Chapter 7 Estimation of effects: simple and more complex

This exercise deals with analysis of metric and binary response variables. We start with simple estimation of effects of a binary, categorical or a numeric explanatory variable, the explanatory or exposure variable of interest. Then evaluation of potential modification and/or confounding by other variables is considered by stratification by and adjustment/control for these variables. For such tasks we utilize functions lm() and glm() which can be used for more general linear and generalized linear models. Finally, more complex spline modelling for the effect of a numeric exposure variable is illustrated.

7.1 Response and explanatory variables

Identifying the response or outcome variable correctly is the key to analysis. The main types are:

  • Metric or continuous (a measurement with units).
  • Binary (“yes” vs. “no”, coded 1/0), or proportion.
  • Failure in person-time, or incidence rate.

All these response variable are numeric.

Variables on which the response may depend are called explanatory variables or regressors. They can be categorical factors or numeric variables. A further important aspect of explanatory variables is the role they will play in the analysis.

  • Primary role: exposure.
  • Secondary role: confounder and/or effect-measure modifier.

The word effect is used here as a general term referring to ways of contrasting or comparing the expected values of the response variable at different levels of an explanatory variable. The main comparative measures or effect measures are:

  • Differences in means for a metric response.
  • Ratios of odds for a binary response.
  • Ratios of rates for a failure or count response.

Other kinds of contrasts between exposure groups include (a) ratios of geometric means for positive-valued
metric outcomes, (b) differences and ratios between proportions (risk difference and risk ratio), and (c) differences between incidence or mortality rates.

Note that in spite of using the causally loaded word effect, we treat outcome regression modelling here primarily with descriptive or predictive aims in mind. Traditionally, these types of models have also been used to estimate causal effects of exposure variables from the pertinent regression coefficients. More serious causal analysis is introduced in the lecture and practical on Tuesday morning, and modern approaches to estimate causal effects will be considered on Thursday afternoon.

7.2 Data set births

We shall use the births data to illustrate different aspects in estimating effects of various exposures on a metric response variable bweight = birth weight, recorded in grams.

  1. Load the packages needed in this exercise and the data set, and look at its content
library(Epi)
library(mgcv)
data(births)
str(births)
  1. We perform similar housekeeping tasks as in a previous exercise. 
births$hyp <- factor(births$hyp, labels = c("normal", "hyper"))
births$sex <- factor(births$sex, labels = c("M", "F"))
births$maged <- cut(births$matage, breaks = c(22, 35, 44), right = FALSE)
births$gest4 <- cut(births$gestwks,
  breaks = c(20, 35, 37, 39, 45), right = FALSE)
  1. Have a look at univariate summaries of the different variables in the data; especially the location and dispersion of the distribution of bweight.
summary(births)
with(births, sd(bweight))

7.3 Simple estimation with lm() and glm()

We are ready to analyze the effect of maternal hypertension hyp on bweight. A binary explanatory variable, like hyp, leads to an elementary two-group comparison of group means for a metric response.

  1. Comparison of two groups is commonly done by the conventional \(t\)-test and the associated confidence interval.
with(births, t.test(bweight ~ hyp, var.equal = TRUE))

The \(P\)-value refers to the test of the null hypothesis that there is no effect of hyp on birth weight (somewhat implausible null hypothesis in itself!). However, t.test() does not provide the point estimate for the effect of hyp; only the test result and a confidence interval. – The estimated effect of hyp on birth weight, measured as a difference in means between hypertensive and normotensive mothers, is \(3199-2768 = 431\) g.

  1. The same task can easily be performed by lm() or by glm(). The main argument in both is the model formula, the left hand side being the response variable and the right hand side after \(\sim\) defines the explanatory variables and their joint effects on the response. Here the only explanatory variable is the binary factor hyp. With glm() one specifies the family, i.e. the assumed distribution of the response variable. However, in case you use lm(), this argument is not needed, because lm() fits only models for metric responses assuming Gaussian distribution.
m1 <- glm(bweight ~ hyp, family = gaussian, data = births)
summary(m1)
  1. Note the amount of output that summary() method produces. The point estimate plus confidence limits can, though, be concisely obtained by function ci.lin() found in Epi package.
round(ci.lin(m1)[, c(1, 5, 6)], 1)

7.4 Stratified effects, and interaction or effect-measure modification

We shall now examine whether and to what extent the effect of hyp on bweight, i.e. the mean difference between hypertensive and normotensive mothers, varies by sex without assigning causal interpretation to the estimated contrasts.

  1. The following interaction plot shows how the mean bweight depends jointly on hyp and gest4
par(mfrow = c(1, 1))
with(births, interaction.plot(sex, hyp, bweight))

At face value it appears that the mean difference in bweight between hypertensive and normotensive mothers is somewhat bigger in boys than in girls.

  1. Let us get numerical values for the mean differences in the two levels of sex. Stratified estimation of effects can be done by lm() as follows:
m3 <- lm(bweight ~ sex / hyp, data = births)
round(ci.lin(m3)[, c(1, 5, 6)], 1)

The estimated effects of hyp in the two strata defined by sex thus are \(-496\) g in boys and \(-380\) g among girls. The error margins of the two estimates are quite wide, though.

  1. An equivalent model with an explicit product term or interaction term between sex and hyp is fitted as follows:
m3I <- lm(bweight ~ sex + hyp + sex:hyp, data = births)
round(ci.lin(m3I)[, c(1, 4, 5, 6)], 2)

From this output you would find a familiar estimate \(-231\) g for girls vs. boys among normotensive mothers and the estimate \(-496\) g contrasting hypertensive and normotensive mothers in the reference class of sex, i.e. among boys. The remaining coefficient is the estimate of the interaction effect such that \(116.6 = -379.8 -(-496.4)\) g describes the contrast in the effect of hyp on bweight between girls and boys.

The \(P\)-value \(0.46\) as well as the wide confidence interval about zero of this interaction parameter suggest good compatibility of the data with the null hypothesis of no interaction between hyp and sex. Thus, there is insufficient evidence against the possibility of effect(-measure) modification by sex on the effect of hyp. On the other hand, this test is not very sensitive given the small sample size. Thus, in spite of obtaining a “non-significant” result, the possibility of a real effect-measure modification cannot be ignored based on these data only.

7.5 Controlling or adjusting for the effect of hyp for sex

The estimated effects of hyp: \(-496\) in boys and \(-380\) in girls, look quite similar (and the \(P\)-value against no interaction was quite large, too). Therefore, we may now proceed to estimate the overall effect of hyp controlling for – or adjusting for – sex.

  1. Adjustment is done by adding sex to the model formula:
m4 <- lm(bweight ~ sex + hyp, data = births)
ci.lin(m4)[, c(1, 5, 6)]

The estimated effect of hyp on bweight adjusted for sex is thus \(-448\) g, which is a weighted average of the sex-specific estimates. It is slightly different from the unadjusted estimate \(-431\) g, indicating that there was no essential confounding by sex in the simple comparison of means. Note also, that the model being fitted makes the assumption that the estimated effect is the same for boys and girls.

Many people go straight ahead and control for variables which are likely to confound the effect of exposure without bothering to stratify first, but often it is useful to examine the possibility of effect-measure modification before that.

7.6 Numeric exposure, simple linear regression and checking assumptions

If we wished to study the effect of gestation time on the baby’s birth weight then gestwks is a numeric exposure variable.

  1. Assuming that the relationship of the response with gestwks is roughly linear (for a continuous response), % or log-linear (for a binary or failure rate response) we can estimate the linear effect of gestwks with lm() as follows:
m5 <- lm(bweight ~ gestwks, data = births)
ci.lin(m5)[, c(1, 5, 6)]

We have fitted a simple linear regression model and obtained estimates of the two regression coefficient: intercept and slope. The linear effect of gestwks is thus estimated by the slope coefficient, which is \(197\) g per each additional week of gestation.

At this stage it will be best to make some visual check concerning our model assumptions using plot(). In particular, when the main argument for the generic function plot() is a fitted lm object, it will provide you some common diagnostic graphs.

  1. To check whether bweight goes up linearly with gestwks try
with(births, plot(gestwks, bweight))
abline(m5)
  1. Moreover, take a look at the basic diagnostic plots for the fitted model.
par(mfrow = c(2, 2))
plot(m5)

What can you say about the agreement with data of the assumptions of the simple linear regression model, like linearity of the systematic dependence, homoskedasticity and normality of the error terms?

7.7 Penalized spline model

We shall now continue the analysis such that the apparently curved effect of gestwks is modelled by a penalized spline, based on the recommendations of Martyn in his lecture today.

You cannot fit a penalized spline model with lm() or glm(). Instead, function gam() in package mgcv can be used for this purpose. Make sure that you have loaded this package.

  1. When calling gam(), the model formula contains expression ‘s(X)’ for any explanatory variable X, for which you wish to fit a smooth function
mPs <- mgcv::gam(bweight ~ s(gestwks), data = births)
summary(mPs)

From the output given by summary() you find that the estimated intercept is equal to the overall mean birth weight in the data. The estimated residual variance is given by Scale est. or from subobject sig2 of the fitted gam object. Taking square root you will obtain the estimated residual standard deviation: \(445.2\) g.

mPs$sig2
sqrt(mPs$sig2)

The degrees of freedom in this model are not computed as simply as in previous models, and they typically are not integer-valued. However, the fitted spline seems to consume only a little more degrees of freedom as an 3rd degree polynomial model would take.

  1. A graphical presentation of the fitted curve together with the confidence and prediction intervals is more informative. Let us first write a short function script to facilitate the task. We utilize function matshade() in Epi, which creates shaded areas, and function matlines() which draws lines joining the pertinent end points over the \(x\)-values for which the predictions are computed.
plotFitPredInt <- function(xval, fit, pred, ...) {
  matshade(xval, fit, lwd = 2, alpha = 0.2)
  matshade(xval, pred, lwd = 2, alpha = 0.2)
  matlines(xval, fit, lty = 1, lwd = c(3, 2, 2), col = c("black", "blue", "blue"))
  matlines(xval, pred, lty = 1, lwd = c(3, 2, 2), col = c("black", "brown", "brown"))
}
  1. Finally, create a vector of \(x\)-values and compute the fitted/predicted values as well as the interval limits at these points from the fitted model object utilizing function predict(). This function creates a matrix of three columns: (1) fitted/predicted values, (2) lower limits, (3) upper limits and make the graph:
nd <- data.frame(gestwks = seq(24, 45, by = 0.25))
pr.Ps <- predict(mPs, newdata = nd, se.fit = TRUE)
str(pr.Ps) # with se.fit=TRUE, only two columns: fitted value and its SE
fit.Ps <- cbind(
  pr.Ps$fit,
  pr.Ps$fit - 2 * pr.Ps$se.fit,
  pr.Ps$fit + 2 * pr.Ps$se.fit
)
pred.Ps <- cbind(
  pr.Ps$fit, # must add residual variance to se.fit^2
  pr.Ps$fit - 2 * sqrt(pr.Ps$se.fit^2 + mPs$sig2),
  pr.Ps$fit + 2 * sqrt(pr.Ps$se.fit^2 + mPs$sig2)
)
par(mfrow = c(1, 1))
with(births, plot(bweight ~ gestwks,
  xlim = c(24, 45),
  cex.axis = 1.5, cex.lab = 1.5
))
plotFitPredInt(nd$gestwks, fit.Ps, pred.Ps)

Compare this with the graph on slide 20 of the lecture we had. Are you happy with the end result?

7.8 Analysis of binary outcomes

Instead of investigating the distribution and determinants of birth weight as such, it is common in perinatal epidemiology to consider occurrence of low birth weight; whether birth weight is \(< 2.5\) kg or not. Variable lowbw with values 1 and 0 in the births data represents that dichotomy. Some analyses on lowbw were already conducted in a previous practical. Here we illustrate further aspects of effect estimation and modelling binary outcome.

  1. We start with simple tabulation of the prevalence of lowbw by maternal hypertension
stat.table(
  index = list(hyp, lowbw),
  contents = list(count(), percent(lowbw)),
  margins = TRUE, data = births
)

It seems that the prevalence for hypertensive mothers is about 18 percent points higher, or about three times as high as that for normotensive mothers

  1. The three comparative measures of prevalences can be estimated by glm() with different link functions:
binRD <- glm(lowbw ~ hyp, family = binomial(link = "identity"), data = births)
round(ci.lin(binRD)[, c(1, 2, 5:6)], 3)
binRR <- glm(lowbw ~ hyp, family = binomial(link = "log"), data = births)
round(ci.lin(binRR, Exp = TRUE)[, c(1, 2, 5:7)], 3)
binOR <- glm(lowbw ~ hyp, family = binomial(link = "logit"), data = births)
round(ci.lin(binOR, Exp = TRUE)[, c(1, 2, 5:7)], 3)

Check that these results were quite compatible with the “about” estimates given in the previous item. How well is the odds ratio approximating the risk ratio here?

  1. The prevalence of low birth weight is expected to be inversely related to gestational age (weeks), as is evident from simple tabulation
stat.table(
  index = list(gest4, lowbw),
  contents = list(count(), percent(lowbw)),
  margins = TRUE, data = births
)
  1. Let’s jump right away to spline modelling of this relationship
binm1 <- mgcv::gam(lowbw ~ s(gestwks), family = binomial(link = "logit"), data = births)
summary(binm1)
plot(binm1)

Inspect the figure. Would you agree, that the logit of the prevalence of outcome is almost linearly dependent on gestwks?

  1. Encouraged by the result of the previous item, we continue the analysis with glm() and assuming logit-linearity
binm2 <- glm(lowbw ~ I(gestwks - 40), family = binomial(link = "logit"), data = births)
round(ci.lin(binm2, Exp = TRUE)[, c(1, 2, 5:7)], 3)

Inspect the results. How do you interpret the estimated coefficients and their exponentiated values?

  1. Instead of fitted logits, it can be more informative to plot the fitted prevalences against gestwks, in which we utilize the previously created data frame nd
predm2 <- predict(binm2, newdata = nd, type = "response")
plot(nd$gestwks, predm2, type = "l")

The curve seems to cover practically the whole range of the outcome probability scale with a relatively steep slope between 33 to 37 weeks.