R

In my last post, I made the point that p-values should not necessarily be considered sufficient evidence (or evidence at all) in drawing conclusions about associations we are interested in exploring. When it comes to contingency tables that represent the outcomes for two categorical variables, it isn’t so obvious what measure of association should augment (or replace) the \(\chi^2\) statistic.

I described a model-based measure of effect to quantify the strength of an association in the particular case where one of the categorical variables is ordinal. This can arise, for example, when we want to compare Likert-type responses across multiple groups. The measure of effect I focused on - the cumulative proportional odds - is quite useful, but is potentially limited for two reasons. First, the proportional odds assumption may not be reasonable, potentially leading to biased estimates. Second, both factors may be nominal (i.e. not ordinal), it which case cumulative odds model is inappropriate.

I frequently find myself in discussions with collaborators about the merits of reporting p-values, particularly in the context of pilot studies or exploratory analysis. Over the past several years, the American Statistical Association has made several strong statements about the need to consider approaches that measure the strength of evidence or uncertainty that don’t necessarily rely on p-values. In 2016, the ASA attempted to clarify the proper use and interpretation of the p-value by highlighting key principles “that could improve the conduct or interpretation of quantitative science, according to widespread consensus in the statistical community.” These principles are worth noting here in case you don’t make it over to the original paper:

I am always saying that simulation can help illuminate interesting statistical concepts or ideas. Such an exploration might provide some insight into the concept of the design effect, which underlies clustered randomized trial designs. I’ve written about clustered-related methods so much on this blog that I won’t provide links - just peruse the list of entries on the home page and you are sure to spot a few. But, I haven’t written explicitly about the design effect.

I am currently wrestling with how to analyze data from a stepped-wedge designed cluster randomized trial. A few factors make this analysis particularly interesting. First, we want to allow for the possibility that between-period site-level correlation will decrease (or decay) over time. Second, there is possibly additional clustering at the patient level since individual outcomes will be measured repeatedly over time. And third, given that these outcomes are binary, there are no obvious software tools that can handle generalized linear models with this particular variance structure we want to model. (If I have missed something obvious with respect to modeling options, please let me know.)

The ROC (receiver operating characteristic) curve visually depicts the ability of a measure or classification model to distinguish two groups. The area under the ROC (AUC), quantifies the extent of that ability. My goal here is to describe as simply as possible a process that serves as a foundation for the ROC, and to provide an interpretation of the AUC that is defined by that curve.

I’m participating in the design of a new study that will evaluate interventions aimed at reducing both pain and opioid use for patients on dialysis. This study is likely to be somewhat complicated, possibly involving multiple clusters, multiple interventions, a sequential and/or adaptive randomization scheme, and a composite binary outcome. I’m not going into any of that here.

There is one issue that should be fairly generalizable to other studies. It is likely that individual measures will be collected repeatedly over time but the primary outcome of interest will be the measure collected during the last follow-up period. I wanted to explore what, if anything, can be gained by analyzing all of the available data rather than focusing only the final end point.

I am contemplating adding a new distribution option to the package simstudy that would allow users to define a new variable as a mixture of previously defined (or already generated) variables. I think the easiest way to explain how to apply the new mixture option is to step through a few examples and see it in action.

I am involved with a very interesting project - the NIA IMPACT Collaboratory - where a primary goal is to fund a large group of pragmatic pilot studies to investigate promising interventions to improve health care and quality of life for people living with Alzheimer’s disease and related dementias. One of my roles on the project team is to advise potential applicants on the development of their proposals. In order to provide helpful advice, it is important that we understand what we should actually expect to learn from a relatively small pilot study of a new intervention.

A couple of months ago, I was contacted about the possibility of creating a simple function in simstudy to generate a large dataset that could include possibly 10’s or 100’s of potential predictors and an outcome. In this function, only a subset of the variables would actually be predictors. The idea is to be able to easily generate data for exploring ridge regression, Lasso regression, or other “regularization” methods. Alternatively, this can be used to very quickly generate correlated data (with one line of code) without going through the definition process.

I am collaborating with a number of folks who think a lot about palliative or supportive care for people who are facing end-stage disease, such as advanced dementia, cancer, COPD, or congestive heart failure. A major concern for this population (which really includes just about everyone at some point) is the quality of life at the end of life and what kind of experiences, including interactions with the health care system, they have (and don’t have) before death.