easy. Schierholz (2019) introduces an alternative method, which yields far better results. His technique relies on anonymized training data from various surveys (Antoni et al., 2010; Rohrbach-Schmidt and Hall, 2013; Lange et al., 2017; Hoffmann et al., 2018; Trappmann et al. 2010), which are provided as part of this package.
Training the model will take several minutes...
simBasedModelWordwise <- trainSimilarityBasedReasoning2(anonymized_data = surveyCountsWordwiseSimilarity,
num.allowed.codes = 1291,
coding_index_w_codes = coding_index_w_codes,
preprocessing = list(stopwords = NULL, stemming = NULL, strPreprocessing = TRUE, removePunct = FALSE),
dist.type = "wordwise",
dist.control = list(method = "osa", weight = c(d = 1, i = 1, s = 1, t = 1)),
threshold = c(max = NA, use = 1),
simulation.control = list(n.draws = 250, check.normality = FALSE)
)
simBasedModelSubstring <- trainSimilarityBasedReasoning2(anonymized_data = surveyCountsSubstringSimilarity,
num.allowed.codes = 1291,
coding_index_w_codes = coding_index_w_codes,
preprocessing = list(stopwords = NULL, stemming = NULL, strPreprocessing = TRUE, removePunct = FALSE),
dist.type = "substring",
dist.control = NA,
threshold = NA,
simulation.control = list(n.draws = 250, check.normality = FALSE)
)
Now predict possible codes
resWordwise <- predictSimilarityBasedReasoning(simBasedModelWordwise, text_input)
resSubstring <- predictSimilarityBasedReasoning(simBasedModelSubstring, text_input)
We actually trained two different models. We could use either one, but the results become better if we combine them. If the output from one model yields poor predictions, we simply use the other.
## first way of combining both models: use resWordwise if the five most likeliest categories from resWordwise have a chance of more than 50% to be correct, otherwise use resSubstring
wordwiseModelUsefulIds <- resWordwise[, .(prob = sum(head(.SD[order(pred.prob, decreasing = TRUE), pred.prob], 5))), by = id][prob > 0.5, id]
## second way of combining both models: select for each case the prediction method which returns highest probability (see function selectMaxProbMethod)
wordwiseProb <- resWordwise[, .SD[which.max(pred.prob), .(wordwiseProb = pred.prob)], by = id]$wordwiseProb
substringProb<- resSubstring[, .SD[which.max(pred.prob), .(substringProb = pred.prob)], by = id]$substringProb
wordwiseModelUsefulIds <- which(wordwiseProb > substringProb)
## use the second way of combining (it is the one performing better, shown below)
resCombined <- rbind(resWordwise[id %in% wordwiseModelUsefulIds], resSubstring[!(id %in% wordwiseModelUsefulIds)])[order(id)]
Let's look at the results again. The two most probable codes for each answer are shown next. The algorithm is rather convinced of its predictions for answers 1 and 3, it has some confidence in its predictions for answers 2 and 5, but for answer 4 even the most likeliest code has just a 7% chance to be correct.
resCombined[, head(.SD[order(pred.prob, decreasing = TRUE)], 2), by = id]
## id pred.code pred.prob ans
## 1: 1 71402 0.8928066326 Bürokauffrau
## 2: 1 72213 0.0200043173 Bürokauffrau
## 3: 2 83124 0.6698570028 Stadtjugendpfleger
## 4: 2 -9999 0.0002560075 Stadtjugendpfleger
## 5: 3 83112 0.8308403476 Erzieherhelfer im Lehrlingswohnheim.
## 6: 3 83132 0.0237893319 Erzieherhelfer im Lehrlingswohnheim.
## 7: 4 43423 0.0744033856 Mitarbeit bei einer Filmproduktion
## 8: 4 71393 0.0488194384 Mitarbeit bei einer Filmproduktion
## 9: 5 24222 0.5933915448 Abschleifer
##10: 5 -9999 0.0003152365 Abschleifer
## Note that the code `-9999` has a special meaning: It is always predicted with very low pred.prob and is a placeholder for all codes that have no evidence of being appropriate. As an example, consider the second answer: There are 1291 allowed codes in the target classification. Code 83124 has probability 0.67 and every other code has probability 0.000256, for a total probability of 0.67 + 1290 * 0.000256 = 1.
There are two main use cases for this algorithm: (semi-)automated coding and computer-assisted coding.
With (semi-)automated coding, the computer automatically selects the code that is most likely (at least if pred.prob is above a user-defined threshold). Semi-automated can be done es follows.
threshold <- 0.8 # only for demonstration purposes
resCombined[, .SD[which.max(pred.prob), list(ans, predicted = ifelse(pred.prob > threshold, pred.code, "no prediction"), pred.prob)], by = id]
With computer-assisted coding, the computer suggests a number of codes (say, 10) and a human coder would select the most appropriate one from the suggested list. For this mode, a well-designed user interface would facilitate the human work enormously, but this is beyond the scope of this package.
## look at the top ten predictions for the first answer
head(resCombined[id == 1][order(pred.prob, decreasing = TRUE)], 10)
For Programmers
Don't expect that this package will solve your problem without effort. You will probably need to do some programming to make this package suit your application.
To get started, run the example code in ?predictLogisticRegressionWithPenalization. This function implements penalized multinomial logistic regression (algorithm 5 in the paper) and produces the kind of output you can get with this package. A fast and reasonable way to train a classifier with your own data would be to follow the steps ?predictGweonsNearestNeighbor.
If you know the algorithm you want to use, look at the examples of the predict-function for this algorithm (e.g., ?predictXgboost).
Additional algorithms not mentioned in the paper are described in my dissertation at https://madoc.bib.uni-mannheim.de/50617/ The example code in ?selectMaxProbMethod is a good start to see these algorithms at work. This function implements the Maximum Probability Algorithm (algorithm 10 in the dissertation and just a different implementation of Similarity-based Reasoning as shown above).
Evaluation
Here is a short example code to evaluate how well Similarity-based Reasoning works with new data. Simply replace the data set occupations with your own data and adjust the number of observations n accordingly.
n <- 250 # number of observations
proc.occupations <- removeFaultyAndUncodableAnswers_And_PrepareForAnalysis(occupations, colNames = c("orig_answer", "orig_code"), allowed.codes = kldb2010PlusFive$code, allowed.codes.titles = kldb2010PlusFive$title)
# make predictions
resWordwise2 <- expandPredictionResults(predictSimilarityBasedReasoning(simBasedModelWordwise, proc.occupations), allowed.codes = kldb2010PlusFive$code, method.name = "wordwiseSimilarityOsa1111")
resSubstring2 <- expandPredictionResults(predictSimilarityBasedReasoning(simBasedModelSubstring, proc.occupations), allowed.codes = kldb2010PlusFive$code, method.name = "substringSimilarity")
# second way of combining both models
res.combined <- rbind(resWordwise2, resSubstring2); class(res.combined) <- class(resWordwise2)
res.max <- selectMaxProbMethod(res.combined, combined.methods = c("wordwiseSimilarityOsa1111", "substringSimilarity"), k = 1, new.method.name = "maxProbAmong1")
res.combined <- rbind(res.combined, res.max); class(res.combined) <- class(resWordwise2)
# and create a set of results
produceResults(res.combined, k = 1, n = n, num.codes = 1291)
To be comparable with results from Schierholz (2019) and Schierholz and Schonlau (2021), this code was tested with what they call dataset E. Here are the results:
| Method | Agreement Rate (%) at 100% Production Rate |
| -------- | -------------- |
| Coding with the Alphabetical Dictionary | 30.55 |
| Similarity-based Reasoning | |
| Substring Similiarity | 50.56 |
| Wordwise Similiarity | 48.31 |
| Combining Models: First Strategy (Sequential) | 51.79 |
| Combining Models: Second Strategy (Maximum Probability) | 52.91 |
| | |
| Best algorithm (Schierholz and Schonlau) | 56.20 |
Schierholz and Schonlau (2021) report for automated coding that their best algorithm matches the human-assigned code for about 56% of all answers. Due to privacy concerns, we cannot make the data for this algorithm publicly available, making the algorithm unusable. To overcome this issue, anonymized training data is provided that can be used with Similarity-based Reasoning. The gap between Similarity-based Reasoning and Schierholz and Schonlau's best algorithm is rather small, confirming that Similarity-based Reasoning is a very decent algorithm.
The performance from Substring Similiarity and Wordwise Similiarity is exactly identical to the performance reported in Schierholz (2019). This is despite minor changes in the training data to anonymize it and shows that the anonymization process has no effect on the performance. Likewise (results not shown), any updates in the coding index Gesamtberufsliste_der_BA.xlsx have little effect for the algorithms presented here (two versions were tested: the older version of Gesamtberufsliste_der_BA.xlsx was released in January 2019 and contained almost 28,000 job titles; the newer version was released in July 2021 and contains less than 19,000 entries)
Some improvements are achieved when combining Substring Similiarity and Wordwise Similiarity. We described two combining strategies and the second one accomplishes higher agreement rates. It is therefore recommended for any practical application.
References
- Antoni, M., Drasch, K., Kleinert, C., Matthes, B., Ruland, M. and Trahms, A. (2010): Arbeiten und Lernen im Wandel * Teil 1: Überblick über die Studie, FDZ-Methodenreport 05/2010, Forschungsdatenzentrum der Bundesagentur für Arbeit im Institut für Arbeitsmarkt- und Berufsforschung, Nuremberg.
- Lange, C., Finger, J., Allen, J., Born, S., Hoebel, J., Kuhnert, R., Müters, S., Thelen, J., Schmich, P., Varga, M., von der Lippe, E., Wetzstein, M., Ziese, T. (2017): Implementation of the European Health Interview Survey (EHIS) into the German Health Update (GEDA), Archives of Public Health, 75, 1–14.
- Hoffmann, R., Lange, M., Butschalowsky, H., Houben, R., Schmich, P., Allen, J., Kuhnert, R., Schaffrath Rosario, A., Gößwald, A. (2018): KiGGS Wave 2 Cross-Sectional Study—Participant Acquisition, Response Rates and Representativeness, Journal of Health Monitoring, 3, 78–91. (only wave 2)
- Rohrbach-Schmidt, D., Hall, A. (2013): BIBB/BAuA Employment Survey 2012, BIBB-FDZ Data and Methodological Reports Nr. 1/2013. Version 4.1, Federal Institute for Vocational Education and Training (Research Data Centre), Bonn.
- Schierholz, Malte (2019): New methods for job and occupation classification. Dissertation, Mannheim. https://madoc.bib.uni-mannheim.de/50617/, pp. 206-208 and p. 268, pp. 308-320
- Trappmann, M., Beste, J., Bethmann, A., Müller, G. (2013): The PASS Panel Survey after Six Waves, Journal for Labour Market Research, 46, 275–281. (only wave 10)
malsch/occupationCoding documentation built on March 14, 2024, 8:09 a.m.
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Overview
This is an R package for occupation coding. It accompanies the following publication:
Malte Schierholz, Matthias Schonlau (2021): Machine Learning for Occupation Coding — a Comparison Study. Journal of Survey Statistics and Methodology 9(5), https://doi.org/10.1093/jssam/smaa023
Two groups of users will profit from this package: - Applied users, who have German survey answers and want to run machine learning for automated occupation coding - Programmers, who want to compare various machine learning algorithms for occupation coding and set up a coding pipeline that works best for them
Installation
The easiest way to install this package.
install.packages("remotes")
remotes::install_github("malsch/occupationCoding")
Applied German users will also want to download the file Gesamtberufsliste_der_BA.xlsx from https://download-portal.arbeitsagentur.de/ (requires free registration).
Usage
For Applied Users (German only, coding according to the 2010 German Classification of Occupation)
We will demonstrate how to code answers using the Alphabetical Dictionary (algorithm 1 in the paper) and using Similarity-based Reasoning (not included in the paper, due to space restrictions).
To prepare, load the coding index Gesamtberufsliste_der_BA.xlsx. Enter the path were you saved Gesamtberufsliste_der_BA.xlsx.
path_to_file <- "./Gesamtberufsliste_der_BA.xlsx" # change path
coding_index_w_codes <- prepare_German_coding_index_Gesamtberufsliste_der_BA(path_to_file,
count.categories = FALSE)
Imagine five persons from your survey answered as follows. We wish to code these answers according to the 2010 German Classification of Occupation.
text_input <- c("Bürokauffrau", "Stadtjugendpfleger", "Erzieherhelfer im Lehrlingswohnheim.",
"Mitarbeit bei einer Filmproduktion", "Abschleifer")
Coding with the Alphabetical Dictionary
Consult the coding index. For typical survey data, there is an entry in the coding index for about ~45% of all answers. In rare cases, especially when identical job titles are used in various industries, there can be more than one entry in the processed coding index.
(res <- predictWithCodingIndex(text_input, coding_index = coding_index_w_codes))
## id ans pred.code
## 1: 1 Bürokauffrau 71402
## 2: 2 Stadtjugendpfleger 83124
## 3: 3 Erzieherhelfer im Lehrlingswohnheim. <NA>
## 4: 4 Mitarbeit bei einer Filmproduktion <NA>
## 5: 5 Abschleifer 24222,21212
Similarity-based Reasoning
This was easy. Schierholz (2019) introduces an alternative method, which yields far better results. His technique relies on anonymized training data from various surveys (Antoni et al., 2010; Rohrbach-Schmidt and Hall, 2013; Lange et al., 2017; Hoffmann et al., 2018; Trappmann et al. 2010), which are provided as part of this package.
Training the model will take several minutes...
simBasedModelWordwise <- trainSimilarityBasedReasoning2(anonymized_data = surveyCountsWordwiseSimilarity,
num.allowed.codes = 1291,
coding_index_w_codes = coding_index_w_codes,
preprocessing = list(stopwords = NULL, stemming = NULL, strPreprocessing = TRUE, removePunct = FALSE),
dist.type = "wordwise",
dist.control = list(method = "osa", weight = c(d = 1, i = 1, s = 1, t = 1)),
threshold = c(max = NA, use = 1),
simulation.control = list(n.draws = 250, check.normality = FALSE)
)
simBasedModelSubstring <- trainSimilarityBasedReasoning2(anonymized_data = surveyCountsSubstringSimilarity,
num.allowed.codes = 1291,
coding_index_w_codes = coding_index_w_codes,
preprocessing = list(stopwords = NULL, stemming = NULL, strPreprocessing = TRUE, removePunct = FALSE),
dist.type = "substring",
dist.control = NA,
threshold = NA,
simulation.control = list(n.draws = 250, check.normality = FALSE)
)
Now predict possible codes
resWordwise <- predictSimilarityBasedReasoning(simBasedModelWordwise, text_input)
resSubstring <- predictSimilarityBasedReasoning(simBasedModelSubstring, text_input)
We actually trained two different models. We could use either one, but the results become better if we combine them. If the output from one model yields poor predictions, we simply use the other.
## first way of combining both models: use resWordwise if the five most likeliest categories from resWordwise have a chance of more than 50% to be correct, otherwise use resSubstring
wordwiseModelUsefulIds <- resWordwise[, .(prob = sum(head(.SD[order(pred.prob, decreasing = TRUE), pred.prob], 5))), by = id][prob > 0.5, id]
## second way of combining both models: select for each case the prediction method which returns highest probability (see function selectMaxProbMethod)
wordwiseProb <- resWordwise[, .SD[which.max(pred.prob), .(wordwiseProb = pred.prob)], by = id]$wordwiseProb
substringProb<- resSubstring[, .SD[which.max(pred.prob), .(substringProb = pred.prob)], by = id]$substringProb
wordwiseModelUsefulIds <- which(wordwiseProb > substringProb)
## use the second way of combining (it is the one performing better, shown below)
resCombined <- rbind(resWordwise[id %in% wordwiseModelUsefulIds], resSubstring[!(id %in% wordwiseModelUsefulIds)])[order(id)]
Let's look at the results again. The two most probable codes for each answer are shown next. The algorithm is rather convinced of its predictions for answers 1 and 3, it has some confidence in its predictions for answers 2 and 5, but for answer 4 even the most likeliest code has just a 7% chance to be correct.
resCombined[, head(.SD[order(pred.prob, decreasing = TRUE)], 2), by = id]
## id pred.code pred.prob ans
## 1: 1 71402 0.8928066326 Bürokauffrau
## 2: 1 72213 0.0200043173 Bürokauffrau
## 3: 2 83124 0.6698570028 Stadtjugendpfleger
## 4: 2 -9999 0.0002560075 Stadtjugendpfleger
## 5: 3 83112 0.8308403476 Erzieherhelfer im Lehrlingswohnheim.
## 6: 3 83132 0.0237893319 Erzieherhelfer im Lehrlingswohnheim.
## 7: 4 43423 0.0744033856 Mitarbeit bei einer Filmproduktion
## 8: 4 71393 0.0488194384 Mitarbeit bei einer Filmproduktion
## 9: 5 24222 0.5933915448 Abschleifer
##10: 5 -9999 0.0003152365 Abschleifer
## Note that the code `-9999` has a special meaning: It is always predicted with very low pred.prob and is a placeholder for all codes that have no evidence of being appropriate. As an example, consider the second answer: There are 1291 allowed codes in the target classification. Code 83124 has probability 0.67 and every other code has probability 0.000256, for a total probability of 0.67 + 1290 * 0.000256 = 1.
There are two main use cases for this algorithm: (semi-)automated coding and computer-assisted coding.
With (semi-)automated coding, the computer automatically selects the code that is most likely (at least if pred.prob is above a user-defined threshold). Semi-automated can be done es follows.
threshold <- 0.8 # only for demonstration purposes
resCombined[, .SD[which.max(pred.prob), list(ans, predicted = ifelse(pred.prob > threshold, pred.code, "no prediction"), pred.prob)], by = id]
With computer-assisted coding, the computer suggests a number of codes (say, 10) and a human coder would select the most appropriate one from the suggested list. For this mode, a well-designed user interface would facilitate the human work enormously, but this is beyond the scope of this package.
## look at the top ten predictions for the first answer
head(resCombined[id == 1][order(pred.prob, decreasing = TRUE)], 10)
For Programmers
Don't expect that this package will solve your problem without effort. You will probably need to do some programming to make this package suit your application.
To get started, run the example code in ?predictLogisticRegressionWithPenalization. This function implements penalized multinomial logistic regression (algorithm 5 in the paper) and produces the kind of output you can get with this package. A fast and reasonable way to train a classifier with your own data would be to follow the steps ?predictGweonsNearestNeighbor.
If you know the algorithm you want to use, look at the examples of the predict-function for this algorithm (e.g., ?predictXgboost).
Additional algorithms not mentioned in the paper are described in my dissertation at https://madoc.bib.uni-mannheim.de/50617/ The example code in ?selectMaxProbMethod is a good start to see these algorithms at work. This function implements the Maximum Probability Algorithm (algorithm 10 in the dissertation and just a different implementation of Similarity-based Reasoning as shown above).
Evaluation
Here is a short example code to evaluate how well Similarity-based Reasoning works with new data. Simply replace the data set occupations with your own data and adjust the number of observations n accordingly.
n <- 250 # number of observations
proc.occupations <- removeFaultyAndUncodableAnswers_And_PrepareForAnalysis(occupations, colNames = c("orig_answer", "orig_code"), allowed.codes = kldb2010PlusFive$code, allowed.codes.titles = kldb2010PlusFive$title)
# make predictions
resWordwise2 <- expandPredictionResults(predictSimilarityBasedReasoning(simBasedModelWordwise, proc.occupations), allowed.codes = kldb2010PlusFive$code, method.name = "wordwiseSimilarityOsa1111")
resSubstring2 <- expandPredictionResults(predictSimilarityBasedReasoning(simBasedModelSubstring, proc.occupations), allowed.codes = kldb2010PlusFive$code, method.name = "substringSimilarity")
# second way of combining both models
res.combined <- rbind(resWordwise2, resSubstring2); class(res.combined) <- class(resWordwise2)
res.max <- selectMaxProbMethod(res.combined, combined.methods = c("wordwiseSimilarityOsa1111", "substringSimilarity"), k = 1, new.method.name = "maxProbAmong1")
res.combined <- rbind(res.combined, res.max); class(res.combined) <- class(resWordwise2)
# and create a set of results
produceResults(res.combined, k = 1, n = n, num.codes = 1291)
To be comparable with results from Schierholz (2019) and Schierholz and Schonlau (2021), this code was tested with what they call dataset E. Here are the results:
| Method | Agreement Rate (%) at 100% Production Rate | | -------- | -------------- | | Coding with the Alphabetical Dictionary | 30.55 | | Similarity-based Reasoning | | | Substring Similiarity | 50.56 | | Wordwise Similiarity | 48.31 | | Combining Models: First Strategy (Sequential) | 51.79 | | Combining Models: Second Strategy (Maximum Probability) | 52.91 | | | | | Best algorithm (Schierholz and Schonlau) | 56.20 |
Schierholz and Schonlau (2021) report for automated coding that their best algorithm matches the human-assigned code for about 56% of all answers. Due to privacy concerns, we cannot make the data for this algorithm publicly available, making the algorithm unusable. To overcome this issue, anonymized training data is provided that can be used with Similarity-based Reasoning. The gap between Similarity-based Reasoning and Schierholz and Schonlau's best algorithm is rather small, confirming that Similarity-based Reasoning is a very decent algorithm.
The performance from Substring Similiarity and Wordwise Similiarity is exactly identical to the performance reported in Schierholz (2019). This is despite minor changes in the training data to anonymize it and shows that the anonymization process has no effect on the performance. Likewise (results not shown), any updates in the coding index Gesamtberufsliste_der_BA.xlsx have little effect for the algorithms presented here (two versions were tested: the older version of Gesamtberufsliste_der_BA.xlsx was released in January 2019 and contained almost 28,000 job titles; the newer version was released in July 2021 and contains less than 19,000 entries)
Some improvements are achieved when combining Substring Similiarity and Wordwise Similiarity. We described two combining strategies and the second one accomplishes higher agreement rates. It is therefore recommended for any practical application.
References
- Antoni, M., Drasch, K., Kleinert, C., Matthes, B., Ruland, M. and Trahms, A. (2010): Arbeiten und Lernen im Wandel * Teil 1: Überblick über die Studie, FDZ-Methodenreport 05/2010, Forschungsdatenzentrum der Bundesagentur für Arbeit im Institut für Arbeitsmarkt- und Berufsforschung, Nuremberg.
- Lange, C., Finger, J., Allen, J., Born, S., Hoebel, J., Kuhnert, R., Müters, S., Thelen, J., Schmich, P., Varga, M., von der Lippe, E., Wetzstein, M., Ziese, T. (2017): Implementation of the European Health Interview Survey (EHIS) into the German Health Update (GEDA), Archives of Public Health, 75, 1–14.
- Hoffmann, R., Lange, M., Butschalowsky, H., Houben, R., Schmich, P., Allen, J., Kuhnert, R., Schaffrath Rosario, A., Gößwald, A. (2018): KiGGS Wave 2 Cross-Sectional Study—Participant Acquisition, Response Rates and Representativeness, Journal of Health Monitoring, 3, 78–91. (only wave 2)
- Rohrbach-Schmidt, D., Hall, A. (2013): BIBB/BAuA Employment Survey 2012, BIBB-FDZ Data and Methodological Reports Nr. 1/2013. Version 4.1, Federal Institute for Vocational Education and Training (Research Data Centre), Bonn.
- Schierholz, Malte (2019): New methods for job and occupation classification. Dissertation, Mannheim. https://madoc.bib.uni-mannheim.de/50617/, pp. 206-208 and p. 268, pp. 308-320
- Trappmann, M., Beste, J., Bethmann, A., Müller, G. (2013): The PASS Panel Survey after Six Waves, Journal for Labour Market Research, 46, 275–281. (only wave 10)
R Package Documentation
Browse R Packages
We want your feedback!
Note that we can't provide technical support on individual packages. You should contact the package authors for that.
Embedding an R snippet on your website
Add the following code to your website.
For more information on customizing the embed code, read Embedding Snippets.