Frequently Asked Questions

How are standardized mean differences computed in cobalt?

Most questions below are related to this one, so here I will try to explain in complete detail how standardized mean differences (SMDs) are computed.

First, it is important to know that by default, mean differences for binary covariates are not standardized. That means in the Diff.Adj column, etc., what you are seeing the is raw difference in proportion for binary variables. (By raw, I mean unstandardized, but weights may be applied if relevant.) To request SMDs for binary covariates, set binary = "std" in the call to bal.tab() or love.plot(). See also the question below.

For continuous covariates, the standardized mean difference is computed as \[ \text{SMD} = \frac{\bar{x}_1 - \bar{x}_0}{s^*} \] where \(\bar{x}_1\) is the mean of the covariate in the treated group, \(\bar{x}_0\) is the mean of the covariate in the control group, and \(s^*\) is a standardization factor (not necessarily a standard deviation!). After matching or weighting, the weighted standardized mean difference is computed as \[ \text{SMD}^w = \frac{\bar{x}_{1w} - \bar{x}_{0w}}{s^*} \] where \(\bar{x}_{1w}\) is the weighted mean of the covariate in the treated group, i.e., \(\bar{x}_{1w} = \frac{1}{\sum_{i:A_i = 1}{w_i}}\sum_{i:A_i=1}{w_ix_i}\), and similarly for the control group. Critically, the standardization factor \(s^*\) is the same before and after weighting. I will repeat, the standardization factor \(s^*\) is the same before and after weighting. I don’t mean it has the same formula, it mean it is literally the same value. I explain in more detail in a question below why this is the case.

How is the standardization factor computed? This depends on the argument to s.d.denom supplied to bal.tab() or love.plot(). When s.d.denom is not supplied, this is determined by the argument supplied to estimand, and when that is not supplied, the estimand is guess based on the form of the weights, if any. By default, with no weights supplied and no argument to s.d.denom or estimand, s.d.denom is set to "pooled", and a note will appear saying so. That note doesn’t appear if weights are supplied or balance is assessed on the output of a another package, as the estimand, and therefore s.d.denom, can be determined automatically.

Below are the formulas for the standardization factor corresponding to each value of s.d.denom:

where \(s_1\) is the standard deviation of the treated group, \(s_0\) is the standard deviation of the control group, \(s\) is the standard deviation of the whole sample ignoring treatment group membership, \(s_w\) is the weighted standard deviation of the whole sample, \(n_1\) and \(n_0\) are sizes of the treated and control groups, respectively, and \(n = n_1 + n_0\).1 For continuous covariates, the unweighted standard deviation is computed as usual, i.e., as \[ s = \sqrt{\frac{1}{n-1}\sum_i{(x_i - \bar{x})^2}} \] and the weighted standard deviation is computed as \[ s = \sqrt{\frac{\sum_{i} w_{i}}{(\sum_{i} w_{i})^2 - \sum_{i=1}^{n} w^2_{i}}\sum_i{w_i(x_i - \bar{x}_w)^2}} \]For binary covariates, the unweighted standard deviation is computed as \(s = \sqrt{\bar{x}(1-\bar{x})}\) and the weighted standard deviation is computed as \(s = \sqrt{\bar{x}_w(1-\bar{x}_w)}\).

When sampling weights are supplied, all standard deviations in the standardization factor are computed incorporating the sampling weights. When s.d.denom = "weighted", the standardization factor is computed using the weights used to balance the sample (i.e., the matching or weighting weights), even for the unadjusted sample. Remember, the standardization is ALWAYS the same before and after adjustment.

I know some of these formulas seem overly complicated for such simple statistics, but they are required to keep things consistent and not dependent on the scale of the weights.

Why are mean differences not standardized for binary covariates?

Ultimately, bias in the treatment effect estimate is a function of imbalance. That bias is indifferent to whether you measure that imbalance using a standardized or unstandardized mean difference. The reason we use SMDs is that covariates naturally are on a variety of different scales, and when trying to quickly assess whether a sample is balanced, it is productive to unify the scales of the covariates. That way, balance on a covariate measured with large numbers (e.g., days in hospital or prior earnings in dollars) can be assessed alongside balance on a covariate measured with small numbers (e.g., number of comorbidities or years of education).

With binary covariates, though, they are already on a comprehensible scale, so there is no need to standardize. In addition, the scale is intuitive for people; a difference in proportion of .1 when both groups have 100 people means that there is an imbalance of 10 people on the covariate. Are 10 people being different enough to cause bias in the estimate? That can be assessed substantively without needing to take the additional step of translating the variable’s scale into something meaningful.

Another important reason why mean difference are not standardized is that it is possible for two covariates with the same imbalance to have vastly different mean differences. For example, consider the following dataset. X1 and X2 both have a mean difference of .1; if they both affected the outcome equally, then each would contribute to the bias in the estimate to the same extent.

treat <- rep(1:0, each = 20)
X1 <- c(rep(0:1, c(1, 19)), rep(0:1, c(3, 17)))
X2 <- c(rep(0:1, c(9, 11)), rep(0:1, c(11, 9)))

bal.tab(treat ~ X1 + X2,
        binary = "raw",
        disp = "means",
        s.d.denom = "treated")
#> Balance Measures
#>      Type M.0.Un M.1.Un Diff.Un
#> X1 Binary   0.85   0.95     0.1
#> X2 Binary   0.45   0.55     0.1
#> 
#> Sample sizes
#>     Control Treated
#> All      20      20

But if we standardized the mean differences, not only do we move away from an actually interpretable statistic (i.e., what does it mean to divide by the standard deviation of a binary variable?), we see that the standardized mean difference vary by a huge amount, with X1 having twice the imbalance of X2.

bal.tab(treat ~ X1 + X2,
        binary = "std",
        s.d.denom = "treated")
#> Balance Measures
#>      Type Diff.Un
#> X1 Binary  0.4588
#> X2 Binary  0.2010
#> 
#> Sample sizes
#>     Control Treated
#> All      20      20

Why does this happen? The standard deviation of a binary variable is a function of its mean (in particular, it is \(s = \sqrt{p(1-p)}\)) where \(p\) is the mean of the variable). That means information about the mean of the variable, which is unrelated to imbalance, contaminates the standardized mean difference, which is supposed to measure imbalance. In this case, standardizing the mean difference only adds confusion and reduces interpretability. That is why, by default, mean differences for binary variables are unstandardized by default.

You can always change this by setting binary = "std" in the call to bal.tab() or setting set.cobalt.options(binary = "std") to change the option for the whole session. One advantage of using standardized mean differences for binary variables is that they are always larger than the raw mean difference (because the standardization factor is always less than 1), which means if you use the standardized mean difference as your balance criterion, you will always seek better balance than using the raw mean differences. The balance statistics computed by bal.compute() that involve the standardized mean difference standardize all variables, including binary variables.

Why do you use the same standardization factor before and after adjustment?

It is important to remember that bias is a function of the difference in means of a covariate and standardization is a tool to aid in balance assessment. As a tool, it should reflect imbalance accurately (i.e., without incorporating extraneous information), but there is no statistical “truth” about the nature of the standardization factor. I use the same standardization factor before and after adjustment as recommended by Stuart (2008). The rationale is that by isolating the SMD to reflect changes in the difference in means, why can more accurately assess improvement in balance rather than combining information about the difference in means with information about the variability of the covariate, which may change in a variety of ways after adjustment. I describe a specific example of how allowing the standardization factor to change can cause problems here.

How do I extract the balance tables from the bal.tab() object?

The output of a call to bal.tab() is a bal.tab object, which has several components depending on the features of the dataset (e.g., whether the data are multiply imputed or clustered or whether the treatment is binary or multi-category, etc.). In the most basic case, a binary or continuous treatment with no clustering, no multiple imputations, a single time point, and subclassification is not used, the balance table is stored in the Balance component of the output object. Let’s take a look:

data("lalonde")

b <- bal.tab(treat ~ age + educ + race + married + re74,
             data = lalonde, s.d.denom = "treated",
             disp = "means", stats = c("m", "v"))

# View the structure of the object
str(b, give.attr = FALSE)
#> List of 2
#>  $ Balance     :'data.frame':    7 obs. of  9 variables:
#>   ..$ Type       : chr [1:7] "Contin." "Contin." "Binary" "Binary" ...
#>   ..$ M.0.Un     : num [1:7] 28.03 10.235 0.203 0.142 0.655 ...
#>   ..$ M.1.Un     : num [1:7] 25.8162 10.3459 0.8432 0.0595 0.0973 ...
#>   ..$ Diff.Un    : num [1:7] -0.3094 0.055 0.6404 -0.0827 -0.5577 ...
#>   ..$ V.Ratio.Un : num [1:7] 0.44 0.496 NA NA NA ...
#>   ..$ M.0.Adj    : num [1:7] NA NA NA NA NA NA NA
#>   ..$ M.1.Adj    : num [1:7] NA NA NA NA NA NA NA
#>   ..$ Diff.Adj   : num [1:7] NA NA NA NA NA NA NA
#>   ..$ V.Ratio.Adj: num [1:7] NA NA NA NA NA NA NA
#>  $ Observations:'data.frame':    1 obs. of  2 variables:
#>   ..$ Control: num 429
#>   ..$ Treated: num 185

b$Balance
#>                Type       M.0.Un       M.1.Un     Diff.Un V.Ratio.Un M.0.Adj
#> age         Contin.   28.0303030 2.581622e+01 -0.30944526  0.4399955      NA
#> educ        Contin.   10.2354312 1.034595e+01  0.05496466  0.4958934      NA
#> race_black   Binary    0.2027972 8.432432e-01  0.64044604         NA      NA
#> race_hispan  Binary    0.1421911 5.945946e-02 -0.08273168         NA      NA
#> race_white   Binary    0.6550117 9.729730e-02 -0.55771436         NA      NA
#> married      Binary    0.5128205 1.891892e-01 -0.32363132         NA      NA
#> re74        Contin. 5619.2365064 2.095574e+03 -0.72108381  0.5181285      NA
#>             M.1.Adj Diff.Adj V.Ratio.Adj
#> age              NA       NA          NA
#> educ             NA       NA          NA
#> race_black       NA       NA          NA
#> race_hispan      NA       NA          NA
#> race_white       NA       NA          NA
#> married          NA       NA          NA
#> re74             NA       NA          NA

It’s not a very pretty object, which is why the print() method makes it look nicer. If you are willing to process this table yourself, you can easily extract it from the bal.tab() output and do what you want with it, e.g., saving it to a CSV file or making a pretty table using another package. Although I have been working on a way to do this more easily (i.e., to create a publication-ready table from a bal.tab object), it might be a while because the main purpose of this package is balance assessment, not formatting for publication (although I did put a lot of work into love.plot() to make it customizable for publication).

How are balance statistics computed when using subclassification?

Subclassification involves creating strata (usually based on the propensity score), within which covariates are ideally balanced. bal.tab() lets you assess balance both within and across subclasses.

One must always remember that the standardized mean difference uses the standardization factor computed in the original sample, i.e., prior to subclassification. Let’s take a look below using MatchIt:

# PS Subclassification
msub <- MatchIt::matchit(treat ~ age + educ + race + married + re74,
                         data = lalonde, method = "subclass",
                         estimand = "ATE", min.n = 4)

# Balance in the first subclass
bal.tab(msub, which.sub = 1, binary = "std")
#> Balance by subclass
#>  - - - Subclass 1 - - - 
#>                 Type Diff.Adj
#> distance    Distance   0.1574
#> age          Contin.  -1.0433
#> educ         Contin.  -0.2759
#> race_black    Binary   0.0000
#> race_hispan   Binary   0.0000
#> race_white    Binary   0.0000
#> married       Binary  -1.1135
#> re74         Contin.  -1.8353

Let’s see where the number -1.0433 came from (the standardized mean difference for age). We compute the mean of age in each treatment group in subclass 1, and then divide it by the pooled standard deviation of age (because we requested the ATE) in the original sample.

m0 <- mean(lalonde$age[lalonde$treat == 0 & msub$subclass == 1])
m1 <- mean(lalonde$age[lalonde$treat == 1 & msub$subclass == 1])

s0 <- sd(lalonde$age[lalonde$treat == 0])
s1 <- sd(lalonde$age[lalonde$treat == 1])

(m1 - m0) / sqrt((s1^2 + s0^2) / 2)
#> [1] -1.043294

A common mistake is to compute the standard deviation within each subclass. There are a few reasons why this is bad: 1) it suffers from the same problem that changing the standardization factor does with matching or weighting, i.e., that balance can appear to be worse because the standardization factor shrank even as the means got closer together; 2) when there is no or little variation of a covariate within a subclass, which is desirable, the standardization factor will be tiny, making the SMD potentially appear huge; and 3) the same variable will use different standardization factors across subclasses, which means the same difference in means, which contribute to bias equally, will have different balance statistics.

A related question is how the balance statistics are computed across subclasses to compute an overall balance statistic for the sample. For (standardized) mean differences, it is as easy as computing the average of the statistic across subclasses, where the statistics are weighted corresponding to the number of units in the subclass in the target group (e.g., the treated units for the ATT, all units for the ATE, etc.). Below I’ll demonstrate how to do that manually for the age covariate:

# SMDs across subclasses for age
smds <- sapply(1:6, function(s) {
    m0 <- mean(lalonde$age[lalonde$treat == 0 & msub$subclass == s])
    m1 <- mean(lalonde$age[lalonde$treat == 1 & msub$subclass == s])
    
    s0 <- sd(lalonde$age[lalonde$treat == 0])
    s1 <- sd(lalonde$age[lalonde$treat == 1])
    
    (m1 - m0) / sqrt((s1^2 + s0^2) / 2)
})

# Sample size in each subclass
ns <- table(msub$subclass)

# Summary SMD for age
weighted.mean(smds, ns)
#> [1] -0.2354095

bal.tab(msub)
#> Balance measures across subclasses
#>                 Type Diff.Adj
#> distance    Distance   0.1081
#> age          Contin.  -0.2354
#> educ         Contin.   0.0075
#> race_black    Binary   0.0535
#> race_hispan   Binary  -0.0420
#> race_white    Binary  -0.0115
#> married       Binary  -0.1160
#> re74         Contin.  -0.3200
#> 
#> Sample sizes by subclass
#>           1   2  3   4   5   6 All
#> Control 102 100 88  72  39  28 429
#> Treated   4   4  9  30  62  76 185
#> Total   106 104 97 102 101 104 614

This works for mean differences but not other statistics. So the way cobalt actually does this is compute stratification weights, and then compute the balance statistics using the stratification weights in the full sample. Stratification weights are first computed by computing the proportion of treated units in each sample, and then using the formulas to compute propensity score weights from propensity scores. Here’s how I do that manually for age:

# Compute proportion of treated units in each subclass
prop1 <- sapply(1:6, function(s) mean(lalonde$treat[msub$subclass == s]))

# Assign to each unit
ps <- prop1[msub$subclass]

# Compute ATE weights
w <- ifelse(lalonde$treat == 1, 1 / ps, 1 / (1 - ps))

# Compute weighted KS statistic
col_w_ks(lalonde$age, treat = lalonde$treat,
         weights = w)
#> [1] 0.1658923

bal.tab(msub, stats = "ks")
#> Balance measures across subclasses
#>                 Type KS.Adj
#> distance    Distance 0.2187
#> age          Contin. 0.1659
#> educ         Contin. 0.0627
#> race_black    Binary 0.0535
#> race_hispan   Binary 0.0420
#> race_white    Binary 0.0115
#> married       Binary 0.1160
#> re74         Contin. 0.3038
#> 
#> Sample sizes by subclass
#>           1   2  3   4   5   6 All
#> Control 102 100 88  72  39  28 429
#> Treated   4   4  9  30  62  76 185
#> Total   106 104 97 102 101 104 614

Why don’t I get the same balance statistics when using cobalt as I do when using tableone?

tableone is another package that provides tools for balance assessment. One strength that the package has is its beautiful, publication-ready tables that include summary statistics for the covariates, clean variable names, and clean headings. But it does not incorporate best practices in balance assessment in favor of transparency. This differs from the ethos of cobalt, which is to provide highly customizable balance statistics that reflect best practices and use well-reasoned decisions. This is not an insult to tableone but is meant to reflect the different purposes cobalt and tableone have. They should not be used interchangeably or expect to yield identical results because they use different formulas for computing certain statistics, most notably the SMD.

Below are some of the reasons why SMDs might differ between tableone and cobalt:

In practice, these differences will be small. Obviously, I recommend using cobalt instead for balance assessment, and I recommend reporting the balance statistics cobalt produces. That said, if you understand what tableone is doing and are okay with the choices it makes, there is no denying that it can produce beautiful tables.

References

Stuart, Elizabeth A. 2008. “Developing Practical Recommendations for the Use of Propensity Scores: Discussion of A Critical Appraisal of Propensity Score Matching in the Medical Literature Between 1996 and 2003’ by Peter Austin, Statistics in Medicine.” Statistics in Medicine 27 (12): 2062–65. https://doi.org/10.1002/sim.3207.
Yang, Dongsheng, and Jarrod E Dalton. 2012. “A Unified Approach to Measuring the Effect Size Between Two Groups Using SAS®,” 6.

  1. For multi-category treatments, all standardization factors are computed using the full data, not just the groups being compared. For example, the pooled standard deviation involves computing the mean of all the group-specific variances, not just the two being compared. Similarly, in the "hedges" formula, \(n-2\) is replaced with \(n-k\), where \(k\) is the number of treatment groups.↩︎