Thoughts: Hazards of dependence on surrogate endpoints

Thoughts: Hazards of dependence on surrogate endpoints


HK: Professor Yudkin et al. (The idolatry of the surrogate (2011)) offer a very important insight into the use of surrogate endpoints in medical science.   They argue that this practice - so widespread, especially in diabetes care - is tantamount to losing sight of the patient's interest.   It's a very readable piece -- a challenge to much of our work in public health & chronic-disease epidemiology.   It adds to the growing chorus of criticisms directed at the "guidelines" we (used to) cherish.

YC:  I like to read this kind of article.We are fool only when we think surrogate is a real/only cause of the disease. I agree that the hard event is the most important for evaluation of a treatment. However, the authors seem ignoring the whole spectrum of development of science, which is from short range (quick surrogate) to long range (hard event). If a drug even cannot control the level of meaningful surrogate such as glucose and blood pressure, I don’t think that the drug is good for controlling of the hard event beyond the surrogate.


BTW, the example of rosiglitazone exactly demonstrates that the scientists or physicians (maybe I need say good scientists ^_^) didn’t forget the target beyond the glucose and stop digging up more evidences. Anyhow, this is a heads-up for someone.

Selected References of The idolatry of the surrogate (2011):

• ebook: Institute of Medicin (2010).Evaluation of Biomarkers and Surrogate Endpoints in Chronic Disease.
• Yudkin JS, Richter B, Gale EAM. Intensified glucose lowering in type 2 diabetes: time for a reappraisal (2010); Intensified glucose control in type 2 diabetes—whose agenda? Lancet (2011)
• Emerging Risk Factors Collaboration (2010). C-reactive protein concentration and risk of coronary heart disease, stroke, and mortality.
• O’Connor PJ, Bodkin NL, Fradkin J, Glasgow RE, Greenfield S, Gregg E, et al (2011). Diabetes performance measures: current status and future directions.
• Lehman R, Yudkin JS and Krumholz HM (2010). Licensing drugs for diabetes.

The Fat Trap - NYTimes.com

The Fat Trap
Source: NYTimes.com
For 15 years, Joseph Proietto has been helping people lose weight. When these obese patients arrive at his weight-loss clinic in Australia, they are determined to slim down. And most of the time, he says, they do just that, sticking to the clinic's program and dropping excess pounds. But then, almost without exception, the weight begins to creep back. In a matter of months or years, the entire effort has come undone, and the patient is fat again. "It has always seemed strange to me," says Proietto, who is a physician at the University of Melbourne. "These are people who are very motivated to lose weight, who achieve weight loss most of the time without too much trouble and yet, inevitably, gradually, they regain the weight." ...
Full text: here

Obesity

Obesity
Source: TheLancet.com Published August 26, 2011
"This four-part Series critically examines what we know about the global obesity pandemic: its drivers, its economic and health burden, the physiology behind weight control and maintenance, and what science tells us about the kind of actions that are needed to change our obesogenic environment and reverse the current tsunami of risk factors for chronic diseases in future generations."


"The first paper looks at the global drivers of the epidemic; the second paper analyses obesity trends in the USA and UK, and their impact on prevalence of diseases and healthcare spending. The third paper introduces a new web-based bodyweight simulation model, that incorporates metabolic adaptations that occur with decreasing bodyweight; and the final paper assesses the interventions needed to halt and reverse the epidemic. Its authors conclude that the changes needed are likely to require many sustained interventions at several levels, but that national governments should take the lead. "
Full Text: Here

How to get the CPS for the NHANES?

Tips - NHANES: Where and how to get the CPS population for the NHANES?

The Current Population Survey (CPS) is a monthly survey of about 50,000 households conducted by the Bureau of the Census for the Bureau of Labor Statistics. The survey has been conducted for more than 50 years.

The NCHS of CDC used the the civilian noninstitutionalized U.S. population information of CPS at specific time point for the post-stratification to match the population control totals for each sampling subdomain, usually the post-stratification structure defined by age, sex, and race/ethnicity. For NHANES III, the structure has 12 age groups, 2 sex groups, and 4 race groups (Non-Hispanic black, Non-Hispanic black, Mexican American, and Other); this means there are 96 cells. You can find the response rate and CPS population for different surveys here. or you can find it from original NHANES demographic data using the same age, sex, and race/ethnicity structure and interview weights [for example, Stata: .svy: tab agesexracegrp, count obs format(%12.0f)]. Theoretically, to a picky epidemiologist/statistician, CPS population information is important to get correct national total estimates, dealing with missing data (reweight), or doing bootstrap analysis of complex survey data.

Meta-Analysis for Linear and Nonlinear Dose-Response Relations: Examples, an Evaluation of Approximations, and Software

Meta-Analysis for Linear and Nonlinear Dose-Response Relations: Examples, an Evaluation of Approximations, and Software
Two methods for point and interval estimation of relative risk for log-linear exposure-response relations in meta-analyses of published ordinal categorical exposure-response data have been proposed by Nicola Orsini (Stata and SAS code), who is the author of Stata ado of -glst-.

Circulation's Diabetes Mellitus Studies

Circulation's Diabetes Mellitus Studies 2009 - 2010
"The following articles are being highlighted as part of Circulation's Topic Review series. This series will summarize the most important manuscripts, as selected by the editors, published in Circulation and the Circulation subspecialty journals. The studies included in this article represent the articles related to diabetes mellitus that were published in Circulation in 2009 and 2010. ..."
full text: here

Piece-wise Regression

Tips: Piecewise/Segmented Regression Related

• Ryan SE, Porth LS (2007). A tutorial on the piecewise regression approach applied to bedload transport data.(pdf) A very good tutorial article from U.S. Forest Service.
• Nonlinear relationships: 
• additivity vs. non-additivity, linearity vs. non-linearity.
• a few types of non-linearity modeling: polynomial models, exponential models, piecewise regression models 
• Example(Stata vs. SAS): If we are looking for the relation of AGE and BMI. Visually there is a reflection/change point/break point at age around 65. We may create two regression: BMI=a1 + b1*AGE for persons age<65, and BMI=a2 + b2*AGE for persons age >= 65. To make the regression continuous at the reflection point: a1 + b1*(age = 65) = a2 + b2*(age = 65), so a2 = a1 + (age=65)*(b1 - b2).
• Stata: .nl (BMI = cond(AGE < {k}, {a1} + {b1}*AGE, {a1} + {k}*({b1} - {b2}) + {b2}*AGE)), initial(a1 1 b1 1 b2 1 k 60) // here k = reflection point of age,  {} = name of expected parameters of the model.
• SAS: PROC NLIN; PARMS a1=1 b1=1 b2=1 k=60; IF AGETHEN DO; MODEL BMI=a1 + b1*AGE; END; ELSE IF AGE>=k THEN DO; MODEL BMI=a1 + k*(b1 - b2) + b2*AGE; END; RUN;

• Stata Commands: -nl-, -svy nl-, -ml-
• Stata Commands:  -mkspline- creates variables containing a linear spline or a restricted cubic spline of a variable, and  makes it easy to create the new variables for piecewise regression models.
• Example 1 - continuous dependent variable:
• .use http://www.stata-press.com/data/r12/mksp1
• .mkspline age1 20 age2 30 age3 40 age4 50 age5 60 age6 = age, marginal
• .regress lninc educ age1-age6, vce(robust) //with knots only
• .gen agegt20=(age>=20)
• .regress lninc educ age1 age2 agegt20, vce(robust) //with one knots and one jump
• Example 2 - dichotomous dependent variable: 
• .use http://www.stata-press.com/data/r12/mksp2
• .mkspline dose 5 = dosage, displayknots
• .logistic outcome dose1-dose5
• Stata: How can I find where to split a piecewise regression?
• Stata: How can I run a piecewise regression in Stata?
• Stata STB: Piece-wise exponential survival models.
• Stata STB: The piecewise linear spline transformation 27-29(pdf).
• Stata Journal: Multivariable modeling with cubic regression splines: A principled approach (pdf)
• Stata/SPSS: Nonlinear relationships (pdf)
• Stata FAQs: Nonlinear regression; How can I use Stata to solve a system of nonlinear equations?
• Stata FAQs: How do I estimate a nonlinear model using -ml-? 
• Stata (UCLA): Simple linear and nonlinear models using Stata's -ml- command.
• I think Chow Test and Generalized Additive Model (GAM) [Wikipedia, StatSoft, an article of Hastie and Tibshirani(pdf)] have the similar concepts.
• Software: Stata (GAM), SAS (PROC GAM, tutorial (pdf)), R (gam).
• R!: R for Ecologists: Putting Together a Piecewise Regression (2012)
• SAS: A tutorial on the piecewise regression approach applied to bedload transport data (pdf)
• SAS(Rick Wicklin):Segmented regression models in SAS
• SAS(Rick Wicklin): Nonsmooth models and spline effects, Shared Concepts and Topics: EFFECT, ESTIMATE
• NCI: Joinpoint Regression Software

Doing bootstrap/jackknife in Stata

Doing Repeated Replication Methods (Bootstrap/Jackknife) for complex survey data

• Bootstrap sampling and estimation, Survey data analysis in Stata
• Starting from Stata 10, you can just use -svy jackknife- instead of creating jackknife weight at first.
• By using information on PSUs and strata, -svy jackknife:- will automatically adjust the sampling weights to produce the replicates using the delete-1 jackknife method
• if you want do delete-k jackknife, you need provide the replicate weight variables using -svyset-
• by default jackknife variance is computed by using deviations of the replicates from their mean. If you want to variance from the observed value of statistics based on the entire dataset, you need use -svy jackknife, mse- (mean square error)option. The -mse- method providea larger variance estimation because of the addition of the familiar sqared bias tern in the mean square error
• The Stata 14 has "strata()" and "cluster()" options and sounds like for the complex survey data, but seemingly it cannot deal with the sampling weight correctly.
• for example: bootstrap "stata_program" _b, strata(strata) cluster(psu) reps(1000) // no weight in the statement
• I tend to avoid to use this bootstrap command for complex survey data
• I will use -bsweights- creates the bootstrap resampling weights then used the weights by -svyset-, -svy bootstrap- or -bs4rw-. Resampling variance estimation for complex survey data - The Stata Journal. -bsweights- was written by Stas Kolenikov (Survey bootstrap and bootstrap weights).
• or, SURVWGT: a Stata module to create and manipulate survey weights by Nickolas Winter
• Jackknife for simple random sampling data:
• jknife r(mean): summarize mpg
• Jackknife for complex survey data: 
• webuse nhanes2, replace
• svyset psu [pw=finalwgt], strata(strata)
• svy jackknife slope=_b[height] constant=_b[_cons]: regress weight height
• Jackknife for complex survey data using a user-written program:
• svy jackknfie _b[, options]: intcens
• User-written program using Jackknife for complex survey data.
• Notes: 
• -svy jackknife- is allowed as long as they follow standard Stata -syntax-, allow the 
-if- qualifier, and allow -pweights- and -iweights-.
• Notes: "anything", "namelist", "name", "weight, "if", "in", "varlist", "using", "exp", etc. are special macros. "anything" is used to tell the -syntax- command what can appear immediately after the name of the command. "anything" could be anything like SILLY in this example, which passes an arguement into the program. Things between "[" and "]" are optional. In order to use -margins-, -set buildfvinfo- needs set as -on-.
• Here is a modified example in the manual: with replication-based variance estimators
• program mymargins, eclass
    syntax anything [if] [iw pw]
    if "`weight'" != "" {
      local wgtexp "[`weight' `exp']"
    }
    set buildfvinfo on
    `anything' `if' `wgtexp'
    margins race, post
  end
  global myanything "logistic highbp height weight i.race c.age##c.age" //!!! using `anything' with caution !!!
  svy jackknife _b: mymargins $myanything
• UCLA: How do I write my own bootstrap program?
• SSCC: Bootstrapping in Stata
• Stata Journal(2003): Bootstrapped standard errors
• Schmidheiny(2016): The Bootstrap
• How can I analyze multiple mediators in Stata? 
• Stata Journal(2004): From the help desk: Some bootstrapping techniques
• 'svr' is a module/package to compute estimates with survey replication (SVR) based standard errors written by Nick Winter. -survwgt-, one of commands of 'svr' creates sets of replicate weights for complex sampling data including balanced repeated replication (BRR) and several version of the survey jackknife (JK*). In addition, survwgt performs oststratification, raking, and non-response adjustments to survey weights. Starting from Stata 10, you can just use -svy jackknife- instead of creating jackknife weight at first.
• The jackknife (n-1) estimate of the standard error is equal to , where n is the total number of observations (or clusters). The factor in the jackknife's standard error is about n times (inflation factor) larger than bootstrap [1/(B-1)].
• The delete-d jackknife estimate of the standard error is using (n-d)/C(n,d) instead of (n-1)/n
• Efron (1981). Nonparametric Estimates of Standard Error: The Jackknife, the Bootstrap and OtherMethods
• McIntosh."The Jackknife Estimation Method"
• UCLA:How can I sample from a dataset with frequency weights?

Journal Article - Heart Disease and Stroke Statistics—2012 Update from the AHA

Journal Article - Heart Disease and Stroke Statistics—2012 Update from the AHA

"Each year, the American Heart Association (AHA), in conjunction with the Centers for Disease Control and Prevention, the National Institutes of Health, and other government agencies, brings together the most up-to-date statistics on heart disease, stroke, other vascular diseases, and their risk factors and presents them in its Heart Disease and Stroke Statistical Update. The Statistical Update is a valuable resource for researchers, clinicians, healthcare policy makers, media professionals, the lay public, and many others who seek the best national data available on disease morbidity and mortality and the risks, quality of care, medical procedures and operations, and costs associated with the management of these diseases in a single document." … 

Full text: here (pdf)

Is a Chow test the correct test to determine whether data can be pooled together?

Tips - Stata: Is a Chow test the correct test to determine whether data can be pooled together?

Source: Stata FAQs by Willam Gould

A Chow test is simply a test of whether the coefficients estimated over one group of the data are equal to the coefficients estimated over another, and you would be better off to forget the word Chow and remember that definition.

History: In the days when statistical packages were not as sophisticated as they are now, testing whether coefficients were equal was not so easy. You had to write your own program, typically in FORTRAN. Chow showed a way you could perform a Wald test based on statistics that were commonly reported, and that would produce the same result as if you performed the Wald test.

Full text: here

Other related articles about Chow test:

• Stata: How can I compute the Chow test statistic? (Chow tests are the same as the Wald test "test" command reports)
• Stata: Can you explain Chow tests?
• Chow test - Wikipedia
• Stata: In Stata, how do I conduct the Chow Test?
• Gregory C. Chow (1960). "Tests of Equality Between Sets of Coefficients in Two Linear Regressions"
• My thoughts: I use 'suest' and 'Chow' test to test differences of coefficients. [Stata 12 help - Contents; the Stata Journal archives]

Role of Environmental Chemicals in Diabetes 1 and Obesity

Role of Environmental Chemicals in Diabetes 1 and Obesity: A 2 National Toxicology Program Workshop Report

By Kristina A. Thayer, Jerry Heindel, John R. Bucher, Michael A. Gallo 

full text: here (pdf)

CHE: Practice Prevention Columns and Fact Sheets 

CHE-Alaska: The Role of Environmental Chemicals in the Development of Diabetes and Obesity 

Tips - Stata: How do I fit a linear regression with interval (inequality) constraints in Stata?

Tips - Stata: How do I fit a linear regression with interval (inequality) constraints in Stata?

Source: Stata FAQs by Isabel Canette

If you need to fit a linear model with linear constraints, you can use the Stata command cnsreg. If you need to fit a nonlinear model with interval constraints, you can use the -ml- command, as explained at http://www.stata.com/support/faqs/stat/intconst.html. However, if you have a linear regression, the simplest way to include these kinds of constraints is by using the -nl- command.

 …

 Full text: here

Tips: Stata - Is a Chow test the correct test to determine whether data can be pooled together?

Tips - Stata: Is a Chow test the correct test to determine whether data can be pooled together?

Source: Stata FAQs by Willam Gould

A Chow test is simply a test of whether the coefficients estimated over one group of the data are equal to the coefficients estimated over another, and you would be better off to forget the word Chow and remember that definition.

History: In the days when statistical packages were not as sophisticated as they are now, testing whether coefficients were equal was not so easy. You had to write your own program, typically in FORTRAN. Chow showed a way you could perform a Wald test based on statistics that were commonly reported, and that would produce the same result as if you performed the Wald test.

Full text: here

Other related article:

How can I compute the Chow test statistic? http://www.stata.com/support/faqs/stat/chow.html

Tips - Stata: a few useful ado-related commands

Tips - Stata: the handy Stata command/function

• 'statsby', 'tabstat', 'scalar themean=r(mean)'.
• 'contrast'(How to get orthogonal polynomial coefficient), 'pwmean', 'pwcompare', and 'margins'.
• p. and q. of 'Contrast' (Orthogonal polynomials) allow to partition the effects of a factor variable into linear, quadratic, cubic, and higher-order polynomial components (I like to use p., q. assumes having equal space between groups). They are only meaningful with factor variables that have a natural ordering in the levels. For exmaple: .contrast p(2 3 4).bmigrp, noeffects
• User defined contrast of race(3 levels) and age (2 levels) without comparing the middle race group: .contrast {race#age -1 -1 0 0 1 1}
• 'destring', 'tostring','string()': Convdrting between numeric variable and string/character variable.
• 'duplicate': Report, tag, or drop duplicate observations.
• 'postfile' posts results in Stata dataset.
• ... [Contents of Stata Help]

Tips - Stata: outputting/exporting Stata results

Tips - Stata: outputting/exporting Stata results

• 'outreg2' by Roy Wada: The 'outreg2' is based on the earlier package 'outreg' but has much more features. 'outreg2' related programs include 'seeout', 'shellout', 'logout', and 'mkest'. You can use command 'ssc install [prgname]' to install these programs. For example,
• .cd \temp
• .svy, subpop(pop):logit diagdm i.survey##i.age3grp
• .margins survey##age3grp,subpop(pop) vce(svy) post
• .outreg2 using demo, stats(coef ci) paren(ci) dec(4) noaster sideway seeout replace
• 'tabout', by Ian Watson: is a table building program for oneway and twoway tables of frequencies and percentages, and for summary tables. The output from tabout matches Stata's tabulate, most of tabstat and some of table. When the 'svy' option is used, you can also get (se ci lb ub) for survey data.
• 'estout', by Ben Jann: How can I use -estout- to make regression tables that look like those in journal articles?
• 'estout' vs. 'outreg' and 'outreg2', by Roy Wada.
• 'logout', by Roy Wada: provides a fast and easy way to convert log or ASCII files into various output formats compatible with Word, Excel, LaTeX, or Stata datafile.
• 'outreg' by John Gallup: Can I make regression tables that look like those in journal articles?
• Making publication-style tables in STATA (pdf)
• How do I export tables from Stata? 
• UCLA Academic Technology Services has provided much richer information on how to use Stata and other statistic software.
• Resources for learning Stata