Karen R. Harker, MLS, MPH Collection Assessment Librarian UNT Libraries IS IT REALLY THAT BAD? Verifying the extent of full-text linking problems
I find searching for journal articles and actually finding full articles to a very difficult. Sometimes it just links to another site and then to a fragment of an article. It's hard to find links to online articles--some links say they can't find it, but sometimes you Frustrated can with still locate "Find it and Full-Text" I don't know when why the main "find links" page doesn't bring it up. "it doesn't work ; if it's not 100% perfect, there is really no point in Sometimes, if a link is not provided offering the service. to an article in the search results, it is very difficult to find. The article linker often comes up with no results even though it says the article is in UNT's collection.
What have you done?
R A N D O M Link-Checking Selection Cannot Be Predicted
What is required? Intermediate to advanced Excel (but NOT programming)
Knowledge About your collection About the problem
Clear questions
Think about... your problem your collection your link resolver your people
Brainstorm
Come away with
What Kind of Research Questions? What is important to you? Just a few Compared to what? Be Specific
Such as Are links from EBSCO more successful than links from Ovid? Is Serials Solutions 360 link resolver more likely to get to the full-text from our key resources than EBSCO s? Is full-text linking better or worse compared to last year? What is the chance that a client will get the full-text of an article on the first click?
Start with the Results
Comparing One Source With Another Sources Full-Text No Full-Text Total EBSCO 90 10 100 Ovid 85 15 100 Totals 175 25 200 Confidence Level Chi-Square Test Target Chi-Square 0.90 0.161 1.00-0.90=0.10 Is Chi-Square test < Target? No Is Ovid Significantly Different from EBSCO? No
Comparing One Source With the Average Sources Full-Text No Full-Text Total EBSCO 90 10 100 Average 75 25 100 Totals 165 35 200 Confidence Level Chi-Square Test Target Chi-Square 0.90 0.001 1.00-0.90=0.10 Is Chi-Square test < Target? Yes Is EBSCO Significantly Different from the Average? Yes
Comparing One Target With The Expected or Ideal Rate Targets Full-Text No Full-Text Total EBSCO 90 10 100 Expected or Ideal Rate 95 5 100 Totals 185 15 200 Confidence Level Chi-Square Test Target Chi-Square 0.90 0.022 1.00-0.90=0.10 Is Chi-Square test < Target? Yes Is EBSCO Significantly Different from Ideal Rate? Yes
Random Sampling Review or background, depending on your viewpoint
Sampling Terms Citations in databases to which we have access to articles to which we have full-text access Sampling Population Only Sampling Journal Articles Frame Universe All Citations Sample
Selection Methods Non-probability Convenience sampling The chance of being selected is not known Probability The probability of any one citation being selected is known.
Simple Random Sampling Every citation that meets the criteria has an equal chance of being selected. See Demo. NOTE: Articles vary greatly by source, target and year.
Stratified Sampling Every citation in discrete homogeneous groups has an equal chance of being chosen. Try Demo again Useful to zero-in on a possible problem Stratify by source, target & year, but would be time consuming
Sampling Population Strata Samples Selected from Each Stratum
Cluster Sampling When the sampling population naturally clusters (e.g. source and targets). The way they cluster doesn t affect your outcome. Divide population into these clusters Randomly select the clusters to be a part of the sampling frame Randomly select sample from selected clusters Useful for very large populations.
Sampling Population Clusters Samples Selected from Selected Clusters
This Methodology Simple randomized cluster 1. Select a sample of ejournals (clusters) 2. Search each database for articles 3. Randomly select a citation (sample) 4. Test and record results Most useful for questions that are focused on the sources.
Other Questions, Other Designs Comparing link-resolvers: Matched-pair 1. Select a sample of ejournals 2. Using one of the link resolvers, search the source for articles in these ejournals. 3. Randomly select a citation 4. Test and record results 5. For the next link resolver, search each source for the same citation (the matched pair). 6. Test and record results.
Other Questions, Other Designs For problems related to targets: Use the same method, but Randomly select ejournals from each target For problems related to environment (browsers, location of user, etc.): Use the same method as the link-resolver, only change the browser or location.
Practical Applications Using Excel to Help You Along
Before we begin For those with Laptops, download files: Excel file: http://digital.library.unt.edu/ark:/67531/metadc96818/ PDF of Steps: http://digital.library.unt.edu/ark:/67531/metadc96827/ Or, just follow along
I need to check how many? May be fewer than you d think Need to know: Sampling strategy Kind of analysis Expected rate Chance of a title being indexed in the source Number of databases or sources to examine Educated Guess
Selecting the Journal Titles (Clusters) 1. Download your ejournals list May want to limit to only those used recently 2. Randomly select the correct number of journals based on sample size 3. Randomly assign each title to the databases or sources you will be searching. Excel Tricks Remove Duplicates Fill Cell assigns a new ID number Sampling method in Data Analysis Randomly selects IDs from your list VLOOKUP gets the titles for the selected IDs RANDBETWEEN Randomly assigns each title to a source to be searched
Search Source Record Result If Not Found Random Number Test Citation Select Citation
Test the Sources 1. Login to the database 2. Search for articles in the first journal 3. If none are found, note this in your results and skip to next title. 4. If some are found, note the total number of articles.
Test Sources If articles are found: 1. Sort the list by author last name (if possible) 2. Note the total number of articles found 3. Enter this number in the Sample Size Calculator worksheet (Random Article Selector) 4. Note the Select this article number 5. In the database, navigate to this article 6. Click on the Find Full-Text button
Full-Text PDF!
Test Sources Note the success of that link in your tally sheet Rinse & repeat: For each journal in the list For each database to be tested
Search Source Record Result If Not Found Random Number Test Citation Select Citation
Tips & Tricks Search by ISSN, if possible. Display the most citations per page If full-text article is in that database, skip title. This is a non-response the Response Rate the Sample Size
Summarize the data Count up all results in each source Create ratios for each result (e.g. Full-text ratio) # Full-Text / # Titles Found Average the ratios
Example Data: Raw Counts
Example Data: Ratios
Test the Results So you have a ratio so what? What does it mean? Is it high? Low? Compared to what? Use the Chi-Squared test to compare the ratios Excel: CHISQ.TEST(actual range, expected range) If the result is less than 0.10 (or 1.00 Confidence Level), then the difference is statistically significant. This may be good or not good, depending on what your comparing against.
Comparing Against an Ideal % Actual range: the # of Successes & # of Failures Expected Range: # expected to be success & # of expected failures Example: CHISQ.TEST(B2:C2, B3:C3) A B C D 1 Sources No Full- Full-Text Text Total 2 EBSCO 90 10 100 3 Expected or Ideal Rate 95 5 100 4 Totals 185 15 200 5 Chi-Square Test 0.022 6 Significantly Different from Expected Count? Yes
Sources to Test Sources Full-Text No Full-Text Total 1 EBSCO 70 21 91 2 Ovid 90 23 113 Totals 160 44 204 Chi-Square Test 0.031632 Significantly Different? Yes Sources to Test Sources Full-Text No Full-Text Total 3 ProQuest 87 18 105 1 EBSCO 70 21 91 Totals 157 39 196 Chi-Square Test 0.032782 Significantly Different? Yes Sources to Test Sources Full-Text No Full-Text Total 2 Ovid 90 23 113 5 Means 82.3333333 20.66666667 103 Totals 172.333333 43.66666667 216 Chi-Square Test 0.322856 Significantly Different? No
No Link Ratio 4% Overall Success Any Link Ratio 16% Any F/T Ratio 80% 120 # Titles by Link Type 100 80 60 40 No Link Any Link Any Full-Text 20 0 EBSCO Ovid ProQuest
Context is King The value of the result depends on what you are measuring and comparing If the difference between two sources is not significant, then they are statistically similar. If the difference between two link resolvers is significant, then one is better than the other. NOTE: This doesn t tell you by how much! Use your best judgment
I am here to help you Karen R. Harker 940-565-2688 karen.harker@unt.edu