FPIN Journal Club
COHORT STUDIES CHECKLIST
Title: Skip this step when checking lipid levels
Journal Club Author and Editor: Kate Rowland, M.D., M.S., Rush-Copley Medical Center
PURL Citation: Wootten M, Stulberg DB, Prasad S, Rowland K. PURLs: skip this step when checking lipid levels. J Fam Pract. 2015 Feb;64(2):113-5. PubMed PMID: 25671529; PubMed Central PMCID: PMC4324304.
Original Article: Doran B, Guo Y, Xu J, Weintraub H, Mora S, Maron DJ, Bangalore S . Prognostic Value of Fasting vs. Non-Fasting Low Density Lipoprotein Cholesterol Levels on Long-term Mortality: Insight from the National Health and Nutrition Survey III (NHANES-III). Circulation. 2014;130:546-553. PMID: 25015340
Cohort studies: Definition
A type of study that attempts to answer “What are the effects of this exposure?”
Relates to studies that compare a group of people with a particular exposure with another group who either have not had the exposure, or have a different level of exposure.
- Prospective: where the exposure is defined and subjects selected before outcomes occur.
- Retrospective: where exposure is assessed after the outcome is known, usually by the examination of medical records.
1. What question did the study attempt to answer?
Patients – Adults participating the National Health and Nutrition Survey (NHANES-III) study
Intervention – fasting >8 hours at the time the cholesterol panel was done
Comparison – fasting <8 hours (including not fasting at all)
Outcome – all cause and cardiovascular mortality
Did the study address an appropriate and clearly focused question Yes No
2. Determining relevance:
a. Did the authors study a clinically meaningful Yes No
and/or a patient oriented outcome?
b. The patients covered by the review similar to your population Yes No
3. Determining Validity:
a. The two groups being studied are selected from Yes No
populations that are similar, except for the factor
under investigation.
This is one of the concerns, and possible sources of bias in a cohort study, and also one of the things that the propensity score matching—the form of analysis the authors undertook—sought to account for
b. The study indicates how many of the people Yes No N/A
asked to take part did so, in each of the groups being
studied.
c. The outcomes are clearly defined. Yes No
d. The assessment of outcome is made blind to Yes No N/A
exposure status
e. The measure of assessment of exposure is reliable. Yes No
f. The main potential confounders are identified and Yes No
taken into account in the design and analysis.
- List the potential confounders:
o Age
o Gender
o Diabetes
o Hypertension
o Smoking history
o Prior CV disease
o Race
o Cholesterol medication use
o Elevated total cholesterol
o Low HDL
o Waist circumference
o Low socioeconomic status
All of these possible confounders were controlled for in this study using a technique called propensity score matching.
4. What are the results?
a. What are the overall results of the study?
From the PURL:
“The primary outcome was all-cause mortality, and the secondary outcome was cardiovascular mortality. The prognostic value of fasting and nonfasting LDL for these outcomes was evaluated as the area under the receiver operator curve (ROC) using the HosmerLemeshow C-statistic.14 (In this case, similar C-statistics indicate that the tests have similar prognostic values.) Kaplan-Meier curves were used to assess survival. The association of LDL with mortality, after adjustment of potential confounders, was evaluated using Cox proportional hazard models. The groups were divided into tertiles based on LDL levels (130 mg/dL).
The risk of cardiovascular mortality also increased with increasing LDL tertiles. As was the case with all-cause mortality, the prognostic value of fasting vs nonfasting status was similar for predicting cardiovascular mortality as observed by similar C-statistics (0.64 [95% CI, 0.62-0.66] vs 0.63 [95% CI, 0.60- 0.65]; P=.49). In addition, fasting vs nonfasting C-statistics were similar for both diabetic and non-diabetic patients.”
b. Are you certain that the overall effect is due to the Yes No
exposure being investigated?
In a cohort study, even a very, very rigorously done one such as this one, there is always the possibility of some unmeasured bias—so certainty is always an impossibility (in fact, even with the best RCT in the world, or the best meta-analysis in the world, certainty is never possible).
c. Are the results statistically significant? Yes No
d. Are the results clinically significant? Yes No
5. Applying the evidence:
a. If the findings are valid and relevant, will this change
your current practice? Yes No
b. Is the change in practice something that can be done in
a medical care setting of a family physician? Yes No
c. Can the results be implemented? Yes No
d. Are there any barrier to immediate implementation? Yes No
In fact, implementing this will remove barriers to care—a nice change!
f. How was this study funded? NHANES is funded by the CDC, and this study received additional support from the New York University Cardiovascular Disease Outcomes group
6. Teaching Point
Propensity score matching is a way to reduce bias within a large cohort.
In this study, we want to know whether the fasting time correlates with mortality rate. Obviously, LOTS of other things are going to correlate with mortality rate—whether or not the participants have heart disease, diabetes, or how old they are, for example. In a propensity score matching analysis, the model takes a subset of the participants’ data and matches them on the factors of interest—in this case, it is anything that the researchers thought could be a possible contributor to mortality rate AND lipid level.
This is a confounder—something that correlates with both your “intervention” and your “outcome” (from your PICO, above), also known as your dependent and independent variables. Confounders create bias, and we do our best to avoid or control for them, because if they are not accounted for, they can make a study appear important when it isn’t, or we can miss something that is actually important. Worst of all is when we don’t even know that a confounder exists, and we interpret the study without controlling for it.
In this study, the researchers matched the two groups for a number of confounders (the list is above) that could increase or decrease lipid levels and correlate with an increase or decrease in mortality[i].
After the matching algorithm was complete, the two cohorts were statistically identical. For example, table 1 demonstrates that the original cohort of 16,000 patients had statistically more women in the nonfasting group (51.35% vs 54.07%, p=0.02). After matching, there were 51.66% vs 53.13%, (p=0.36). Similar changes can be seen in hypertension and diabetes, among others.
In exchange for matching and reducing bias, the trade-off is that not all of the 16,000 original patients are included. The cohort has been reduced to about half that size, 8598 patients. On the other hand, the baseline characteristics of these patients are identical. If fasting status matters, we should see a difference in mortality, because we have presumably accounted for the rest of our confounders. If mortality is the same, we can conclude fasting status doesn’t matter.
This study found that mortality was the same. They did several stratified analyses, looking at additional confounders, and still found that mortality was the same.
These results correspond with another recent study of 209,000 lipid panels performed in Calgary over a six month period, which found that there was almost no change in LDL with fasting. Triglycerides changed modestly with prolonged fasting, but the clinical significance of the changes were minimal.
Reference:
Sidhu, D. Naugler, C. Fasting time and lipid levels in a community based population: a cross-sectional study. Arch Intern Med. 2012;172(22):1707-1710. Available at:
http://archinte.jamanetwork.com/article.aspx?articleid=1391022
[i] For those interested in more detail into this process: The algorithm calculates a “propensity score” based on the average characteristics of the treatment group (or control group, it should work either way), and then attempts to match this score—and thus the average characteristics—by creating a subset of the other group that matches that propensity score. So each group winds up with a matched propensity score that should translate to matching characteristics. In this study, this algorithm worked perfectly, and each of the variables entered into the model (those from the list above, also found in table 1) are matched in the post-match analysis. From once the second subset is created, the outcomes can be compared to see if they are the same or different—or put another way, we know the rate of diabetes is the same because we artificially made it so, next we can examine whether the fasting status and mortality are different in the two groups.