Meta-regression using SAS

A meta-analysis that simultaneously examines multiple predictors of the relationship between two variables is called meta-regression.  It is roughly analogous to linear regression.  The IVs are moderator variables.  That is, they are the variables you believe may predict the size of the effect.  The DV is the effects sizes that you are meta-analyzing. 

 

Overview

The general approach is to

1)     obtain the studies (not discussed here)

2)     calculate the effect sizes and variances by hand (not discussed here),

3)     enter the data into SAS,

4)     conduct the analyses without predictor variables, to estimate the overall effect,

5)     examine whether there is significant heterogeneity of the effect (not discussed here), and

6)     examine moderators. 

 

Calculate effect sizes

The DV is the effect size that characterizes the relationship of X and Y.  Values of the DV are usually z (the Fisher transform of r), the log-odds (the natural log of the odds ratio) or d.  The variance of each effect size is calculated outside of SAS in the usual way. 

 

Data entry

To enter the data into SAS, follow the format below. 

 

DATA test1;

INPUT trial ln_or      est   year  year_m  cohort  design  country  mos;

CARDS;

1     0.348254219 0.121887759 2003  3     1     1     2     2

3     0.055663868 0.007594335 1999  -1    0     1     1     1

4     -0.968310623      0.000455847 2003  3     0     1     3     1

8     -0.199289901      0.001559349 2002  2     0     1     2     1

9     -0.768254653      0.078953618 2003  3     0     1     2     1

32    0.284104251 0.199718382 1998  -2    0     1     2     2

33    0.071692745 0.020341215 2001  1     0     1     2     1

39    0.435497395 0.067556427 1991  -9    0     1     1     2

40    -0.947515921      0.019046944 1996  -4    0     1     3     1

42    0.445625551 0.040364599 2000  0     0     0     1     2

46    0.249910566 0.001183328 2003  3     0     1     1     1

; run;

 

trial is just an arbitrary name for the study.  BUT, the list of studies must be sorted in ascending numerical order, otherwise SAS messes up the stats.   This is an undocumented bug.

 

ln_or is the name for the effect size (but you can choose any name).  In this case, it is the log odds.

 

est is the variance for the effect size.  Note that this is, in essence, the inverse of the sample size.  This variable must be called est. 

 

Data analysis – unadjusted effect size

The syntax below will yield an estimate of the pooled effect size, log odds in this example.  This is a fixed effects model.

 

PROC MIXED method = ml data=test1;

class trial;

model ln_or = /s cl;

repeated / group = trial;

parms/ parmsdata=test1 eqcons = 1 to 11;

run;

 

s requests an estimate of the pooled effect size and cl asks for confidence intervals. 

The parms command tells SAS where to get the variances; it looks for the variable named est

 

The output:

Effect       Estimate       Error      DF    t Value    Pr > |t|     Alpha       Lower       Upper

 

Intercept     -0.5197     0.01585      10     -32.78      <.0001      0.05     -0.5550     -0.4843

 

The intercept is the estimate of the pooled effect size.  Thus, the pooled log odds is -.52 and is significantly different from zero.   (Recall that the natural log of OR = 1 is zero.)

 

Data analysis – adjusted effect size

 

If you want to conduct a meta-regression, then start adding predictors to the model.  To examine moderators, one must have heterogeneity of variance, a topic not addressed here but easily understood from relevant texts on meta-analysis. 

 

Note that the predictors will start changing the intercept and you have to think carefully about whether the intercept will still be interpretable.  For example, using  continuous covariate such as year of study will then estimate the intercept holding year constant at zero (a useless estimate).  So mean centering some variables or other transformations may make sense.

 

Another example is using a categorical covariate.  The intercept will now be estimated for the highest level of the categorical covariate.  Next, SAS will estimate the relative effect size for the other levels of the IV, how many times larger or smaller the other effect sizes are than the referent. 

 

PROC MIXED method = ml data=test1;

class trial country;

model ln_or = country mos /s cl;

repeated / group = trial;

parms/ parmsdata=test1 eqcons = 1 to 11;

run;

 

 

                                  Type 3 Tests of Fixed Effects

 

                                        Num     Den

                          Effect         DF      DF    F Value    Pr > F

                          country         2       7     532.87    <.0001

                          mos             1       7       4.39    0.0743

 

  Effect     country  Estimate     Error    DF  t Value  Pr > |t|   Alpha     Lower     Upper

   Intercept            -1.2613    0.1416     7    -8.91    <.0001    0.05   -1.5961   -0.9265

   country    1          1.1886   0.03819     7    31.12    <.0001    0.05    1.0983    1.2789

   country    2          0.7813   0.04299     7    18.18    <.0001    0.05    0.6797    0.8830

   country    3               0         .     .      .       .           .         .         .

   mos                   0.2935    0.1400     7     2.10    0.0743    0.05  -0.03762    0.6246

 

 

In this output, country moderates the effect size but mos does not.  The estimate of the log odds is -1.2613 for country 3 (Canada), estimated where mos = 0.  The odds ratio for country 1 is exp(1.1886) or 3.3 times larger than for country 3, holding mos = 0. 

 

For the log-odds ratio, multiply the log (estimate) for the intercept by the log (estimate) for any level of a categorical covariate to find the estimate of the OR for that level.

 

For z, add the estimate for the intercept to the estimate) for any level of a categorical covariate to find the estimate of the z for that level.

 

--------------------------------------

                

Most statisticians who write on this topic argue strongly for moving beyond fixed effects models, as shown above, to random effects models. 

  

--------------------------------------

                

References

 

Easy to read intro material:

Wolf, F.M. (1986).  Meta-Analysis: Quantitative Methods for Research Synthesis. Beverly Hills, California: SAGE Publications.

 

A bit harder but definitive:

Cooper, H., & Hedges L. V..  (1994). The handbook of research synthesis. New York: Russell Sage Foundation.  Look at the chapter by Rosenthal, by Fleiss, and by Shadish and Haddock. 

 

Over the top unless you have a serious stats background, but it gives relatively clear SAS code:

Normand, S. L. (1999).  Meta-analysis: formulating, evaluating, combining, and reporting.  Statistics in Medicine, 18(3), 321-59.

van Houwelingen, H. C., Arends, L. R., & Stijnen, T. (2002).  Advanced methods in meta-analysis: multivariate approach and meta-regression.  Statistics in Medicine, 21(4), 589-624.

 

 

 

Back to Noel Brewer’s Home Page