Gerhard Tintner: Variate Differences, Weighted Regression, and Multicollinearity (1952)

Gerhard Tintner’s scientific journal article, Die Anwendung der Variate-Difference-Methode auf die Probleme der gewogenen Regression und der Multikollinearität, links three statistical tasks: estimating observation-error covariances, determining the number of independent linear relations among systematic components, and estimating those relations by weighted regression. Its central move is to use the error covariance matrix to distinguish underlying linear dependence from variability introduced by observation errors. The exposition proceeds through equations and large-sample testing procedures, without an empirical application.

Tintner begins by decomposing each observation into a systematic component, identified with its mathematical expectation, and a random component. The random components have no lag correlation and have constant variances and covariances. Since the subsequent analysis requires these error covariances, he proposes estimating them through the Variate Difference method, under a specific restriction:

Dies ist möglich, wenn die Beobachtungsreihen $X_{it}$ Zeitreihen sind und wenn die systematischen Teile $M_{it}$ glatte Funktionen der Zeit sind.

English translation: This is possible when the observation series $X_{it}$ are time series and when the systematic components $M_{it}$ are smooth functions of time.

Smoothness supplies the rationale for differencing: successive differences diminish the contribution of the systematic component, permitting the random covariance structure to be estimated. Tintner computes cross-products of differences of increasing order, normalized by the number of remaining observations and a binomial coefficient. If these covariance estimates become approximately equal across successive orders, their stabilized value is taken as an approximation to the error covariance. He refers to Oskar Anderson for statistical tests of this stabilization; the article itself sets out the procedure rather than deriving those tests.

The second stage turns multicollinearity into a rank problem. Tintner assumes that (R) independent linear equations hold among the systematic components and seeks to estimate their number. After centering the observations and constructing their sample covariance matrix (A), he solves the determinant equation (\lvert A-\lambda V\rvert=0), where (V) is the error covariance matrix. The roots are ordered from smallest to largest. Their importance is not simply that they describe observed covariance, but that they measure it relative to the covariance attributable to error.

For each candidate number of relations, Tintner forms a test statistic from the sum of the corresponding smallest roots, multiplied by (N-1). Here he adds normality of the errors and explicitly acknowledges a limitation:

Die Verteilung von (11) ist nicht bekannt.

English translation: The distribution of (11) is not known.

The proposed inference therefore rests on asymptotic approximations. Drawing on Hsu, Tintner uses a chi-square approximation with (r(N-p-1+r)) degrees of freedom; drawing on T. W. Anderson, he also gives a standardized normal approximation. If the statistic for (R) roots is not significant but that for (R+1) is significant, he takes (R) as the estimated number of independent linear relations among the systematic components. Multicollinearity is thus treated as a question about latent relations, assessed against an explicit model of observational noise.

The final stage estimates the relations themselves. For each of the (R) smallest roots, Tintner solves the homogeneous linear system defined by (A-\lambda_vV). Its solutions supply the relation coefficients, which may be normalized arbitrarily; the constants are then obtained by requiring each fitted relation to pass through the vector of observation means.

Das ist die Methode der gewogenen Regression /7/. Sie folgt aus einer Anwendung der Methode der maximum likelihood, die in unserem Falle zur Methode der kleinsten Quadrate führt.

English translation: This is the method of weighted regression /7/. It follows from an application of the method of maximum likelihood, which in our case leads to the method of least squares.

The article’s contribution is this compact integration of covariance estimation, rank testing, and coefficient estimation. Weighted regression and multicollinearity become successive parts of one procedure, connected through the error covariance matrix. Its scope remains conditional: the initial estimate requires smooth systematic time-series components and uncorrelated errors across lags, while the proposed significance tests require normality and large samples. These qualifications are integral to the argument, not incidental reservations.