Series Editorsâ€™ Note
We are pleased to add this typescript to the Bone Marrow Transplantation Statistics Series. We realize the term cubic splines may be a bit offputting to some readers, but stay with us and donâ€™t get lost in polynomial equations. What the authors describe is important conceptually and in practice. Have you ever tried to buy a new pair of hiking boots? Getting the correct fit is critical; shoes that are too small or too large will get you in big trouble! Now imagine if hiking shoes came in only 2 sizes, small and large, and your foot size was somewhere in between. You are in trouble. Sailing perhaps?
Transplant physicians are often interested in the association between two variables, say pretransplant measurable residual disease (MRD) test state and an outcome, say cumulative incidence of relapse (CIR). We typically reduce the results of an MRD test to a binary, negative or positive, often defined by an arbitrary cutpoint. However, MRD state is a continuous biological variable, and reducing it to a binary discards what may be important, useful data when we try to correlate it with CIR. Put otherwise, we may miss the trees from the forest.
Another way to look at splines is a technique to make smooth curves out of irregular data points. Consider, for example, trying to describe the surface of an egg. You could do it with a series of straight lines connecting points on the egg surface but a much better representation would be combining groups of points into curves and then combining the curves. To prove this try drawing an egg using the draw feature in Microsoft Powerpoint; you are making splines.
Gauthier and coworkers show us how to use cubic splines to get the maximum information from data points, which may, unkindly, not lend themselves to dichotomization or a best fit line. Please read on. We hope readers will find their typescript interesting and exciting, and that it will give them a new way to think about how to analyse data. And no, a spline is not a bunch of cactus spines.
Robert Peter Gale, Imperial College London, and MeiJie Zhang, Medical College of Wisconsin and CIBMTR.
It is quite common in clinical research, and in the field of hematopoietic cell transplantation (HCT) in particular, to explore the association and predictive ability of a continuous variable with an outcome. For example, one might wish to predict the occurrence of acute graftversushost disease (aGvHD) from the serum concentration of a specific cytokine or maybe the cell count of a newly discovered immunecell subset. Or one might be interested in the association of postHCT blood glucose with the risk of nonrelapse mortality (NRM). In such situations, there are many ways to model the continuous variable of interest.
George Box is largely credited with the aphorism â€śall models are wrong, but some are usefulâ€ť [1]. Indeed, we wish to stress that there is not a single model that is correct, rather that there will almost always be many that are useful. There are, however, similarly many models that are not nearly as useful as they will lead to poor predictions of outcome or fit to the observed data.
Among the more common approaches to modelling continuous data is dichotomising (e.g., biomarker^{high} versus biomarker^{low}) or splitting the variable into several categories. Another common approach is to model the marker as a continuous variable but additionally imposing a linear relationship between the variable and (some function of) the outcome. However, these approaches carry relatively stringent assumptions (see below), and as a result can lead to loss of information. This information loss might impair the predictive ability of the model, or provide a poor fit of the association between the variable and the studied outcome, or damage both. We will try to convince the reader that, at a minimum, alternative modelling approaches should be explored.
The assumption attached to dichotomising a continuous variable into two (or more groups) is that all values of the variable being modelled falling into a common category have the exact same association with outcome. While this assumption maybe approximately true in some cases, it is likely rare that this would hold biologically. For example, if one is interested in modelling age and categorises as 40 and above vs. less than 40, this dichotomisation assumes that a 40year old has the same association with outcome as, say, a 75year old but a different association from, say, a 39year old. It is difficult to imagine that in this example, or in similar scenarios, our assumption would hold true.
Another approach is to model a linear relationship between the continuous variable and outcome. The assumption attached to this modelling is that any change in the variable of a specified size is associated with the same change in outcome, regardless of where the variable â€śstartsâ€ť. For example, if aÂ model suggests that an increase in bilirubin of one unit leads to an increase in the odds of day100 NRM of 30%, this change in odds is the same if bilirubin increases from 1.2 to 2.2â€‰mg/dL orÂ if bilirubin increases from 6.4 to 7.4â€‰mg/dL. It is likely the case that this assumption of linearity would more often be approximately true than the assumption associated with dichotomising data. It is also likely, however, that there are manyÂ instances where this assumption would break down.
As a potential alternative to these modelling strategies (categorising a continuous variable or imposing the assumption of a linear association on a continuous variable), we advocate exploration of nonlinear continuous associations. There are many ways to do this, but we shall focus on one possibilityâ€”restricted cubic splines [2,3,4,5]. As shown in Figs.Â 1a, b, a cubic spline is essentially a piecewise cubic polynomial, where the number of â€śpiecesâ€ť is dictated by the number of windows used. Within each window is effectively a cubic polynomial, and these windows are defined by â€śknotsâ€ť. The mathematics are a bit more complicated than simply fitting a cubic polynomial within each window, as further restrictions need to be imposed so that the spline is continuous (i.e., there is no gap in the spline curve) and â€śsmoothâ€ť at each knot. To better understand this smoothing process, one can imagine trying to model the curved surface of an egg only using straight lines. Intuitively, a much better approach would be to bend or smooth these lines to follow more closely the curvature of the egg (Fig. 2).Â A restricted cubic spline has the additional property that the curve is linear before the first knot and after the last knot. The number of knots used in the spline is determined by the user, but in practice we have found that generally five or fewer knots are sufficient. The location of the knots also needs to be specified by the user, but it is common that the knot with the smallest value is relatively close to the smallest value of the variable being modelled (e.g., the 5th percentile), while the largest knot is in the neighbourhood of the largest value of the variable being modelled (e.g., the 95th percentile).
Consider the scatter plot shown in Fig.Â 1a, with five knots included along the xaxis (at the 5th, 25th, 50th, 75th, and 95th percentiles) and the resulting windows. Within the first and last windows is a simple leastsquares regression line fitting the data. Within each of the interior windows is a cubic polynomial fit to the data. But recall that the cubic spline requires the curve to be continuous and smooth at the knots, so after imposing this condition we get the restricted cubic spline shown in Fig.Â 1b.
A cubic spline with k knots will have k componentsâ€”one constant value (the yintercept), one component that is linear in the variable being modelled (the xvalue), and k2 nonlinear (cubic) components in the modelled variable. In other words, in equation form, y (the outcome) and x (the modelled variable) are associated as
where C_{i}(x) is the cubic component that falls in the ith window, and g is a socalled link function (for example, for logistic regression g is the â€ślogitâ€ť (see below) of the probability of outcome, and for Cox regression g is the socalled loglog transformation of the survivor function).
We can take this idea of a cubic spline to the regression setting, where one assumes that some function of outcome, y, is associated with a continuous variable, x, via the equation specified above. As in any regression setting, the data on outcome and the corresponding value of the covariate for each subject are then used to estimate the coefficients (Î˛_{0} and the Î˛_{i}â€™s) that best fit the observed data. If one is further interested in testing the hypothesis that the modelled association is linear, this can be done by testing that the coefficients associated with the nonlinear components are equal to zero. If we fail to reject this null hypothesis, we might surmise that the association between outcome and the modelled covariate is approximately linear and such a model might be perfectly appropriate as the more complex nonlinear components do not add significant information. If we reject this null hypothesis, however, we conclude that a nonlinear association better describes the data than does a linear association.
Let us consider a few examples, some from simulated data (where we know the true underlying association between a variable and an outcome), and others from real data (where we do not know the true underlying association).
Example 1, Simulated data
Consider a fictitious biomarker (fictitin1) that is associated with the probability of aGvHD after allogeneic HCT (alloHCT). We assume a nonlinear relationship between fictitin1 and the risk of aGvHD. Technically, we assume a nonlinear relationship between fictitin1 and a function of the probability of aGvHD, the socalled logit of aGvHD, or the log of the odds of aGvHD, but visually we will use the probability rather than the logit. This assumedtrue relationship is shown in the solid blue line depicted in Fig.Â 3. We randomly generate a value of fictitin1 from a specified distribution of fictitin1 values, and for each value calculate the corresponding probability of aGvHD based on the assumedtrue association. If this is done n times, we have a simulated data set consisting of n patients, each simulated patient having a fictitin1 value and either having or not having aGvHD. From these â€śobservedâ€ť data, we can fit a logistic regression model by modelling the fictitin1 value in any way that is desired. If we repeat this process m times (say, mâ€‰=â€‰1000 times), we can then take the average of the regression coefficients estimated in each simulated data set (n patients), and use these to plot the â€śaverageâ€ť association under the model that is used to describe the association between fictitin1 and the risk of aGvHD. We consider three different ways to model fictitin1: (1) dichotomising fictitin1 into two groups (e.g., at the median, Fig.Â 3a); (2) categorising fictitin1 into four groups (e.g., quartiles, Fig.Â 3b); and (3) modelling fictitin1 as a restricted cubic spline (Fig.Â 3c).
First, we can appreciate visually that only two or four probabilities can be predicted using the models splitting fictinin1 across its median (Fig.Â 3a) or quartiles (Fig.Â 3b). In contrast, leaving the variable as continuous allows making (potentially unique) predictions for any value of fictitin1 (Fig.Â 3c). In other words, categorising fictitin1 led to loss of information. Second, our predictions (the fitted model shown in red in Fig.Â 3) tend to be further from the â€śtruthâ€ť (the assumedtrue probabilities of aGVHD, depicted in blue in Fig.Â 3) after categorisation, reflected by the red line (our predictions) not overlapping with the blue line (the assumedtrue relationship). In summary, when we modelled fictitin1 using the restricted cubic spline, our predictions were much closer to the assumedtrue probabilities (Fig.Â 3c).
Example 2, Real data
While it is never possible to capture the exact nature of the relationship between a variable and an outcome in the real world, some modelling tools maybe better than others at describing this relationship. Using a dataset of 589 patients who underwent an alloHCT at our institution we will compare different ways of modelling a continuous predictor (in this case the serum glucose concentration measured at day 30 after alloHCT). Here we are interested in predicting the risk of day 200 NRM after alloHCT among patients who survive without relapse to day 30 postHCT. Since our outcome is specified by day 200 and all patients have complete followup by this time, we will use logistic regression for this example as was done above. Similar to our previous example, we consider three different ways to model glucose concentration: (1) dichotomising glucose into two groups (at the median); (2) categorising glucose into four groups (across quartiles); (3) modelling glucose as a restricted cubic spline. Next, we graph the predicted probability of day 200 NRM using these three models (Fig.Â 4).
We again visualise the loss of information associated with categorising the glucose variable across its median (Fig.Â 4a) or quartiles (Fig.Â 4b). Modelling glucose with the restricted cubic spline suggests a strongly nonlinear relationship between glucose and NRM and allows potentially unique predictions for any glucose concentration. As expected, extreme values (both very low and very high) are associated with an increase in the risk of NRM, while a lower risk is predicted for intermediate values (Fig.Â 4c). At odds with human physiology, the models categorising glucose predicted lower NRM for very low concentrations of glucose. More consistent with medical knowledge, an increase in the risk of NRM is predicted by the model using the restricted cubic spline for very low concentrations of glucose. In this example, restricted cubic splines helped us model a complex, nonlinear relationship between glucose concentration and the risk of NRM.
In both the simulated and real data examples above, we utilised logistic regression for illustrative purposes, where the yaxis represents the probability of the outcome under study (acute GVHD in the first, day 200 NRM in the second). One could similarly plot on the yaxis the logit of the probability of failure, or additionally one could plot on the yaxis the odds ratio of failure relative to a particular value of the parameter being modelled. Many outcomes in HCT, however, are modelled as timetoevent endpoints with censored observations contributing to the observed data. In such cases, Cox proportionalhazards models are often fit, allowing for appropriate consideration of censored observations. Cubic splines, or any other method of modelling continuous data as a nonlinear function, can also be used in the timetoevent setting in a manner similar to that in the setting of binary outcomes. In fact, these methods can be used in any regression setting. In the case of Cox regression, plots similar to those for logistic regression can be generated, with the yaxis representing the modelled probability of failure at a particular time or the hazard ratio of failure relative to a particular value of the parameter being modelled.
Restricted cubic splines are not devoid of limitations. In some situations, the use of a large number of knots may lead to â€śoverfittingâ€ť the dataâ€”a phenomenon where the fit of the model corresponds too closely to the observed data, and may therefore fail to provide adequate fit to additional data or predict future observations reliably. In other words, the model may fit â€śnoiseâ€ť more than â€śsignalâ€ť. However, this problem can be managed by evaluating different statistical metrics that are beyond the scope of this editorial. Moreover, the concept of overfitting is not, of course, unique to cubic splines. In general, â€śgood regressionâ€ť practices should be applied with or without the use of restricted cubic splines. We also note that while splines may use more degrees of freedom (the number of parameters that require estimation in a regression model) than a categorical model, a threeknot spline uses three degrees of freedom, as does a quartile model. The fit or predictive ability of cubic splines in the â€śtailsâ€ť of the distribution of the variable being modelled can also deteriorate, this due to sparcity of data. But this is not an issue with splines, per se, rather a problem that exists with any sort of modelling.
Clinicians frequently use cutpoints or threshold values to help the decisionmaking process. Despite their prominence in the field of medicine, their usefulness remains controversial [6]. Cutpoints generally fit only to the currently observed data and therefore rarely replicate across independent studies or data sets. As we have just demonstrated, cutpoints result in the categorisation of a continuous predictor that can be detrimental to risk prediction. That said, there maybe situations in which categorisation of a continuous variable could be helpful. For example, if a particular biomarker has been identified as being important for an outcome and a clinical trial is proposed to target this biomarker, it might be of value to restrict enrolment to patients who have a â€śhighâ€ť level of this biomarker. Categorisation also lends itself to visual presentation of the data that are more natural and familiar to most readers, or an outcome could be based on a particular variable falling within a prescribed window. In any event, we caution the reader to carefully consider the ramifications of categorisation of a continuous variable if, indeed, such categorisation occurs.
In conclusion, restricted cubic splines are a flexible tool to model complex, nonlinear relationships between a continuous variable and an outcome. In general, categorising continuous variables will lead to loss of information and poor predictions (particularly if splitting into only two groups), and this approach should be avoided in most settings [7, 8], or at minimum used with caution. When faced with continuous data, we recommend the exploration of nonlinear associations between the continuous variable being modelled and outcome and we argue that a useful method for such exploration is the use of restricted cubic splines. This approach can be implemented with many statistical software programmes currently available (e.g., R software, â€śrmsâ€ť and â€śsplinesâ€ť package; SAS software). A Shiny web application (https://drjgauthier.shinyapps.io/spliny/) is also offered, allowing interactive visualisations of the use of restricted cubic splines in logistic regression.Â This app modelsÂ various nonlinear relationships and compares predictions between a conventional logistic regression modelÂ and a model using a restricted cubic spline.
References
Box GEP. Robustness in Statistics. Academic Press; 1979. p. 201â€“36. https://doi.org/10.1016/b9780124381506.500182
Stone CGKCY. Additive splines in statistics. In: Proceedings of the American Statistical Association; 1985. p. 45â€“8.
Devlin TF. Spline functions for logistic regression modelling. In: Proceedings of the 11th Annual SAS Users Group International Conference; 1986. p. 646â€“51.
Harrell JFE. Regression modeling strategies. Springer; 2015. https://doi.org/10.1007/9783319194257.
Heinzl H, Kaider A. Gaining more flexibility in Cox proportional hazards regression models with cubic spline functions. Comput Methods Prog Biomed. 1997;54:201â€“8. https://doi.org/10.1016/s01692607(97)000436.
Altman DG, Lausen B, Sauerbrei W, Schumacher M. Dangers of using â€śOptimalâ€ť cutpoints in the evaluation of prognostic factors. JNCI J Natl Cancer Inst. 1994;86:829â€“35. https://doi.org/10.1093/jnci/86.11.829.
Royston P, Altman DG, Sauerbrei W. Dichotomizing continuous predictors in multiple regression: a bad idea. Stat Med. 2006;25:127â€“41. https://doi.org/10.1002/sim.2331.
Altman DG, Royston P. The cost of dichotomising continuous variables. Br Med J. 2006;332:1080 https://doi.org/10.1136/bmj.332.7549.1080.
Author information
Authors and Affiliations
Corresponding author
Ethics declarations
Conflict of interest
The authors declare that they have no conflict of interest.
Additional information
Publisherâ€™s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
Rights and permissions
About this article
Cite this article
Gauthier, J., Wu, Q.V. & Gooley, T.A. Cubic splines to model relationships between continuous variables and outcomes: a guide for clinicians. Bone Marrow Transplant 55, 675â€“680 (2020). https://doi.org/10.1038/s414090190679x
Received:
Accepted:
Published:
Issue Date:
DOI: https://doi.org/10.1038/s414090190679x
This article is cited by

Prognostic impact of changes in aortic stiffness for cardiovascular and mortality outcomes in individuals with type 2 diabetes: the Rio de Janeiro cohort study
Cardiovascular Diabetology (2022)

The relationship between ventilatory ratio (VR) and 28day hospital mortality by restricted cubic splines (RCS) in 14,328 mechanically ventilated ICU patients
BMC Pulmonary Medicine (2022)

Maternal vitamin D during pregnancy and offspring autism and autismassociated traits: a prospective cohort study
Molecular Autism (2022)

Association between the secondstage duration of labor and perinatal outcomes in women with a prior cesarean delivery
BMC Pregnancy and Childbirth (2022)

Intraoperative mild hyperoxia may be associated with improved survival after offpump coronary artery bypass grafting: a retrospective observational study
Perioperative Medicine (2022)