You are here: Vanderbilt Biostatistics Wiki>Main Web>Clinics>ClinicGeneral>GenClinicAnalyses>MondayClinicNotes2017 (15 Jan 2021, DalePlummer)EditAttach

Notes 2017

- What are the speech perceptual characteristics consistent with diagnosis of HD and how do they compare to previous literature on hyperkinetic dysarthria?
- Are there distinct clusters of speech perceptual characteristics within speakers with mild dysarthria due to HD?
- If distinct clusters of speech characteristics exist within speakers with HD, do the individuals who belong to the same cluster also share other disease- or treatment-related features (i.e. type of medications, number of CAG repeats, and the length of disease duration)?

- Do not do statistical tests comparing groups on the 38 items that went into the cluster analysis. It would be reasonable to do tests on things like medication use, etc., that did not go into the cluster analysis.
- To show group differences visually, consider plotting the first two principal components using colors for the different clusters, and/or making a parallel coordinates plot
- With so few patients relative to the number of items, the clusters are likely to be unstable, although if the plots show large separation between groups, this may be less of a concern
- Consider trying a few different clustering methods to see whether they all suggest the same four clusters
- Consider sparse principal components analysis, and either cluster on some or all of the principal components, or use the PCA results to help you decide which variables to cluster on
- For a manuscript, present intra- and inter-rater reliability

- Because some of the data has been published already, focus on two comparisons, using Wilcoxon rank-sum tests: PD vs. HD and PD vs. control
- For comparisons involving race and gender, use descriptive plots rather than statistical tests (because the group sizes are very small)

- First step: see how comparable the responders are to the non-responders (or the whole set of programs) in terms of # residents, # MD/PhD residents, NIH funding amount
- From there we can talk about statistical testing. We will probably want to use a finite-population correction since the whole population = 63 programs
- Regardless of comparability, descriptive statistics (10 out of 23... etc.) will still be interesting to report

- We are concerned because we don't know the number of times each system was tested. If it's not possible to get this information, one possibility might be to simulate data to try to get a sense of the possible scope of the impact of frequency-of-testing
- The overall project seems like a good fit for a VICTR voucher or short-term biostatistics support (its scope is too large for clinic). To inquire about short-term biostats support, email Yu Shyr, Chair. Another possibility might be working with a student (email Jeffrey Blume, director of graduate studies).
- Will need to keep in mind: some systems get swallowed up into other systems.
- Longitudinal data analysis won't be feasible without the complete testing data (we would need the non-violations in addition to the violations).
- Next level (after other issues resolved): geospatial correlation (tricky, though, because of the upstream/downstream issue)

- Pearson’s chi-squared or Fisher’s exact test, depending on n in each category
- Ordinal regression (vs. multinomial?)

- Pearson’s chi-squared or Fisher’s exact test, depending on n in each category
- Logistical regression

- You are welcome to come back to clinic, but as a member of the Nephrology division you are also welcome to work directly with Thomas Stewart
- Recruit patients across a range of SES's; will probably want to limit to patients who are either AA or white, due to likely low numbers in other groups

- You are welcome to come back to clinic, but as a member of the Gastroenterology division you may be able to work directly with Chris Slaughter
- Identifying appropriate controls for this study will be tricky

- for enrolled patients, look at previous 3 years then follow forward 2 years
- 250 children & 300 adults seen at clinic
- plasma & DNA samples
- outcome is pain and acute chest syndrome
- 3 genotypes and plasma biomarkers
- believe 2,2 genotype will have increased pain and chest syndrome (1,1 (26%) 1,2 (55%) 2,2 (19%))
- look at incident rate at 1 yr and 2 yr; poisson model; need to know what difference would be expected between the groups
- sample size - graph of incidence rate that can be detected vs sample size needed
- Simplest approach: find confidence interval formula for a Poisson rate; assume lowest true rate and solve for n such that multiplicative margin of error is 1.5 with 0.95 confidence
- Simplest confidence interval is lambda +- 1.96 * sqrt(lambda / n); once you have an upper limit on lambda can solve for n to give acceptable margin of error for lambda
- Would be better to get the multiplicative margin of error for the ratio of two Poisson rates (to simplify we may assume the sample size in each group is the lowest of the three genotype group sizes)
- See http://statsdirect.com/help/rates/compare_crude_incidence_rates.htm

- Perception of collaboration between doctors and nurses in Guyana
- Pre- and post-team-building exercise, then 4-6w later
- 27 participants with 2 dropouts by the end; 15 nurses, 10 doctors at the end
- Issue of using means vs. proportions for Likert scales; want to look at disagreements of perception before and after training
- Nurses used more spread of answers than doctors
- 2 demographic variables, 15 Likert questions; need to combine into a single global scale for graphical individual profiles and for stat analysis
- Can do a formal analysis of variability of responses within subject, e.g. compute the SD over 15 questions within subject and see if nurses have more variation than doctors
- Main analysis on mean
- Form within-person difference from baseline (paired data)
- Do 2-sample (unpaired) t-test comparing these differences - nurses vs doctors

- Be sure to graph all measurements (on summary score)
- Pre-post design often provides an upper limit to an intervention effect

- Sample size is fixed based on fellowship time. Power and sample size should be calculated accordingly.
- Keep the measure in the continuous form (0-100) instead of dichonimization.
- Consider to have CTSA statistician's early involvement at the design stage. Given this involves design, grant writing, data collection, data analysis, and manuscript preparation, a 90 hour work maybe needed.
- As prediction is involved (identify characters that are related to high measures), model validation should be considered.

- If dropout is not random, either GLS with a serial correlation structure or a linear mixed-effects model would be more appropriate than GEE.
- Do not collapse the diet variables into quintiles; leave them as continuous variables
- For the power calculation, it may be possible to ask for conditional approval to have access to a subset of the data to get estimates of the quantities needed for a power calculation.
- You can do a simplified power calculation with just one wave of data, and argue that the power will be higher when there are more data points per person.
- Possibly useful R packages: longpower (thank you for bringing this to our attention!), pwr (in particular, the pwr.f2.test function).
- Simulation could also be a useful approach, but it would also require some background information about the standard deviations of the variables

- Keep all levels of the "bother" variable (do not collapse it). When using it as an outcome, a good approach might be proportional odds logistic regression
- Consider trying to cluster the variables rather than the people (using factor analysis or another approach)
- A good reference for the preceding points: http://www.springer.com/gb/book/9783319194240
- Developing something like the Crohn's Disease Activity Index, https://en.wikipedia.org/wiki/Crohn%27s_Disease_Activity_Index, may be useful

- Email Hakmook Kang to talk about the possibility of working through the KC biostatistics core to get an estimate of how many children and timepoints you would need to do the flexible-breakpoint approach discussed in the article
- We also discussed an approach using restricted cubic splines. It's possible that this approach would let you use fewer subjects; it may be useful even though you are expecting a linear relationship

- See https://stats.idre.ucla.edu/stata/dae/ordered-logistic-regression/ for an explanation of proportional odds logistic regression and some helpful language for describing the results

- In deciding which categories to collapse, look at the sample overall (not by complication status)
- To increase power, consider treating the outcome as an ordinal, rather than binary, variable if there are enough people in the additional groups
- Look at the cross-tabulation between physician and sling type to see whether it is feasible to include both
- Leave the continuous variables as is (do not categorize them). May want to consider log-transforming age.
- Try variable clustering to see which variables may be collinear/redundant
- Consider combining less important (less interesting) variables into a score
- For binary logistic regression, we generally want to have 10--20 people in the smaller outcome group for every degree of freedom (continuous variable or single category) in the model
- If you apply for VICTR funding, we recommend the larger time amount if you are interested in a publication or presentation. In your application, you can cite these notes as evidence that you have been to a biostatistics clinic.

- Get more information about the survey design (especially number of people surveyed) so that you can compare the response rates in 2012 and 2016. If they are not close to each other, it will be harder to justify comparing the results of the two surveys
- If possible, get info about demographic makeup of the people surveyed in 2012 and 2016 from the organization's records. If, for example, the mean age of respondents is very different from the known mean age of the people surveyed, you will know that in at least that one aspect, the respondents are not representative of the people surveyed.
- Chi-squared tests should be fine if the categories are exhaustive (but this is secondary to the nonresponse issue)
- If possible, get more info about the outcomes and model specifications used for the regressions in Table 3.

- This project investigates speech-language imbalances in children. We are interested in the best way to measure imbalances using five standardized tests. Simple range scatter and standard deviation have been discussed. We are also interested in the best way to analyze whether increased synchrony between the five tests is associated with a decrease in stuttering frequency based on two years of development.

- The objective of this study was to measure the energy expenditure (oxygen consumption O2/kg/min) of adults practicing common yoga movements. For each individual, participants were asked to do movements in a standing position, lying position, and seated position (body orientation). In addition, each movement was done with different variations serially. In addition, participants were asked to walk at low and moderate intensities to compare energy expenditure of a comparative aerobic exercise to yoga.

- This project investigates differences in skin conductance levels in children who stutter and are persisting, children who stuttered and recovered, and children who do not stutter. All children were followed 3-4 times across a two year period. At each visit, skin conductance levels were measured during a neutral video and speaking task, a positive emotion-inducing video and speaking task, and a negative emotion-inducing video and speaking task. We would like to discuss the best statistical models for our hypotheses.

- Note that at each timepoint, there are 7 skin conductance measures (a "baseline" and 6 other measures)

- Recommendations:
- Keep all possible timepoints from all possible subjects. Do not exclude subjects based on their trajectories or baseline characteristics
- Use continuous versions of the stuttering outcomes if possible; at a minimum, collapse the outcomes into 5 ordinal categories
- Use a longitudinal mixed-effects model. Each subject will contribute 1, 2, or 3 rows depending on how many of the timepoints they have. You can model severity as a function of time-1 severity, age, sex, the seven time-1 conductance measures (or a reduction thereof; try a redundancy analysis first), time in days, and squared time in days, with random effects for subject (and possibly time and squared time). We recommend a continuous-time correlation structure, but this might be tricky with the mixed-effects model; generalized least squares might work better.
- If we can get a clear, simple plan and the analysis is not a multi-step analysis and the dataset is clean (and tall and thin, with the relevant time-1 variables and non-identifying subject ID on each row), we may be able to conduct the analysis during a clinic.
- Starting next month, we will be able to take on longer short-term projects for a charge.
- The Kennedy Center statistics core may also be able to do this. If you come back to a clinic, please remind us to invite Hakmook.

- For each overall question category, try a scatterplot of a) the means and b) the standard deviations for each item, with staff values on the x-axis and parent values on the y-axis (or vice-versa). Label each point with the question number or a short phrase to identify it
- Do variable clustering within the staff items and the parent items, to see which items tend to be answered similarly by the same person (hcavar in stata)
- Rather than doing several univariate analyses comparing the relationship between the demographic items and each survey item, do a single regression analysis for each survey item, with all the demographic items included in the model at once. Collapse the categorical items into 2 or at most 3 categories, and just assign numeric values (e.g. 1--5) to the levels in the binned
*continuous*items like distance and treat those as continuous variables (so they will have just one term in the model). Actually, though, drop distance altogether and just use travel time. The overall F-statistic from the regression will tell you whether anything in the model matters. The best approach would be a proportional odds model, but ordinary regression will be next best. - It's ok to take the means of means (across items in a particular category) and talk about those, but there aren't enough data points to warrant a statistical test.

- Instead of doing t-tests, do wilcoxon rank-sum test (only 5 response options)
- Rather than overlaying the parent and staff histograms, show the parent mean as a dot on the staff histograms
- Do the "dot-histograms" by hospital because the hospitals are so different, even if tests comparing hospitals are not significant
- Don't put too much weight on the p-values; this is exploratory research with relatively small sample sizes
- For the two similar staff questions, run a correlation on the responses to help justify using only one of the questions. Use a Spearman rank correlation.
- We don't think it would make sense to take the mean of the responses for the parent "how often" questions
- For any set of questions, it could be interesting to order the means to see which questions had the highest or lowest means, but it wouldn't make sense to do a statistical test comparing the means of the different items.

Edit | Attach | Print version | History: r1 | Backlinks | View wiki text | Edit wiki text | More topic actions

Topic revision: r1 - 15 Jan 2021, DalePlummer

Copyright © 2013-2022 by the contributing authors. All material on this collaboration platform is the property of the contributing authors.

Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback

Ideas, requests, problems regarding Vanderbilt Biostatistics Wiki? Send feedback