Dataset Information

UNEXPECTED PROPERTIES OF BANDWIDTH CHOICE WHEN SMOOTHING DISCRETE DATA FOR CONSTRUCTING A FUNCTIONAL DATA CLASSIFIER.

ABSTRACT: The data functions that are studied in the course of functional data analysis are assembled from discrete data, and the level of smoothing that is used is generally that which is appropriate for accurate approximation of the conceptually smooth functions that were not actually observed. Existing literature shows that this approach is effective, and even optimal, when using functional data methods for prediction or hypothesis testing. However, in the present paper we show that this approach is not effective in classification problems. There a useful rule of thumb is that undersmoothing is often desirable, but there are several surprising qualifications to that approach. First, the effect of smoothing the training data can be more significant than that of smoothing the new data set to be classified; second, undersmoothing is not always the right approach, and in fact in some cases using a relatively large bandwidth can be more effective; and third, these perverse results are the consequence of very unusual properties of error rates, expressed as functions of smoothing parameters. For example, the orders of magnitude of optimal smoothing parameter choices depend on the signs and sizes of terms in an expansion of error rate, and those signs and sizes can vary dramatically from one setting to another, even for the same classifier.

SUBMITTER: Carroll RJ

PROVIDER: S-EPMC4191932 | biostudies-literature | 2013 Dec

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

UNEXPECTED PROPERTIES OF BANDWIDTH CHOICE WHEN SMOOTHING DISCRETE DATA FOR CONSTRUCTING A FUNCTIONAL DATA CLASSIFIER.

Carroll Raymond J RJ Delaigle Aurore A Hall Peter P

Annals of statistics 20131201 6

The data functions that are studied in the course of functional data analysis are assembled from discrete data, and the level of smoothing that is used is generally that which is appropriate for accurate approximation of the conceptually smooth functions that were not actually observed. Existing literature shows that this approach is effective, and even optimal, when using functional data methods for prediction or hypothesis testing. However, in the present paper we show that this approach is no ...[more]

PMID: 25309640

Similar Datasets

Project description:BackgroundDigital technological development in the last 20 years has led to significant growth in digital collection, use, and sharing of health data. To maintain public trust in the digital society and to enable acceptable policy-making in the future, it is important to investigate people's preferences for sharing digital health data.ObjectiveThe aim of this study is to elicit the preferences of the public in different Northern European countries (the United Kingdom, Norway, Iceland, and Sweden) for sharing health information in different contexts.MethodsRespondents in this discrete choice experiment completed several choice tasks, in which they were asked if data sharing in the described hypothetical situation was acceptable to them. Latent class logistic regression models were used to determine attribute-level estimates and heterogeneity in preferences. We calculated the relative importance of the attributes and the predicted acceptability for different contexts in which the data were shared from the estimates.ResultsIn the final analysis, we used 37.83% (1967/5199) questionnaires. All attributes influenced the respondents' willingness to share health information (P<.001). The most important attribute was whether the respondents were informed about their data being shared. The possibility of opting out from sharing data was preferred over the opportunity to consent (opt-in). Four classes were identified in the latent class model, and the average probabilities of belonging were 27% for class 1, 32% for class 2, 23% for class 3, and 18% for class 4. The uptake probability varied between 14% and 85%, depending on the least to most preferred combination of levels.ConclusionsRespondents from different countries have different preferences for sharing their health data regarding the value of a review process and the reason for their new use. Offering respondents information about the use of their data and the possibility to opt out is the most preferred governance mechanism.

Project description:BACKGROUND:The use of shotgun metagenomics to analyse low-complexity microbial communities in foods has the potential to be of considerable fundamental and applied value. However, there is currently no consensus with respect to choice of species classification tool, platform, or sequencing depth. Here, we benchmarked the performances of three high-throughput short-read sequencing platforms, the Illumina MiSeq, NextSeq 500, and Ion Proton, for shotgun metagenomics of food microbiota. Briefly, we sequenced six kefir DNA samples and a mock community DNA sample, the latter constructed by evenly mixing genomic DNA from 13 food-related bacterial species. A variety of bioinformatic tools were used to analyse the data generated, and the effects of sequencing depth on these analyses were tested by randomly subsampling reads. RESULTS:Compositional analysis results were consistent between the platforms at divergent sequencing depths. However, we observed pronounced differences in the predictions from species classification tools. Indeed, PERMANOVA indicated that there was no significant differences between the compositional results generated by the different sequencers (p = 0.693, R2 = 0.011), but there was a significant difference between the results predicted by the species classifiers (p = 0.01, R2 = 0.127). The relative abundances predicted by the classifiers, apart from MetaPhlAn2, were apparently biased by reference genome sizes. Additionally, we observed varying false-positive rates among the classifiers. MetaPhlAn2 had the lowest false-positive rate, whereas SLIMM had the greatest false-positive rate. Strain-level analysis results were also similar across platforms. Each platform correctly identified the strains present in the mock community, but accuracy was improved slightly with greater sequencing depth. Notably, PanPhlAn detected the dominant strains in each kefir sample above 500,000 reads per sample. Again, the outputs from functional profiling analysis using SUPER-FOCUS were generally accordant between the platforms at different sequencing depths. Finally, and expectedly, metagenome assembly completeness was significantly lower on the MiSeq than either on the NextSeq (p = 0.03) or the Proton (p = 0.011), and it improved with increased sequencing depth. CONCLUSIONS:Our results demonstrate a remarkable similarity in the results generated by the three sequencing platforms at different sequencing depths, and, in fact, the choice of bioinformatics methodology had a more evident impact on results than the choice of sequencer did.

Project description:ObjectivesThe objective of this paper is to study under which circumstances wearable and health app users would accept a compensation payment, namely a digital dividend, to share their self-tracked health data.MethodsWe conducted a discrete choice experiment alternative, a separated adaptive dual response. We chose this approach to reduce extreme response behavior, considering the emotionally-charged topic of health data sales, and to measure willingness to accept. Previous experiments in lab settings led to demands for high monetary compensation. After a first online survey and two pre-studies, we validated four attributes for the final online study: monthly bonus payment, stakeholder handling the data (e.g., health insurer, pharmaceutical or medical device companies, universities), type of data, and data sales to third parties. We used a random utility framework to evaluate individual choice preferences. To test the expected prices of the main study for robustness, we assigned respondents randomly to one of two identical questionnaires with varying price ranges.ResultsOver a period of three weeks, 842 respondents participated in the main survey, and 272 respondents participated in the second survey. The participants considered transparency about data processing and no further data sales to third parties as very important to the decision to share data with different stakeholders, as well as adequate monetary compensation. Price expectations resulting from the experiment were high; pharmaceutical and medical device companies would have to pay an average digital dividend of 237.30€/month for patient generated health data of all types. We also observed an anchor effect, which means that people formed price expectations during the process and not ex ante. We found a bimodal distribution between relatively low price expectations and relatively high price expectations, which shows that personal data selling is a divisive societal issue. However, the results indicate that a digital dividend could be an accepted economic incentive system to gather large-scale, self-tracked data for research and development purposes. After the COVID-19 crisis, price expectations might change due to public sensitization to the need for big data research on patient generated health data.ConclusionA continuing success of existing data donation models is highly unlikely. The health care sector needs to develop transparency and trust in data processing. An adequate digital dividend could be an effective long-term measure to convince a diverse and large group of people to share high-quality, continuous data for research purposes.

Dataset Information

UNEXPECTED PROPERTIES OF BANDWIDTH CHOICE WHEN SMOOTHING DISCRETE DATA FOR CONSTRUCTING A FUNCTIONAL DATA CLASSIFIER.

Publications

UNEXPECTED PROPERTIES OF BANDWIDTH CHOICE WHEN SMOOTHING DISCRETE DATA FOR CONSTRUCTING A FUNCTIONAL DATA CLASSIFIER.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets